Documente Academic
Documente Profesional
Documente Cultură
RECENT ADVANCES
IN NATURAL LANGUAGE PROCESSING
RANLP 2019
PROCEEDINGS
Edited by Galia Angelova, Ruslan Mitkov, Ivelina Nikolova, Irina Temnikova
Varna, Bulgaria
2–4 September, 2019
INTERNATIONAL CONFERENCE
RECENT ADVANCES IN
NATURAL LANGUAGE PROCESSING 2019
PROCEEDINGS
Varna, Bulgaria
2–4 September, 2019
Print ISBN 978-954-452-055-7
Online ISBN 978-954-452-056-4
Series Print ISSN 1313-8502
Series Online ISSN ISSN 2603-2813
Designed and Printed by INCOMA Ltd.
Shoumen, BULGARIA
ii
Preface
Welcome to the 12th International Conference on “Recent Advances in Natural Language Processing”
(RANLP 2019) in Varna, Bulgaria, 2-4 September 2019. The main objective of the conference is to
give researchers the opportunity to present new results in Natural Language Processing (NLP) based on
modern theories and methodologies.
The Conference is preceded by the First Summer school on Deep Learning in NLP (29-30 August 2019)
and two days of tutorials (31 August – 1 September 2019).
The Summer School lectures are given by Kyunghyun Cho (New York University), Marek Rei
(University of Cambridge), Tim Rocktäschel (University College London) and Hinrich Schütze (Ludwig
Maximilian University, Munich). Training in practical sessions is provided by Heike Adel (University
of Stuttgart), Alexander Popov (Institute of Information and Communication Technologies, Bulgarian
Academy of Sciences), Omid Rohanian and Shiva Taslimipoor (University of Wolverhampton).
Tutorials are given by the following lecturers: Preslav Nakov (Qatar Computing Research Institute,
HBKU), Valia Kordoni (Humboldt University, Berlin), Antonio Miceli Barone (University of Edinburgh)
and Sheila Castilho (Dublin City University), Vlad Niculae and Tsvetomila Mihaylova (Institute of
Telecommunications, Lisbon).
This year 18 regular papers, 37 short papers, 95 posters, and 7 demos have been accepted for presentation
at the conference. The proceedings cover a wide variety of NLP topics, including but not limited to:
deep learning; machine translation; opinion mining and sentiment analysis; semantics and discourse;
named entity recognition; coreference resolution; corpus annotation; parsing and morphology; text
summarisation and simplification; event extraction; fact checking and rumour analysis; NLP for
healthcare; and NLP for social media.
In 2019 RANLP hosts four post-conference workshops on influential NLP topics: the 2nd Workshop
on Human-Informed Translation and Interpreting Technology (HiT-IT 2019), the 12th Workshop on
Building and Using Comparable Corpora (BUCC), the Multiling 2019 Workshop: Summarization Across
Languages, Genres and Sources as well as an Workshop on Language Technology for Digital Historical
Archives with a Special Focus on Central-, (South-)Eastern Europe, Middle East and North Africa.
We would like to thank all members of the Programme Committee and all additional reviewers. Together
they have ensured that the best papers were included in the Proceedings and have provided invaluable
comments for the authors.
Finally, special thanks go to the University of Wolverhampton, the Institute of Information and
Communication Technologies at the Bulgarian Academy of Sciences, the Bulgarian National Science
Fund, Ontotext and IRIS.AI for their generous support of RANLP.
iii
The International Conference RANLP–2019 is organised by:
Ontotext AD
IRIS.AI
Workshop Coordinator:
Kiril Simov, Bulgarian Academy of Sciences, Bulgaria
Tutorial Coordinator:
Preslav Nakov, Qatar Computing Research Institute, HBKU, Qatar
Proceedings Printing:
Nikolai Nikolov, INCOMA Ltd., Shoumen, Bulgaria
v
Programme Committee:
Ahmed Abdelali (Hamad Bin Khalifa University, Qatar)
Cengiz Acarturk (Middle East Technical University, Turkey)
Guadalupe Aguado-de-Cea (Polytechnic University of Madrid, Spain)
Luis Alfonso Ureña López (University of Jaén, Spain)
Hassina Aliane (Research Center on Scientific and Technical Information, Algeria)
Pascal Amsili (University of Paris Diderot, France)
Galia Angelova (Bulgarian Academy of Sciences, Bulgaria)
Riza Batista-Navarro (University of Manchester, United Kingdom)
Kalina Bontcheva (University of Sheffield, United Kingdom)
Svetla Boytcheva (Bulgarian Academy of Sciences, Bulgaria)
António Branco (University of Lisbon, Portugal)
Chris Brew (Digital Operatives)
Nicoletta Calzolari (Italian National Research Council, Italy)
Sheila Castilho (Dublin City University, Ireland)
Key-Sun Choi (Korea Advanced Institute of Science and Technology, South Korea)
Kenneth Church (Baidu, United States of America)
Kevin Cohen (University of Colorado School of Medicine, United States of America)
Gloria Corpas Pastor (University of Málaga, Spain)
Dan Cristea (University of Las, i, Romania)
Antonio Ferrández Rodrı́guez (University of Alicante, Spain)
Fumiyo Fukumoto (University of Yamanashi, Japan)
Prószéky Gábor (Pázmány Péter Catholic University & Bionics, Hungary)
Alexander Gelbukh (National Polytechnic Institute, Mexico)
Yota Georgakopoulou (Athena Consultancy, Greece)
Ralph Grishman (New York University, United States of America)
Veronique Hoste (Ghent University, Belgium)
Diana Inkpen (University of Ottawa, Canada)
Hitoshi Isahara (Toyohashi University of Technology, Japan)
Miloš Jakubı́ček (Lexical Computing Ltd)
Alma Kharrat (Microsoft)
Udo Kruschwitz (University of Essex, United Kingdom)
Sandra Kübler (Indiana University, United States of America)
Katia Lida Kermanidis (Ionian University, Greece)
Natalia Loukachevitch (Lomonosov Moscow State University, Russia)
Eid Mohamed (Doha Institute for Graduate Studies, Qatar)
Emad Mohamed (University of Wolverhampton, United Kingdom)
Johanna Monti (University of Naples L’Orientale, Italy)
Andrés Montoyo (University of Alicante, Spain)
Alessandro Moschitti (Amazon)
Rafael Muñoz Guillena (University of Alicante, Spain)
Preslav Nakov (Qatar Computing Research Institute, Qatar)
Roberto Navigli (Sapienza University of Rome, Italy)
Raheel Nawaz (Manchester Metropolitan University, United Kingdom)
Mark-Jan Nederhof (University of St Andrews, United Kingdom)
Ivelina Nikolova (Bulgarian Academy of Sciences, Bulgaria)
Kemal Oflazer (Carnegie Mellon University, Qatar)
Maciej Ogrodniczuk (Polish Academy of Sciences, Poland)
vi
Constantin Orasan (University of Wolverhampton, United Kingdom)
Petya Osenova (Sofia University and Bulgarian Academy of Sciences, Bulgaria)
Sebastian Padó (Stuttgart University, Germany)
Noa P. Cruz Diaz (Artificial Intelligence Excellence Center, Bankia, Spain)
Liviu P. Dinu (University of Bucharest, Romania)
Pavel Pecina (Charles University, Czech Republic)
Stelios Piperidis (Athena Research Center, Greece)
Massimo Poesio (University of Essex, United Kingdom)
Horacio Rodrı́guez (Polytechnic University of Catalonia, Spain)
Paolo Rosso (Polytechnic University of Valencia, Spain)
Vasile Rus (The University of Memphis, United States of America)
Frédérique Segond (Viseo)
Kiril Simov (Bulgarian Academy of Sciences, Bulgaria)
Vilelmini Sosoni (Ionian University, Greece)
Keh-Yih Su (Institute of Information Science, Academia Sinica, Taiwan)
Stan Szpakowicz (University of Ottawa, Canada)
Hristo Tanev (European Commission, Belgium)
Shiva Taslimipoor (University of Wolverhampton, United Kingdom)
Irina Temnikova (Sofia University, Bulgaria)
Dan Tufis, (Romanian Academy of Sciences, Romania)
Aline Villavicencio (University of Essex, United Kingdom and Federal University of Rio
Grande do Sul, Brazil)
Yorick Wilks (Florida Institute for Human and Machine Cognition, United States of
America )
Mai Zaki (American University of Sharjah, United Arab Emirates)
Marcos Zampieri (University of Wolverhampton, United Kingdom)
Michael Zock (University of Aix-Marseille, France)
Reviewers:
Ahmed AbuRa’ed (University Pompeu Fabra, Spain)
Mattia A. Di Gangi (University of Trento, Italy)
Itziar Aldabe (University of Paı́s Vasco, Spain)
Ahmed Ali (Hamad Bin Khalifa University, Qatar)
Ahmed Amine Aliane (Research Center on Scientific and Technical Information, Alge-
ria)
Le An Ha (University of Wolverhampton, United Kingdom)
Atefeh (Anna) Farzindar (University of Southern California, United States of America)
João António Rodrigues (University of Lisboa, Portugal)
Pepa Atanasova (University of Copenhagen, Denmark)
Mohammed Attia (George Washington University, United States of America)
Parnia Bahar (Aachen University, Germany)
Belahcene Bahloul (University of Khemis Miliana, Algeria)
Eduard Barbu (University of Tartu, Estonia)
Alberto Barrón-Cedeño (University of Bologna, Italy)
Leonor Becerra (Jean Monnet University, France)
Andrea Bellandi (National Research Council, Italy)
Fernando Benites (ZHAW School of Engineering, Switzerland)
Victoria Bobicev (Technical University of Moldova, Moldova)
vii
Antonina Bondarenko (Lipetsk State Technical University, Russia)
Aurélien Bossard (University Paris 8, France)
Aljoscha Burchardt (German Research Centre for Artificial Intelligence, Germany)
Lindsay Bywood (University of Westminster, United Kingdom)
Ruket Cakici (Middle East Technical University, Turkey)
Iacer Calixto (University of Amsterdam, Netherlands and New York University, United
States of America)
Pablo Calleja (Polytechnic University of Madrid, Spain)
Erik Cambria (Nanyang Technological University, Singapore)
Kai Cao (New York University, United States of America)
Thiago Castro Ferreira (Tilburg University, Netherlands)
Yue Chen (Queen Mary University of London, United Kingdom)
Mihaela Colhon (University of Craiova, Romania)
Daniel Dakota (Indiana University, United States of America)
Kareem Darwish (Hamad Bin Khalifa University, Qatar)
Orphee De Clercq (Ghent University, Belgium)
Kevin Deturck (Viseo)
Asma Djaidri (University of Science and Technology Houari Boumediene, Algeria)
Mazen Elagamy (Staffordshire University, United Kingdom)
Can Erten (University of York, United Kingdom)
Luis Espinosa Anke (Cardiff University, United Kingdom)
Kilian Evang (University of Düsseldorf, Germany)
Richard Evans (University of Wolverhampton, United Kingdom)
Stefan Evert (Friedrich–Alexander University, Germany)
Anna Feherova (University of Wolverhampton, United Kingdom)
Mariano Felice (University of Cambridge, United Kingdom)
Corina Forascu (The Alexandru Ioan Cuza University, Romania)
Vasiliki Foufi (University of Geneva, Switzerland)
Thomas Francois (Université catholique de Louvain, Belgium)
Adam Funk (University of Sheffield, United Kingdom)
Björn Gambäck (Norwegian University of Science and Technology, Norway)
Aina Garı́ Soler (The Computer Science Laboratory for Mechanics and Engineering Sci-
ences, France)
Federico Gaspari (Dublin City University, Ireland)
José G. C. de Souza (eBay)
Goran Glavaš (University of Mannheim, Germany)
Darina Gold (University of Duisburg-Essen, Germany)
Reshmi Gopalakrishna Pillai (University of Wolverhampton, United Kingdom)
Rohit Gupta (University of Wolverhampton, United Kingdom)
Amir Hazem (Nantes University, France)
Tomáš Hercig (University of West Bohemia, Czech Republic)
Yasser Hifny (University of Helwan, Egypt)
Diliara Iakubova (Kazan Federal University, Russia)
Adrian Iftene (The Alexandru Ioan Cuza University, Romania)
Camelia Ignat (European Commission, Belgium)
Dmitry Ilvovsky (National Research University Higher School of Economics, Russia)
Miloš Jakubı́ček (Masaryk University, Czech Republic)
Arkadiusz Janz (Wroclaw University of Science and Technology, Poland)
viii
Héctor Jiménez-Salazar (The Metropolitan Autonomous University, Mexico)
Olga Kanishcheva (National Technical University, Ukraine)
Georgi Karadzhov (Sofia University, Bulgaria)
David Kauchak (Pomona College, United States of America)
Yasen Kiprov (Sofia University, Bulgaria)
Jan Kocoń (Wroclaw University of Science and Technology, Poland)
Sarah Kohail (Hamburg University, Germany)
Yannis Korkontzelos (Edge Hill University, United Kingdom)
Venelin Kovatchev (University of Barcelona, Spain)
Peter Krejzl (University of West Bohemia, Czech Republic)
Sudip Kumar Naskar (Jadavpur University, India)
Maria Kunilovskaya (University of Tyumen, Russia)
Andrey Kutuzov (University of Oslo, Norway)
Sobha Lalitha Devi (Anna University, India)
Gabriella Lapesa (Universität Stuttgart, Institut für Maschinelle Sprachverarbeitung, Ger-
many)
Todor Lazarov (Bulgarian Academy of Sciences, Bulgaria)
Els Lefever (Ghent University, Belgium)
Ladislav Lenc (University of West Bohemia, Czech Republic)
Elena Lloret (University of Alicante)
Pintu Lohar (Dublin City University, Ireland)
Epida Loupaki (Aristotle University of Thessaloniki, Greece)
Lieve Macken (Ghent University, Belgium)
Mireille Makary (University of Wolverhampton, United Kingdom)
Michał Marcińczuk (Wroclaw University of Technology, Poland)
Angelo Mario Del Grosso (National Research Council of Italy, Italy)
Federico Martelli (Babelscape, Italy)
Patricia Martin Chozas (Polytechnic University of Madrid, Spain)
Eugenio Martı́nez-Cámara (University of Granada, Spain)
Irina Matveeva (NexLP, Unites States of America)
Flor Miriam Plaza del Arco (University of Jaén, Spain)
Arturo Montejo-Ráez (University of Jaén, Spain)
Paloma Moreda Pozo (University of Alicante, Spain)
Diego Moussallem (University of Paderborn, Germany)
Sara Moze (University of Wolverhampton, United Kingdom)
Nona Naderi (University of Toronto, Canada)
Marcin Oleksy (Wrocław University of Science and Technology, Poland)
Antoni Oliver (The Open University of Catalonia, Spain)
Mihaela Onofrei (University of Iasi, Romania)
Arzucan Özgür (Bogazici University, Tyrkey)
Santanu Pal (Saarland University, Germany)
Alexander Panchenko (University of Hamburg, Germany)
Sean Papay (University of Stuttgart, Germany)
Ljudmila Petković (University of Belgrade, Serbia)
Maciej Piasecki (Wroclaw University of Science and Technology, Poland)
Paul Piwek (The Open University, United Kingdom)
Alistair Plum (University of Wolverhampton, United Kingdom)
Alberto Poncelas (Dublin City University, Ireland)
ix
Alexander Popov (Bulgarian Academy of Sciences, Bulgaria)
Maja Popović ( Dublin City University, Ireland)
Dan Povey (Johns Hopkins University, United States of America)
Ondřej Pražák (University of West Bohemia, Czech Republic)
Prokopis Prokopidis (Research and Innovation Center in Information, Greece)
Tharindu Ranasinghe (University of Wolverhampton, United Kingdom)
Natalia Resende (Dublin City University, Ireland)
Pattabhi RK Rao (Anna University, India)
Omid Rohanian (University of Wolverhampton, United Kingdom)
Josef Ruppenhofer (Institute for the German Language, Germany)
Pavel Rychlý (Masaryk University, Czech Republic)
Magdaléna Rysová (Charles University, Czech Republic)
Branislava Šandrih (Belgrade University, Serbia)
Estela Saquete (University of Alicante, Spain)
Leah Schaede (Indiana University, United States)
Ineke Schuurman (University of Leuven, Belgium)
Olga Seminck (Paris Diderot University, France)
Nasredine Semmar (Laboratory for Integration of Systems and Technology, France)
Matthew Shardlow (Manchester Metropolitan University, United Kingdom)
Artem Shelmanov (Russian Academy of Sciences, Russia)
Dimitar Shterionov (Dublin City University, Ireland)
Jennifer Sikos (University of Stuttgart, Germany)
João Silva (University of Lisboa, Portugal)
Vasiliki Simaki (Lancaster University, United Kingdom)
Sunayana Sitaram (Microsoft Research, India)
Mihailo Skoric (Researcher, Serbia)
Felix Stahlberg (University of Cambridge, Department of Engineering, United Kingdom)
Kenneth Steimel (Indiana University, United States)
Sebastian Stüker (Karlsruhe Institute of Technology, Germany)
Yoshimi Suzuki (Shizuoka University, Japan)
Liling Tan (Nanyang Technological University, Singapore)
Segun Taofeek Aroyehun (National Polytechnic Institute, Mexico )
Laura Tolos, i (Self employed data scientist)
Elena Tutubalina (Kazan Federal University, Russia)
Eleni Tziafa (National and Kapodistrian Unversity of Athens, Greece)
Antonio Valerio Miceli Barone (University of Edinburgh, United Kingdom)
Mihaela Vela (Saarland University, Germany)
Cristina Vertan (University of Hamburg, Germany)
Manuel Vilares Ferro (University of Vigo, Spain)
Veronika Vincze (University of Szeged, Hungary)
Pidong Wang (National University of Singapore, Singapore)
Michael Wiegand (Heidelberg University, Germany)
Victoria Yaneva (University of Wolverhampton, United Kingdom)
Kristina Yordanova (University of Rostock, Germany)
Juntao Yu (Queen Mary University of London, United Kingdom)
Wajdi Zaghouani (Hamad Bin Khalifa University, Qatar)
Kalliopi Zervanou (Eindhoven University of Technology, Netherlands)
Inès Zribi (University of Sfax, Tunisia)
x
Table of Contents
Bilingual Low-Resource Neural Machine Translation with Round-Tripping: The Case of Persian-Spanish
Benyamin Ahmadnia and Bonnie Dorr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Quasi Bidirectional Encoder Representations from Transformers for Word Sense Disambiguation
Michele Bevilacqua and Roberto Navigli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
xi
Learning Sentence Embeddings for Coherence Modelling and Beyond
Tanner Bohn, Yining Hu, Jinhang Zhang and Charles Ling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Risk Factors Extraction from Clinical Texts Based on Linked Open Data
Svetla Boytcheva, Galia Angelova and Zhivko Angelov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Parallel Sentence Retrieval From Comparable Corpora for Biomedical Text Simplification
Rémi Cardon and Natalia Grabar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Sparse Victory – A Large Scale Systematic Comparison of Count-Based and Prediction-Based Vectoriz-
ers for Text Classification
Rupak Chakraborty, Ashima Elhence and Kapil Arora . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
OlloBot - Towards A Text-Based Arabic Health Conversational Agent: Evaluation and Results
Ahmed Fadhil and Ahmed AbuRa’ed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
xii
Summarizing Legal Rulings: Comparative Experiments
Diego Feijo and Viviane Moreira . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
Automatic Question Answering for Medical MCQs: Can It Go Further than Information Retrieval?
Le An Ha and Victoria Yaneva . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 418
From the Paft to the Fiiture: A Fully Automatic NMT and Word Embeddings Method for OCR Post-
Correction
Mika Hämäläinen and Simon Hengchen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
Investigating Terminology Translation in Statistical and Neural Machine Translation: A Case Study on
English-to-Hindi and Hindi-to-English
Rejwanul Haque, Md Hasanuzzaman and Andy Way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
Beyond English-Only Reading Comprehension: Experiments in Zero-Shot Multilingual Transfer for Bul-
garian
Momchil Hardalov, Ivan Koychev and Preslav Nakov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
xiii
Tweaks and Tricks for Word Embedding Disruptions
Amir Hazem and Nicolas Hernandez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
Emoji Powered Capsule Network to Detect Type and Target of Offensive Posts in Social Media
Hansi Hettiarachchi and Tharindu Ranasinghe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
EoANN: Lexical Semantic Relation Classification Using an Ensemble of Artificial Neural Networks
Rayehe Hosseini Pour and Mehrnoush Shamsfard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
Opinions Summarization: Aspect Similarity Recognition Relaxes the Constraint of Predefined Aspects
Nguyen Huy Tien, Le Tung Thanh and Nguyen Minh Le . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
Word Sense Disambiguation Based on Constrained Random Walks in Linked Semantic Networks
Arkadiusz Janz and Maciej Piasecki . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
Is Similarity Visually Grounded? Computational Model of Similarity for the Estonian Language
Claudia Kittask and Eduard Barbu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 542
Multi-Level Analysis and Recognition of the Text Sentiment on the Example of Consumer Opinions
Jan Kocoń, Monika Zaśko-Zielińska and Piotr Miłkowski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560
xiv
Resolving Pronouns for a Resource-Poor Language, Malayalam Using Resource-Rich Language, Tamil.
Sobha Lalitha Devi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
Semantic Role Labeling with Pretrained Language Models for Known and Unknown Predicates
Daniil Larionov, Artem Shelmanov, Elena Chistova and Ivan Smirnov . . . . . . . . . . . . . . . . . . . . . . 620
The Impact of Semantic Linguistic Features in Relation Extraction: A Logical Relational Learning Ap-
proach
Rinaldo Lima, Bernard Espinasse and Frederico Freitas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
v-trel: Vocabulary Trainer for Tracing Word Relations - An Implicit Crowdsourcing Approach
Verena Lyding, Christos Rodosthenous, Federico Sangati, Umair ul Hassan, Lionel Nicolas, Alexan-
der König, Jolita Horbacauskiene and Anisia Katinskaia. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .675
Jointly Learning Author and Annotated Character N-gram Embeddings: A Case Study in Literary Text
Suraj Maharjan, Deepthi Mave, Prasha Shrestha, Manuel Montes, Fabio A. González and Thamar
Solorio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
Inforex — a Collaborative System for Text Corpora Annotation and Analysis Goes Open
Michał Marcińczuk and Marcin Oleksy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
Turning Silver into Gold: Error-Focused Corpus Reannotation with Active Learning
Pierre André Ménard and Antoine Mougeot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 759
xv
NLP Community Perspectives on Replicability
Margot Mieskes, Karën Fort, Aurélie Névéol, Cyril Grouin and Kevin Cohen . . . . . . . . . . . . . . . . 769
incom.py - A Toolbox for Calculating Linguistic Distances and Asymmetries between Related Languages
Marius Mosbach, Irina Stenger, Tania Avgustinova and Dietrich Klakow . . . . . . . . . . . . . . . . . . . . 811
From Image to Text in Sentiment Analysis via Regression and Deep Learning
Daniela Onita, Liviu P. Dinu and Adriana Birlutiu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 863
Toponym Detection in the Bio-Medical Domain: A Hybrid Approach with Deep Learning
Alistair Plum, Tharindu Ranasinghe and Constantin Orăsan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 913
xvi
Combining SMT and NMT Back-Translated Data for Efficient NMT
Alberto Poncelas, Maja Popović, Dimitar Shterionov, Gideon Maillette de Buy Wenniger and Andy
Way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 923
Enhancing Unsupervised Sentence Similarity Methods with Deep Contextualised Word Representations
Tharindu Ranasinghe, Constantin Orăsan and Ruslan Mitkov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 995
Analysing the Impact of Supervised Machine Learning on Automatic Term Extraction: HAMLET vs
TermoStat
Ayla Rigouts Terryn, Patrick Drouin, Veronique Hoste and Els Lefever . . . . . . . . . . . . . . . . . . . . . 1013
Persistence Pays Off: Paying Attention to What the LSTM Gating Mechanism Persists
Giancarlo Salton and John Kelleher . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1053
Development and Evaluation of Three Named Entity Recognition Systems for Serbian - The Case of
Personal Names
Branislava Šandrih, Cvetana Krstev and Ranka Stankovic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
Moral Stance Recognition and Polarity Classification from Twitter and Elicited Text
Wesley Santos and Ivandré Paraboni . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1070
xvii
The "Jump and Stay" Method to Discover Proper Verb Centered Constructions in Corpus Lattices
Bálint Sass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1077
Automated Text Simplification as a Preprocessing Step for Machine Translation into an Under-Resourced
Language
Sanja Štajner and Maja Popović . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1142
Augmenting a BiLSTM Tagger with a Morphological Lexicon and a Lexical Category Identification Step
Steinþór Steingrímsson, Örvar Kárason and Hrafn Loftsson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1162
Comparison of Machine Learning Approaches for Industry Classification Based on Textual Descriptions
of Companies
Andrey Tagarev, Nikola Tulechki and Svetla Boytcheva . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1170
Text-Based Joint Prediction of Numeric and Categorical Attributes of Entities in Knowledge Bases
V Thejas, Abhijeet Gupta and Sebastian Padó . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1195
It Takes Nine to Smell a Rat: Neural Multi-Task Learning for Check-Worthiness Prediction
Slavena Vasileva, Pepa Atanasova, Lluís Màrquez, Alberto Barrón-Cedeño and Preslav Nakov1230
xviii
Deep Learning Contextual Models for Prediction of Sport Events Outcome from Sportsmen Interviews
Boris Velichkov, Ivan Koychev and Svetla Boytcheva . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1241
Exploiting Frame-Semantics and Frame-Semantic Parsing for Automatic Extraction of Typological In-
formation from Descriptive Grammars of Natural Languages
Shafqat Mumtaz Virk, Azam Sheikh Muhammad, Lars Borin, Muhammad Irfan Aslam, Saania
Iqbal and Nazia Khurram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1248
Towards Adaptive Text Summarization: How Does Compression Rate Affect Summary Readability of L2
Texts?
Tatiana Vodolazova and Elena Lloret . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1266
ETNLP: A Visual-Aided Systematic Approach to Select Pre-Trained Embeddings for a Downstream Task
Son Vu Xuan, Thanh Vu, Son Tran and Lili Jiang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1286
Bigger versus Similar: Selecting a Background Corpus for First Story Detection Based on Distributional
Similarity
Fei Wang, Robert J. Ross and John D. Kelleher . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1313
xix
Multilingual Dynamic Topic Model
Elaine Zosa and Mark Granroth-Wilding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1389
xx
Table Structure Recognition Based on Cell Relationship, a Bottom-Up
Approach
1
Proceedings of Recent Advances in Natural Language Processing, pages 1–8,
Varna, Bulgaria, Sep 2–4 2019.
lenged the researchers over the last two decades. spatial distance of neighboring words.
An ample amount of research work has been car- Klampfl et al. (2014) experimented with two un-
ried out leading to a diverse list of approaches in- supervised approaches for table recognition and
cluding heuristics, rule-engine, and recently ma- showcased the importance of spatial distances be-
chine learning based proposals. tween words of a table using vertical and horizon-
We believe, generalizing the patterns across the tal histogram projection of words coordinates.
variety of table layouts in diverse type of doc- Experiments using rule-engine has been pro-
uments, is best solved by machine learning ap- posed by Shigarov (2015), considers the physi-
proach. We propose a bottom-up approach for ta- cal layout of a rendered table, and the logical lay-
ble structure recognition as a cell relation extrac- out representing the relationships between the ele-
tion task between the table text tokens using deep ments of a table, differently. Another work of Shi-
learning. The way a human understands the ta- garov et al. (2016) shows promising results in rec-
bles can be analogous to the proposed approach. ognizing the columns and rows of tables by using
Often, relation extraction task involves classifica- the word and line distances, the order of appear-
tion of an entity pair to a set of known relations, ance of text chunks. The methodology makes use
using documents containing mentions of the en- of configurable thresholds in its heuristic decision
tity pair (Kumar, 2017). By considering the table making.
recognition task as a relationship extraction prob- The heuristic and rule-based solutions make
lem, we introduce a novel approach suitable for various assumptions on the visual, type and con-
several document understanding solutions. tent, structural details of tables and the thresholds
The proposed method deals with the basic used in the algorithms. These assumptions may
building blocks of any table, the table cells. With not hold on heterogeneous documents and may
this approach, we hope to solve the column and even break the system.
row spanning, the presence or absence of borders,
Perez-Arriaga et al. (2016) have made use of
and other challenges mentioned earlier. The table
both k-nearest neighbor and layout heuristics,
recognition system operates at token-level and in-
making it a hybrid methodology to recognize the
volves learning the complex patterns in order to
table structure. The method groups the words
extract the cell relationships among the table text
into rows and columns using spatial distances of
tokens using deep learning.
words heuristically. Interestingly, the spatial dis-
tance thresholds are learned using the k-nearest
2 Related Work
neighbor algorithm. Their work also proposes a
According to the well-known ICDAR 2013 Table heuristic method to identify the headers of the ta-
Competition (Gbel et al., 2013), the problem of ble. Deep learning based semantic segmentation
table understanding can be split into table location has been used by Schreiber et al. (2017) where
detection, table structure recognition, and table in- an image of a document is fed to the neural net-
terpretation. Each of these sub-problems has at- work to identify the rows and columns of a ta-
tracted a great deal of attention from researchers ble. However, the work makes use of a heuristic
and has extensive work. post-processing step to improve the table structure
A peek at the literature shows that many heuris- recognition.
tic solutions have been proposed for table struc- Clinchant et al. (2018) have made an exten-
ture recognition. Most of those work consider sive comparison of three different Machine Learn-
the white space and layout analysis. Yildiz et al. ing approaches to recognize the table structure
(2005) propose an algorithm to recognize the in hand-written register books. The method first
columns of a table using distances between lines recognizes the cell locations and then groups the
and then identify the cells to find the rows. The al- cells into rows. The experimentations do a thor-
gorithm makes a few assumptions about the struc- ough comparison of CRF, a variation of Graph-CN
ture of the tables. The work of Krüpl and Her- called Edge-CN, and conventional Logistic Re-
zog (2006) takes a bottom-up approach towards gression algorithms. However, the method works
structure recognition using heuristics but works on on already recognized headers and columns of the
browser-rendered documents. The methodology table and addresses only row recognition task.
aggregates words into columns by considering the To the best of our knowledge, most of the re-
2
lated work of table recognition try to identify the guistics, determining whether two or more expres-
columns and rows of tables first and then locate the sions in a text refer to the same person or thing,
intersections of rows and columns as table cells. is a relation extraction task. We take the idea of
A few heuristics based works have considered the relation extraction and formulate the task of ta-
grouping of words into blocks and then aggregat- ble structure recognition as identifying the cell-
ing blocks into rows and columns. A common relationship among all the content of individual
downside of these methodologies is that they fail cells of a table. When the tokens that are part of
to capture the information about rows and columns a table are considered as the smallest possible ele-
spanning multiple table cells. Few of the heuristic ments of a table, the relation extraction task will be
approaches do try to solve this issue however, they to identify whether given two tokens of table text,
fail to generalize the solution. belong to the same cell or not. If two tokens be-
We propose a purely bottom-up approach by long to the same cell, then those two tokens have a
building the table structure by recognizing the in- belong-to-same-cell relation. In our experiments,
dividual cells of the table and their location in the this task of binary relationship extraction is con-
document. The task of recognizing the table cells sidered as a binary classification problem.
is addressed as cell-relation extraction between the For every pair of tokens, the goal of binary rela-
tokens present in the table. tionship classifier is to determine whether the two
tokens belong to the same cell or not. An impor-
3 Methodology tant thing to note here is that this relationship be-
tween the two tokens is transitive. If a token A is
In this section, we first explain how we modeled
related to the token B and the token B is related
the table structure recognition as a relation extrac-
to the token C, then the token A is related to C.
tion task, then the training data preparation, and
Hence, we don’t need to generate feature vectors
finally describe how the binary relationship clas-
of all the possible pairs of tokens in a cell to deter-
sification is modeled using a Multilayer Feedfor-
mine all the tokens of a cell, we only need to make
ward Neural Network.
sure that all the tokens of a cell are connected via
3.1 Cell Relationship Extraction in Table a chain of such transitive dependencies.
Structure Recognition Once we predict the belong-to-same-cell rela-
tion between token pairs, we group all the table
Humans will recognize the table structure even tokens into different cells. This is a simple task of
without a need for borders, based on visual clues, aggregating different token pairs into their respec-
spatial distances and the content of the cells. tive cells. Using this data of all the table cells and
These visual clues present in the tables help the the tokens in each of the cells, we can model the
readers to recognize the location of table cells eas- recognition of rows and columns of the table again
ily, by bringing all the words of a cell together both as a relation extraction task between the pairs of
visually and semantically. The proposed method is table cells themselves. However, in this work, we
based on this idea of identification of cell relation- concentrate only on cell recognition.
ship among the table words. The first step towards
the recognition of rows or columns is the identifi- 3.2 Data Preparation
cation of table cells and thus the whole process of
table structure recognition is a bottom-up process. Detecting Tables and Training Data
This reasoning is based on the underlying defini- Generation
tion of any table: Unit of a table is a cell, hori- The detection of the location of tables in PDF doc-
zontal and vertical alignment of cells forms rows uments is the first task in the process of table ex-
and columns, respectively. Tokens are generated traction and this location information is prerequi-
by using white-space and new-line characters as site to our system. There are several open source
delimiters. In this paper the terms token and word and free of charge tools for detecting table loca-
are used interchangeably. tions in PDF documents. In our experiments, we
Relation extraction is a well-known task in Nat- used Tabula to obtain the location details of a table
ural Language Processing, which deals with clas- in the document. The location of a table is repre-
sifying whether a given set of n samples have any sented by five values, pageNum, (startX, startY)
of m different relationships. For example, in lin- and (endX, endY). All the coordinates are assigned
3
Figure 1: Features used in cell-relation extraction. A) T00, T01, T02,... are the tiles in the first row of tile
matrix. B) T00, T10, T20,... are the tiles in the first column of tile matrix. C) (x1,y1) and (x2,y2) are the
start and end coordinates of a token. D) sameCell=1, Pair of tokens which have sameCell relationship.
E) samceCell=0, Pair of tokens which do not have sameCell relationship.
with respect to a Cartesian plane centered at the for every token, we create a pair of current to-
top-left corner of the document page. We used an ken with every token, which is located within an
off-the-shelf library to get the content in a PDF imaginary rectangular window around the current
document inside a given region, that, along with token. The size of this imaginary rectangular win-
the coordinates of characters (x,y), provides the dow will help us determine the number of pairs of
font style for every character in the document. The tokens to generate.
characters along with their coordinates and font Training sample is a vector of all the features of
styles are further aggregated into tokens, by using a pair of tokens as denoted by 1.
the white-space and new-line characters as delim-
V = [W1 Fi , W2 Fi , sameCell] (1)
iters.
With the help of table location and the loca- Where, W1 Fi are n features of first token, W2 Fi
tion of individual tokens in the document, only are n features of second token, and sameCell is the
those tokens which are within the given table loca- target class indicating True if the two tokens be-
tion are collected by comparing their correspond- long to the same cell, False otherwise (see Figure
ing coordinate values. Specifically, all the to- 1).
kens, whose x coordinate is between startX and The target class, sameCell is captured using Da-
endX and whose y coordinate is between startY toin’s WYSIWYG annotation tool, that allows to
and endY are collected as table text tokens. select a sequence of words on the PDF document
and tag those words as a table cell. The training
After collecting all the tokens from table text data is generated using the annotated PDF doc-
using the table coordinates, we generate the train- uments and the target label sameCell is assigned
ing data for the binary relationship classification. accordingly for all the pair of words.
Training data requires a pair of tokens and a tar-
get label indicating whether or not those two to- Cell Relationship Features
kens belong to the same cell or not. Once we have For each token in the table, we generate a set of lo-
a list of all the tokens that are part of the table, cational and visual features. Use of semantic fea-
4
tures of tabular data along with the mentioned fea- neighbor or right neighbor. Similarly, we identify
tures of this work can be one of the future works whether that token is nearer to the top neighbor or
with the intention of improving the accuracy of the bottom neighbor (see Figure 1). We have used the
system. absence of neighborhood tokens as a set of four
We group the features used in the relation learn- categorical features as well, indicating whether or
ing task as four categories as below. not a given token has left, right, top and bottom
Location and Tile features(LTF). The absolute neighbor token.
location of a token is important evidence to in- Type and Style features(TSF). Another signifi-
dicate that it is indeed part of the table. Along cant visual clue used by humans in determining
with the absolute location, a more generalized po- whether two words belong to the same cell or not
sitional information of tokens relative to the docu- is the content and the styles used in the words. A
ment makes the contextualization and localization binary feature representing whether a token is a
of tokens easier for a reader. number or not was used to capture the data simi-
larity within a row or a column. For every pair of
To capture these relative location and distance
tokens, the comparison of font size and bold styles
information, for each token in the table, we con-
are used to indicate whether the two tokens have a
sider (x,y) of starting of the token, and (x,y) of
similar font style or not. Use of semantic features
the ending of the token in documents (see Figure
of the content of words could be another important
1). In order to incorporate contextual information
clue in differentiating the words into cells.
about a token, we split the entire document page
into an imaginary matrix of tiles of size (n X m) We find that Neighborhood, Clustering and
and for each tile, we assign a tile number. For Alignment features play a critical role in distin-
each table token, based on its coordinates we find guishing the tokens that do not belong to the same
in which tile the token is located, and we include cell. All of the feature generation techniques are
its tile number, the row, and columns of the tile as based on the coordinates of each of the tokens and
features of that token (see Figure 1). the coordinates of the table itself. The number of
tiles and the number of neighboring tokens are the
Neighborhood features(NF). The position of the
parameters which can be tuned to achieve better
surrounding tokens of a given token indicates the
table structure recognition accuracy.
relative position of a token to its neighbors and
captures the empty spatial distance around a to- 3.3 Relation Classification Using Multilayer
ken. For a given table token, we find a list of n Feedforward Neural Network
nearest tokens in all the four directions, left, right,
We have used a Multilayer Feedforward Neural
top and bottom based on the neighboring tokens’
Network to model the binary relationship classifier
spatial distances with respect to the current token
in the experiments. In order to learn the complex
(see Figure 1). The horizontal and vertical rela-
patterns that exist in the table layouts, and gen-
tive distances between these neighboring tokens
eralize these patterns we decided that deep learn-
are used as features. The Location and Tile fea-
ing is the right tool. Working at token-level, we
tures of neighboring tokens are also included as
have huge training data as well and deep neural
part of the given token’s feature set. This feature
networks seemed a right candidate for the task.
ensures that there exists a chain of transitive de-
The generated training data is fed into the Mul-
pendency connecting all the tokens of a cell.
tilayer Feedforward Neural Network that uses relu
Clustering and Alignment features(CAF). A hu- activation function in the hidden layers and a sig-
man reader makes use of the relative closeness moid activation function in the output layer. The
and horizontal and vertical alignments of a given models were trained using Adam optimizer and Bi-
word, especially when a table is not completely nary cross-entropy loss function as defined in 2
bordered, to decide which cell the word belongs (Zhang, 2019).
to. The proximity of a pair of tokens and the pres-
ence of neighboring tokens for each token in four Loss = − [y log(p) + (1 − y) log(1 − p)] (2)
different directions captures the information about
the relative closeness. Where y is a binary indicator of correct predic-
Among all the neighborhood tokens, we iden- tion of a sample, p is the predicted probability for
tify whether a given token is nearer to the left a training sample.
5
Feature Parameter Value
Number of left, right, top and bottom neighbor tokens 1
Window size for token pair generation 30 x 30 pixels
Number of tiles 20 tile rows x 20 tile columns
The input feature vector of N dimension is fed number of tokens that are in the same cell will be
into the network and the sigmoid output value is less than the number of tokens that are not in the
decoded as binary classes, 0 indicating that the two same cell.
tokens do not belong to the same cell, 1 indicating Measuring how many predicted cells are actual
that the two tokens belong to the same cell. In cells in a given table, would be a more explana-
our experiments, the Multilayer Feedforward neu- tory metric for evaluation. However, if one to-
ral network has been built using Keras backed by ken among all the tokens of a cell is wrongly pre-
Tensorflow, for quick experimentation and devel- dicted by the relationship classification model as
opment. belonging to a different cell, then measuring the
correctness of this prediction at a cell-level would
4 Experiments and Results be challenging. So we decided to use the accu-
4.1 Dataset and Evaluation Metrics racy of the binary classification model itself as our
evaluation metric. This token-level metric is sim-
Due to the lack of publicly available datasets that pler and straightforward.
suit our methodology, we prepared the training
data on our own. The dataset used for the experi- 4.2 Hyperparameters
ments contains a total of 250 PDF documents, hav- We have experimented with the hyperparameters
ing one table per document. We ensured that the of the neural network architecture itself. Table 3
tables present in our dataset represent the possi- defines the set of hyperparameters used in our ex-
ble diverse type of tables that are most commonly periments. In terms of the number of weights, Set-
used. Our dataset has tables with and without bor- 1 is a simpler network with fewer weights to learn
ders, with and without column headings, with col- and Set-2 is a more complex network.
umn and row spans, with all types of text align-
ments, varying line, and word spacing, and font Hyperparameter Set-1 Set-2
styles. All the PDF documents were annotated us- Number of layers 4 5
ing Datoin’s WYSIWYG annotation tool. Number of Epochs 200 300
Using the parameters listed in Table 1, we cre- Batch size 300 100
ated approximately 0.3 million training samples Learning Rate 0.001 0.001
from all the tokens of 250 tables, containing 83
different features. Training samples are split by a Table 3: Hyperparameters used in Multilayer
9 to 1 ratio for training and testing, keeping ap- Feedforward Neural Network
proximately 30,000 samples for testing.
The distribution of target labels in our training Table 4: Feature sets used in the experiments
and testing dataset is shown in Table 2. The imbal-
ance in the distribution of classes makes sense be- It is important to note here that a smaller batch
cause for every token in the table, within an imag- size and a higher number of epochs do increase the
inary rectangular window around that token, the F1-Scores and help the model to learn more com-
6
Feature Set Class Precision Recall F1-Score
LTF True 77.89% 91.33% 83.51%
False 95.13% 85.91% 90.29%
LTF & NF CAF True 93.65% 89.50% 91.07%
False 94.07% 96.10% 95.73%
LTF, NF CAF & TSF True 94.10% 90.85% 92.62%
False 97.56% 98.27% 98.15%
plex patterns in the data, at the cost of increased that of True class. One possible reason for this
training time. could be, the tokens that are not likely to be in the
same cell will clearly have a distinguishable set of
4.3 Experiments locational and neighboring features. It is clear that
We have experimented with many combinations of the recall of True class is causing the F1-Score to
feature sets and feature generation parameters and be low. Our training dataset has comparably fewer
selected the best three sets of features, as listed in training samples for the True class this could be a
Table 4. The Neighborhood features and Cluster- possible reason for the low recall scores.
ing and Alignment features are combined into one Manual verification of individual table cells
set because both the features measure the togeth- with the prediction of the relation classifier shows
erness of two given table tokens. that the model is able to generalize the cell recog-
In each of these experiments, we have further nition task across a variety of table cells. The
experimented with 2 sets of neural network hyper- cell relationships are identified accurately irre-
parameters listed in Table 3. spective of the presence of borders lines, column,
and row spans and text alignments. The rela-
4.4 Results and Discussion tionship among the tokens of a table is learned
All the experiments with the two sets of neural by the model based on Neighborhood, Clustering
network hyperparameters listed in Table 3, indi- and Alignment features of the tokens. However,
cated that Set-2 outperforms the Set-1. So, we for a few tokens where the neighborhood features
have listed only the results of experiments carried do not have a clear separation with tokens from
out using hyperparameters Set-2 in Table 5. nearby cells, the model combines the tokens from
Multiple experiments indicated that neighbor- adjacent cells, producing wrong predictions. Be-
hood features have sufficient information to cap- cause of the absence of visual separation among
ture the table structure and the use of visual clues the tokens of two closely aligned cells, the model
does increase accuracy. However, one experiment predicts those multiple cells as a single cell.
showed that increasing the number of neighboring
5 Conclusion and Future Work
tokens for each token, reduces the Recall measure
of True class. Increased number of False classes By applying the idea of relation extraction in ta-
could be a possible explanation for this behavior. ble structure recognition task, we have shown the
Also, increasing the number of hidden layers or possibility of high accuracy information extraction
hidden units of the network did not improve the in unstructured documents. Table structure recog-
accuracy further. nition as relation extraction task is a novel ap-
The model achieved an overall accuracy of proach in table extraction process and to the best
97.95% on the test set after training the network of our knowledge has never been explored. We
for about an hour. Clearly, the model is predicting have taken the first step towards this direction and
the cell-relationships on unseen token pairs with have proved that a bottom-up approach of cell re-
very high accuracy. A set of 20 documents con- lationship extraction is the right way towards ta-
taining a variety of tables, which are not part of ble structure recognition task. We have compared
training documents, are considered as a validation three sets of features and showcased the signifi-
set. cance of cognitive features in our experiments.
The F1-Score of False class is much better than For a few of the tables, closely aligned adja-
7
cent cells are wrongly identified as one cell. In- methods from digital scientific articles. D-Lib Mag-
corporating the semantic features of the content of azine https://doi.org/10.1045/november14-klampfl.
the words, especially using Natural Language Pro- Bernhard Krüpl and Marcus Herzog. 2006. Visually
cessing, will enrich the feature vector and should guided bottom-up table detection and segmentation
help the model to do better generalizations. Ex- in web documents. In WWW.
ploring the different layout and visual features and Shantanu Kumar. 2017. A survey of deep learn-
improving the accuracy of the proposed method ing methods for relation extraction. CoRR
could be one of the possible future works. abs/1705.03645.
Building on top of cell-relationship recognition
Martha O. Perez-Arriaga, Trilce Estrada, and Soraya
work, we hope to explore the table structure ex- Abad-Mota. 2016. Tao: System for table detec-
traction further. The knowledge of table cells can tion and extraction from pdf documents. In FLAIRS
be used to build up the rest of the table structures Conference.
from bottom-up. We believe that the relation ex- S. Schreiber, S. Agne, I. Wolf, A. Dengel, and
traction methodologies apply to other document S. Ahmed. 2017. Deepdesrt: Deep learning for
understanding tasks and we hope to explore them detection and structure recognition of tables in
as well. document images. In 2017 14th IAPR Inter-
national Conference on Document Analysis and
Recognition (ICDAR). volume 01, pages 1162–
1167. https://doi.org/10.1109/ICDAR.2017.192.
References
Alexey Shigarov. 2015. Table under-
Rahul Anand, Hye-Young Paik, and Cheng Wang.
standing using a rule engine. Expert
2019. Integrating and querying similar tables
Systems with Applications 42:929937.
from pdf documents using deep learning. CoRR
https://doi.org/10.1016/j.eswa.2014.08.045.
abs/1901.04672.
Alexey Shigarov, Andrey Mikhailov, and An-
S. Clinchant, H. Djean, J. Meunier, E. M. Lang, and drey Altaev. 2016. Configurable table struc-
F. Kleber. 2018. Comparing machine learning ap- ture recognition in untagged pdf documents.
proaches for table recognition in historical register https://doi.org/10.1145/2960811.2967152.
books. In 2018 13th IAPR International Workshop
on Document Analysis Systems (DAS). pages 133– Burcu Yildiz, Katharina Kaiser, and Silvia Miksch.
138. https://doi.org/10.1109/DAS.2018.44. 2005. pdf2table: A method to extract table infor-
mation from pdf files. pages 1773–1785.
Andreiwid Correa and Pr-ola Zander. 2017. Unleash-
ing tabular content to open data: A survey on pdf Jiawei Zhang. 2019. Gradient descent based optimiza-
table extraction methods and tools. pages 54–63. tion algorithms for deep learning models training.
https://doi.org/10.1145/3085228.3085278. CoRR abs/1903.03614.
Azka Gilani, Shah Rukh Qasim, Imran Malik, and
Faisal Shafait. 2017. Table detection using deep
learning. https://doi.org/10.1109/ICDAR.2017.131.
8
Identification of Good and Bad News on Twitter
9
Proceedings of Recent Advances in Natural Language Processing, pages 9–17,
Varna, Bulgaria, Sep 2–4 2019.
labels. The dataset is publicly accessible and
can be used for further research1 .
10
Category Topics Collected Annotated
Ebola 892 852
Health
Hiv 663 630
Hurricane Harvey 2,073 1,997
Natural Disaster
Hurricane Irma 795 772
Macerata oohmm 668 625
Terrorist Attack
Stockholm Attack 743 697
AGU17 652 592
Geography and Env.
Swachh Bharat 21 21
IOT 627 602
Science and Edu.
Nintendo 78 65
Total 7,212 6,853
Table 1: Categories, their topics, and distributions for the dataset generation.
we retrieve the examples from Twitter using its We ask three annotators5 to classify the selected
API2 . Next, we discard non-English tweets and examples into good and bad news. We also al-
re-tweets. We also remove duplicates based on lowed a third category cannot say. We computed
lower-cased first four words of tweets keeping Fleiss’ kappa Fleiss (1971) on the trial dataset for
only the first one. Thereafter, we filter only those the three annotators. The value is 0.605 which in-
tweets which can be regarded as news by using an dicates rather a high agreement. We used 247 in-
in house SVM classifier (Aggarwal, 2019). This stances agreed by all the three annotators as test
classifier is trained on tweets annotated with the questions for the crowdsourcing platform.
labels news and not news. We use this classifier to During the crowd annotation, we showed each
remove not news tweets from the annotation task3 . annotator 5 tweets per page and paid 3 US Cents
We select only tweets where the classifier predic- per tweet. For maintaining quality standards, in
tion probability is greater than or equal to 80%. In addition to the test questions, we applied a re-
Table 1, we provide information about the topics striction so that annotation could be performed
and categories as well as statistics about the col- only by people from English speaking countries.
lected tweets that will be used for annotation (col- We also made sure that each annotation was per-
umn collected). formed maximum by 7 annotators and that an an-
notator agreement of min. 70% was met. Note if
Data Annotation For data annotation, we use
the agreement of 70% was met with fewer anno-
the figure-eight crowdsourcing service4 . Before
tators then the system would not force an anno-
uploading our collected examples, we carried out a
tation to be done by 7 annotators but would fin-
round of trial annotation of 300 randomly selected
ish earlier. The system requires 7 annotators if
instances from our tweet collection corpus. The
the minimum agreement requirement is not met.
aim of the trial annotation was
We only choose instances that are annotated by at-
• to ensure the newsworthiness quality of our least 3 annotators. In addition to the good and bad
collected examples. news categories we also ask annotators to manda-
tory provide their confidence score (range between
• to create test questions to ensure the qual- 0-100%) for the label they have annotated6 . We
ity of the annotators, for the rest of the data, discarded all the tweets where we did not have
which was carried out using crowdsourcing. at least 3 annotators with each having min. 50%
confidence value. We also discarded tweets that
• to test our guidelines described in Section 3.
are annotated by less than three annotators. We
2
https://www.tweepy.org
3 5
Since we want humans to annotate tweets as good and All are post-graduate students who are fluent in English
bad news we apply this approach to filter tweets that are not and use Twitter to post information on a daily basis.
news at all and so avoid our annotators spending valuable 6
We found this strategy better than providing the option
time on annotating tweets that are not our target. cannot say and later allowed us to discard annotations where
4
https://www.figure-eight.com/ the confidence score was less than 50%.
11
use a total 7,212 tweets to annotate. After all fil- Sentiment: We use the textblob9 tool to com-
terings, we remained with 6,853 instances which pute sentiment score over each tweet. The score
were classified as good and bad news. Topic-wise varies between -1 (negative) to 1 (positive).
number of successful annotations are displayed in
the fourth column of Table 1. POS-Tag: This feature includes 36 different
pos-tags (uni-gram) and are used as binary fea-
Inter Annotator Agreement To calculate tures.
agreement between the annotators of the crowd-
Significant terms: Using tf-idf values we also
sourcing annotation results, we select the top
extract the top 300 terms (uni-gram and bi-gram,
three confident annotator labels for each sample.
300 in each case) from the training data and use
Based on this, we record an agreement of 0.614
them as binary features. Note, we extract for good
as Fleiss’ Kappa (Fleiss, 1971) score indicating a
and bad news separate uni-grams and bi-grams.
good agreement among the annotators. We also
claim stability in our annotation task because of Tweet Characteristics: This feature contains
the score similarity with that of trail annotation. tweet specific characterstics such as the number
of favorite counts, tweet replies count and number
5 Method of re-tweets.
We use the ArkTokenizer (Gimpel et al., 2011) to Finally, we also use fasttext based embedding
tokenize the tweets. In addition to tokenization, (Mikolov et al., 2018) vectors which are trained
we do lowercasing and remove digits if available on common crawl having 600 billion tokens.
in text.
5.3 Classifiers
5.2 Features We investigate 8 classifiers for our task including
Multi-Layer Perceptron (MLPC), Support Vector
We extract nine features for each tweet and di- Machine with linear (LSVC) and rbf (SVC) ker-
vide them into Structural, TF-IDF and Embed- nel, K Nearest Neighbour (KNN), Logistic Re-
dings features. gression (LR), Random Forest (RF), XGBoost
(XGB) and Decision Tree (DT). In addition, we
5.2.1 Structural features
also fine-tune BERT-base model (Devlin et al.,
Emoticons: We extract all the emoticons from 2018). Each classifier, except the BERT, has been
the training data and use them as a binary feature, trained and tested on each possible combination of
i.e. does a tweet contain a particular emoticon or the three feature types.
not.
6 Results
Interjections: We use existing list of interjec-
tions7 and use them similar to Emoticons as binary Overall results We performed a stratified 5-
feature. fold cross-validation. We evaluate each result-
ing model on a held-back development dataset
Lexicons: We use existing positive and negative containing 264 good news postings and 764 bad
lexicons8 and use them as a binary feature. news ones. The 5-fold cross validation has been
performed on the training data containing 4,332
7
https://www.vidarholen.net/contents/ bad news and 1,493 good news instances. For
interjections/
8 9
http://www.cs.uic.edu/˜liub/FBS/ https://textblob.readthedocs.io/en/
sentiment-analysis.html dev/
12
Feature set SVC XGB LSVC KNN RF MLPC DT LR
Structural .78 .78 .77 .77 .77 .63 .74 .78
Embeddings .88 .86 .87 .86 .85 .85 .72 .87
TF-IDF .86 .85 .86 .83 .84 .84 .83 .87
Structural + Embeddings .86 .85 .87 .79 .86 .86 .78 .87
Structural + TF-IDF .87 .87 .87 .80 .87 .86 .81 .87
Embeddings + TF-IDF .89 .87 .89 .87 .88 .87 .81 .89
ALL .88 .88 .89 .82 .87 .86 .82 .88
BERT-base model with its pre-trained embedding features: .92
Table 2: F1 (macro) scores of different classifiers on different feature types evaluated on the test data. Multi-
Layer Perceptron (MLPC), Support Vector Machine with linear (LSVC) and rbf (SVC) kernel, K Nearest
Neighbour (KNN), Logistic Regression (LR), Random Forest (RF), XGBoost (XGB) and Decision Tree
(DT).
0.6
the structural features of the task independently
(Figure 3). For this, we use the SVC classifier as it
is one of the best performing traditional methods. 0.5 0.48
From the figure, we see that the significant term 0.44 0.43
0.43 0.43 0.43
feature gives the best performance. The difference
0.4
to the other features is greater than the 23% F1
Em ags
s
ns
Le ics
se ons
te ents
rm
io
o
t
sT
Ch otic
rs
c
te
ct
im
xi
rje
nt
Po
nt
ca
ar
ifi
In
10
We always use the best result for every system.
13
12th layer embedding vectors (having 300 dimen-
0.84
0.85 Tweets sions) for random 100 test points (Figure 5). The
analysis shows that BERT can classify semantics
0.8
of good and bad news instances correctly even the
F1 (Macro)
LR
Be
SV
V
LS
14
Figure 5: t-SNE distribution of random 100 test points with Bert’s performance. The pie chart displays
the percentage of BERT’s misclassifications on these points.
15
John H. Boyd, Jian Hu, and Ravi Jagannathan. 2005. Vishal Kharde, Prof Sonawane, et al. 2016. Senti-
The stock market’s reaction to unemployment news: ment analysis of twitter data: a survey of techniques.
Why bad news is usually good for stocks. Journal arXiv preprint arXiv:1601.06971 .
of Finance 60(2):649–672.
Efthymios Kouloumpis, Theresa Wilson, and Jo-
PHILLIP A. BRAUN, DANIEL B. NELSON, and hanna D. Moore. 2011. Twitter sentiment analysis:
ALAIN M. SUNIER. 1995. Good news, bad The good the bad and the omg! In ICWSM.
news, volatility, and betas. The Journal of Finance
50(5):1575–1603. https://doi.org/10.1111/j.1540- Tomas Mikolov, Edouard Grave, Piotr Bojanowski,
6261.1995.tb05189.x. Christian Puhrsch, and Armand Joulin. 2018. Ad-
vances in pre-training distributed word representa-
Jennifer Conrad, Bradford Cornell, and Wayne R. tions. In Proceedings of the International Confer-
Landsman. 2002. When is bad news really bad ence on Language Resources and Evaluation (LREC
news? The Journal of Finance 57(6):2507–2532. 2018).
https://doi.org/10.1111/1540-6261.00504.
Paul R. Milgrom. 1981. Good news and bad
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and news: Representation theorems and applications.
Kristina Toutanova. 2018. BERT: pre-training The Bell Journal of Economics 12(2):380–391.
of deep bidirectional transformers for lan- http://www.jstor.org/stable/3003562.
guage understanding. CoRR abs/1810.04805.
http://arxiv.org/abs/1810.04805. Igor Mozetič, Miha Grčar, and Jasmina Smailović.
2016. Multilingual twitter sentiment classifica-
Joseph L. Fleiss. 1971. Measuring nomi- tion: The role of human annotators. PloS one
nal scale agreement among many raters. 11(5):e0155036.
Psychological Bulletin 76(5):378–382.
https://doi.org/10.1037/h0031619. Preslav Nakov, Alan Ritter, Sara Rosenthal, Fabrizio
Sebastiani, and Veselin Stoyanov. 2016. Semeval-
Kevin Gimpel, Nathan Schneider, Brendan O’Connor,
2016 task 4: Sentiment analysis in twitter. In Pro-
Dipanjan Das, Daniel Mills, Jacob Eisenstein,
ceedings of the 10th international workshop on se-
Michael Heilman, Dani Yogatama, Jeffrey Flanigan,
mantic evaluation (semeval-2016). pages 1–18.
and Noah A. Smith. 2011. Part-of-speech tagging
for twitter: Annotation, features, and experiments.
Nasir Naveed, Thomas Gottron, Jérôme Kunegis,
In Proceedings of the 49th Annual Meeting of the
and Arifah Che Alhadi. 2011. Bad news travel
Association for Computational Linguistics: Human
fast: A content-based analysis of interesting-
Language Technologies. Association for Computa-
ness on twitter. In Proceedings of the 3rd In-
tional Linguistics, Portland, Oregon, USA, pages
ternational Web Science Conference. ACM, New
42–47. https://www.aclweb.org/anthology/P11-
York, NY, USA, WebSci ’11, pages 8:1–8:7.
2008.
https://doi.org/10.1145/2527031.2527052.
Alec Go. 2009. Sentiment classification using distant
supervision. Yafeng Ren, Yue Zhang, Meishan Zhang, and
Donghong Ji. 2016. Context-sensitive twitter sen-
Lars Kai Hansen, Adam Arvidsson, Finn Aarup timent classification using neural network. In Thir-
Nielsen, Elanor Colleoni, and Michael Etter. 2011. tieth AAAI Conference on Artificial Intelligence.
Good friends, bad news - affect and virality in twit-
ter. In Communications in Computer and Informa- Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017.
tion Science, Springer Berlin Heidelberg, pages 34– Semeval-2017 task 4: Sentiment analysis in twitter.
43. https://doi.org/10.1007/978-3-642-22309-9 5. In Proceedings of the 11th international workshop
on semantic evaluation (SemEval-2017). pages 502–
Tony Harcup and Deirdre ONeill. 2017. What is 518.
news? Journalism Studies 18(12):1470–1488.
https://doi.org/10.1080/1461670X.2016.1150193. Paul Rozin and Edward B. Royzman. 2001. Negativity
bias, negativity dominance, and contagion. Person-
Wendy M. Johnston and Graham C. L. Davey. ality and Social Psychology Review 5(4):296–320.
1997. The psychological impact of negative https://doi.org/10.1207/S15327957PSPR0504 2.
tv news bulletins: The catastrophizing of per-
sonal worries. British Journal of Psychol- Aliaksei Severyn and Alessandro Moschitti. 2015.
ogy 88(1):85–91. https://doi.org/10.1111/j.2044- Unitn: Training deep convolutional neural network
8295.1997.tb02622.x. for twitter sentiment classification. In Proceedings
of the 9th international workshop on semantic eval-
Michael A. Kamins, Valerie S. Folkes, and uation (SemEval 2015). pages 464–469.
Lars Perner. 1997. Consumer responses
to rumors: Good news, bad news. Jour- Stuart N. Soroka. 2006. Good news and bad news:
nal of Consumer Psychology 6(2):165–187. Asymmetric responses to economic information.
https://doi.org/10.1207/s15327663jcp0602 03. The Journal of Politics 68(2):372–385.
16
Stuart N. Soroka. 2006. Good news and bad
news: Asymmetric responses to economic infor-
mation. The Journal of Politics 68(2):372–385.
https://doi.org/10.1111/j.1468-2508.2006.00413.x.
Laurens van der Maaten and Geoffrey E. Hinton. 2008.
Visualizing data using t-sne.
17
Bilingual Low-Resource Neural Machine Translation with
Round-Tripping: The Case of Persian-Spanish
18
Proceedings of Recent Advances in Natural Language Processing, pages 18–24,
Varna, Bulgaria, Sep 2–4 2019.
two weak translation models, and makes improve- but improvements are small. The production of
ment through the incorporation of unlabeled data synthetic parallel texts bears resemblance to data
into the training dataset. augmentation techniques, where datasets are often
In the round-tripping approach, the two transla- augmented with rotated, scaled, or otherwise dis-
tion tasks (forward and backward) together make torted variants of the (limited) training set (Rowley
a closed loop, i.e., one direction produces infor- et al., 1998).
mative feedback for training the TM for the other A similar avenue of research is self-training
direction, and vice versa. The feedback signals— (McClosky et al., 2006). The self-training ap-
which consist of the language model likelihood of proach as a bootstrapping method typically refers
the output model and the reconstruction error of to the scenario where the training dataset is en-
the original sentence—drive the process of itera- hanced with training instances with artificially
tive updates of the forward and backward TMs. produced output labels (whereas we start with neu-
For the purpose of evaluation, we apply this ap- ral network based output, i.e., the translation, and
proach to a bilingually low-resource language pair artificially produce an input). We expect that this
(Persian-Spanish) to leverage monolingual data in is more robust towards noise in MT.
a more effective way. By utilizing the round- Hoang et al. (2018) showed that the quality of
tripping approach, the monolingual data play a back translation matters and proposed an iterative
similar role to the bilingual data, effectively re- back translation, in which back translated data are
ducing the requirement for parallel data. In par- used to build better translation systems in forward
ticular, each model provides guidance to the other and backward directions. These, in turn, are used
throughout the learning process. Our results show to reback translate monolingual data. This process
that round-tripping for NMT works well in the is iterated several times.
Persian-Spanish low-resource scenario. By learn- Improving NMT with monolingual source data,
ing from monolingual data, this approach achieves following similar work on phrase-based SMT
comparable accuracy to a NMT approach trained (Schwenk, 2008), remains possible future work.
from the full bilingual data for the two translation Domain adaptation of neural networks via contin-
tasks (forward and backward). ued training has been shown to be effective for
The remainder of this paper is organized as fol- neural language models by (Ter-Sarkisov et al.,
lows; Section 2 presents the previous related work. 2015).
In Section 3, we briefly review the relevant mathe- Round-tripping has already been utilized in
matical background of NMT paradigm. Section 4 SMT by (Ahmadnia et al., 2019). In this work,
describes the round-trip training approach. The forward and backward models produce informa-
experiments and results are presented in Section 5. tive feedback to iteratively update the TMs during
Conclusions and future work are discussed in Sec- the training of the system.
tion 6.
3 Neural Machine Translation
2 Related Work NMT consists of an encoder and a decoder. Fol-
lowing (Bahdanau et al., 2015), we adopt an
The integration of monolingual data for NMT
attention-based encoder-decoder model, i.e., one
models was first proposed by (Gülçehre et al.,
that selectively focuses on sub-parts of the sen-
2015), who train monolingual LMs independently,
tence during translation. Consider a source sen-
and then integrate them during decoding through
tence X = {x1 , x2 , ..., xJ } and a target sentence
rescoring of the beam (adding the recurrent hid-
Y = {y1 , y2 , ..., yI }. The problem of translation
den state of the LM to the decoder state of the
from the source language to the target is solved
encoder-decoder network). In this approach, an
by finding the best target language sentence ŷ that
additional controller mechanism controls the mag-
maximizes the conditional probability:
nitude of the LM signal. The controller parameters
and output parameters are tuned on further paral- ŷ = arg max P (y|x) (1)
y
lel training data, but the LM parameters are fixed
during the fine tuning stage. The conditional word probabilities given the
Jean et al. (2015) also report on experiments source language sentence and preceding target lan-
with reranking of NMT output with a 5-gram LM, guage words compose the conditional probability
19
as follows: a stacking LSTM (Hochreiter and Schmidhuber,
I
1997). The goal is to derive a context vector ci
Y
P (y|x) = P (yi |y<i , x) (2) that captures relevant source information to help
i=1
predict the current target word yi .
While these models differ in how the context
where yi is the target word emitted by the decoder vector ci is derived, they share the same subse-
at step i and y<i = (y1 , y2 , ..., yi−1 ). quent steps. ci is calculated as follows:
To compose the model, both the encoder and de-
J
coder are implemented employing Recurrent neu- X
ral Networks (RNNs) (Rumelhart et al., 1986), i.e., ci = αt,j hj (8)
j=1
the encoder converts source words into a sequence
of vectors, and the decoder generates target words where hj is the annotation of source word xj and
one-by-one based on the conditional probability αt,j is a weight for the j th source vector at time
shown in the Equation (2). More specifically, the step t to generate yi :
encoder takes a sequence of source words as inputs
→
− exp (score (di , hj ))
and returns forward hidden vectors hj (1 ≤ j ≤ J) αt,j = PJ (9)
of the forward-RNN: j 0 =1 exp (score (di , hj 0 ))
→
− −−→ The score function above may be defined in a va-
hj = f (hj−1 , x) (3)
riety of ways as discussed by Luong et al. (2015).
Similarly, we obtain backward hidden vectors In this paper, we denote all the parameters to be
←
− optimized in the neural network as Θ and denote C
hj (1 ≤ j ≤ J) of the backward-RNN, in the re-
verse order. as the dataset that contains source-target sentence
←− ←−− pairs for the training phase. Hence, the learning
hj = f (hj−1 , x) (4) objective is to seek the optimal parameters Θ∗ :
20
translate the resulting sentence pairs backward ∇ΘY X E[Rtotal ] =
through the inbound-trip translation to the original E[(1 − α)∇ΘY X log P (S|Ssample ; ΘY X )] (15)
language. This step finds high-quality sentences
throughout the entirety of the generated sentence
pairs. Evaluating the results of this round-tripping Algorithm 1 shows the round-tripping procedure.
approach will provide an indication of the qual-
ity of the two TMs, and will enable their enhance- Algorithm 1 Round-trip training for NMT
ment, accordingly. This process is iterated for sev- Input: Monolingual dataset in source and target
eral rounds until both TMs converge. languages (CX and CY ), initial translation
We define KX as the number of sentences in models in outbound and inbound trips (ΘXY
CX and KY as the number of sentences in CY . and ΘY X ), language models in source and tar-
We take P (.|S; ΘXY ) and P (.|S; ΘY X ) to be two get languages (LMX and LMY ), trade-off pa-
neural TMs in which ΘXY and ΘY X are supposed rameter between 0 and 1 (α), beam search size
as their parameters. We also assume the existence (N ), learning rates (γ1,t and γ2,t ).
of two LMs for languages X and Y , trained in 1: repeat:
advance either by using other resources or using 2: t = t + 1.
the monolingual data (CX and CY ). Each LM 3: Sample sentences SX and SY from CX and
takes a sentence as input and produces a real num- CY respectively.
ber, based on target-language fluency (LM correct- 4: // Update model starting from language X.
ness) together translation accuracy (TM correct- Set S = SX .
ness). This number represents the confidence of 5: // Generate top-N translations using ΘXY .
the translation quality of the sentence in its own Generate sentences Ssample,1 , ..., Ssample,N .
language. 6: for n = 1, ..., N do
We start with a sentence in CX and denote 7: // Set LM score for nth sampled sentence.
Ssample as a translation output sample. This step R1,n = LMY (Ssample,n ).
has a score as follows: 8: // Set TM score for nth sampled sentence.
R2,n = logP (S|Ssample,n ; ΘY X ).
R1 = LMY (Ssample ) (11) 9: // Set total score of nth sampled sentence.
The R1 score indicates the well-formedness of the Rn = αR1,n + (1 − α)R2,n .
output sentence in language Y . 10: end for
Given the translation output Ssample , we em- 11: // SDG computing for ΘXY .
P
ploy the log probability value of s recovered from ∇ΘXY Ê [Rtotal ] = N1 N n=1
the Ssample as the score of the construction: [Rn ∇ΘXY log P (Ssample,n |S; ΘXY )].
12: // SDG computing for ΘY X .
P
R2 = log P (S|Ssample ; ΘY X ) (12) ∇ΘY X Ê [Rtotal ] = N1 N n=1
[(1 − α)∇ΘY X log P (S|Ssample,n ; ΘY X )].
We then adopt the LM score and construction
13: // Model update.
score as the total reward score:
ΘXY ← ΘXY + γ1,t ∇ΘXY Ê[Rtotal ].
Rtotal = αR1 + (1 − α)R2 (13) 14: // Model update.
ΘY X ← ΘY X + γ2,t ∇ΘY X Ê[Rtotal ].
where α is an input hyper-parameter. 15: // Update model starting from language Y .
The total reward score is considered a function Set S = SY .
of S, Ssample , ΘXY and ΘY X . To maximize this 16: Go through lines 5 − 14 symmetrically.
score, we optimize the parameters in the TMs uti- 17: until convergence.
lizing Stochastic Gradient Descent (SGD) (Sutton
et al., 2000). According to the forward TM, we
sample the ssample and then compute the gradi- To achieve reasonable translations we use beam
ent of the expected score (E[Rtotal ]), where E is search. We generate N-best sample translations
taken from Ssample : and use the averaged value on the beam search re-
sults to estimate the true gradient value.1
∇ΘXY E[Rtotal ] =
E[Rtotal ∇ΘXY log P (Ssample |S; ΘXY )] (14) 1
We used beam sizes 500 and 1000.
21
5 Experiments and Results We trained an RNN-based LM (Mikolov et al.,
2010) for each language using its corresponding
We apply the round-trip training approach to bilin-
monolingual corpus. The LM was then fixed and
gual Persian-Spanish translation, and evaluate the
the log-likelihood of a received message was uti-
results. We used the Persian-Spanish small bilin-
lized for scoring the TM.
gual corpora from the Tanzil corpus (Tiedemann,
To start the round-trip training approach, the
2012),2 which contains about 50K parallel sen-
systems are initialized using warm-start TMs
tence pairs. We also used 5K and 10K parallel sen-
trained from initial small bilingual data. The goal
tences extracted from the OpenSubtitles2018 col-
is to see whether the round-tripping augments the
lection (Tiedemann, 2012),3 as the validation and
baseline accuracy. We retrain the baseline systems
test datasets, respectively. Finally, we utilized 70K
by enlarging the initial small bilingual corpus: we
parallel sentences from the KDE4 corpus (Tiede-
add the optimized generated bilingual sentences to
mann, 2012),4 as the monolingual data.
the initial parallel text. The new enlarged transla-
We implemented the DyNet-based model ar-
tion system contains both the initial and optimized
chitecture (Mi et al., 2016) on top of Mantis
generated bilingual text. For each translation task,
(Cohn et al., 2016) which is an implementation of
we train the round-trip training approach.
the attention sequence-to-sequence (Seq-to-Seq)
We employ Bilingual Evaluation Understudy
NMT. For each language, we constructed a vo-
(BLEU) (Papineni et al., 2001) (using multi-
cabulary with the most common 50K words in
bleu.perl script from Moses) as the evaluation
the parallel corpora, and OOV words were re-
metric. BLEU is calculated for individual trans-
place with a special token <UNK>. For mono-
lated segments by comparing them with a data set
lingual corpora, sentences containing at least one
of reference translations. The scores of each seg-
OOV word were removed. Additionally, sentences
ment, ranging between 0 and 100, are averaged
with more than 80 words were removed from the
over the entire evaluation dataset to yield an es-
training set.5 The encoders and decoders make
timate of the overall translation quality (higher is
use of Long Short-Term Memory (LSTM) with
better).
500 embedding dimensions, 500 hidden dimen-
The baseline systems for Persian-Spanish are
sions. For training, we used the SGD algorithm
first trained, while our round-trip method conducts
as the optimizer. The batch size was set as 64
joint training. We summarize the overall perfor-
with 20 batches pre-fetched and sorted by sentence
mances in Table 1:
lengths.
Below we compare the system based on NMT systems Pe-Es Es-Pe
the optimized round-trip training (round-tripping) baseline 31.12 29.56
through two translation systems; the first one is the self-train 29.29 27.36
standard NMT system (baseline), and the second round-trip 34.91 33.43
one is the system that generates pseudo bilingual
sentence pairs from monolingual corpora to assist Table 1: BLEU scores for Persian-Spanish trans-
the training step (self-training). For the pseudo- lation task and vice-versa.
NMT we used the trained NMT model to generate
pseudo bilingual sentence pairs from monolingual As seen in Table 1, the round-tripping systems
text, removed sentences with more than 80 words outperform the others in both translation direc-
(as above), merged the generated data with the tions. In Persian to Spanish translation, the round-
original parallel training data, and then trained the tripping system outperforms the baseline by about
model for testing. Each of the translation systems 3.87 BLEU points and also outperforms the self-
was trained on a single GPU until their perfor- training system by about 6.07 BLEU points. In the
mances stopped improving on the validation set. back translation from Spanish to Persian, the im-
This approach required an LM for each language. provement of the round-tripping outperforms both
2
http://opus.nlpl.eu/Tanzil.php the baseline and self-training by about 3.79 and
3
http://opus.nlpl.eu/OpenSubtitles-v2018.php 5.62 BLEU points, respectively.
4
http://opus.nlpl.eu/KDE4-v2.php These results demonstrate the effectiveness of
5
The average sentence length is 47; an upper bound of 80
ensured exclusion of non-sentential and other spurious mate- the round-trip training approach. The baseline sys-
rial. tems outperform the self-training ones in all cases
22
because of the noise in the generated bilingual sen- ation. Thus, a possible future direction would
tences used by self-training. Upon further exami- be to design and test the round-tripping approach
nation, this result might have been expected given for more tasks beyond MT. We note that round-
that the aim of round-trip training is to optimize tripping is not restricted to two tasks only. Indeed,
the generated bilingual sentences by selecting the the key idea is to form a closed loop so feedback
high-quality sentences to get better performance signals are extracted by comparing the original in-
over self-training systems. When the size of bilin- put data with the final output data. Therefore, if
gual corpus is small, the round-tripping makes a more than two associated tasks form a closed loop,
larger improvement. This outcome is an indication this approach can applied in each task for improve-
that round-trip training approach makes effective ment of the overall model, even in the face of un-
use of monolingual data. labeled data.
Table 2 shows the performance of the base-
line alongside of the enlarged translation systems, Acknowledgments
where the latter leverages the training text of the We would like to express our sincere gratitude
baseline and the round-tripping systems as well. to Dr. Michael W. Mislove (Tulane University,
USA), Dr. Javier Serrano (Universitat Autònoma
NMT systems Pe-Es Es-Pe
de Barcelona, Spain) and Dr. Gholamreza Haffari
baseline 31.12 29.56
(Monash University, Australia) for all their sup-
enlarged 34.21 33.03
port. We also would like to acknowledge the fi-
nancial support received from the School of Sci-
Table 2: BLEU scores comparing the baseline and
ence and Engineering, Tulane University (USA).
enlarged NMT systems for Persian-Spanish trans-
lation task and vice-versa.
References
As seen in Table 2, the BLEU scores of the en-
larged NMT systems are better than the baseline Benyamin Ahmadnia, Gholamreza Haffari, and Javier
Serrano. 2018. Statistical machine translation
ones in both translation directions. The enlarged for bilingually low-resource scenarios: A round-
system in the Persian-Spanish direction outper- tripping approach. In Proceedings of the 3rd IEEE
forms the baseline by about 3.47 BLEU points, International Conference on Machine Learning and
and outperforms the baseline by about 3.09 BLEU Natural Language Processing. pages 261–265.
points in the back translation. The improvements Benyamin Ahmadnia, Gholamreza Haffari, and Javier
show that the optimized round-trip training system Serrano. 2019. Round-trip training approach for
is promising for tackling the training data scarcity bilingually low-resource statistical machine transla-
tion systems. International Journal of Artificial In-
and it also helps to enhance translation quality. telligence 17(1):167–185.
6 Conclusions and Future Work Benyamin Ahmadnia and Javier Serrano. 2017. Em-
ploying pivot language technique through statisti-
In this paper, we applied a round-tripping ap- cal and neural machine translation frameworks: The
proach based on a retraining scenario to tackle case of under-resourced persian-spanish language
training data scarcity in NMT systems. An excit- pair. International Journal on Natural Language
Computing 6(5):37–47.
ing finding of this work is that it is possible to learn
translations directly from monolingual data of the Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben-
two languages. We employed a low-resource lan- gio. 2015. Neural machine translation by jointly
guage pair and verified the hypothesis that, re- learning to align and translate. In Proceedings of
the International Conference on Learning Represen-
gardless of the amount of training resources, this tations.
approach outperforms the baseline. The results
demonstrate that round-trip training is promising Thorsten Brants, Ashok C. Popat, Peng Xu, Franz J.
Och, and Jeffrey Dean. 2007. Large language mod-
and better utilizes the monolingual data. els in machine translation. In Proceedings of the
Many Artificial Intelligence (AI) tasks are natu- Joint Conference on Empirical Methods in Natural
rally in dual form. Some examples are: (1) speech Language Processing and Computational Natural
recognition paired with text-to-speech; (2) image Language Learning. pages 858–867.
captioning paired with image generation; and (3) Trevor Cohn, Cong Duy Vu Huang, Ekaterina Vy-
question answering paired with question gener- molova, Kaisheng Yao, Chris Dyer, and Gholamreza
23
Haffari. 2016. Incorporating structural alignment bi- Henry A. Rowley, Shumeet Baluja, and Takeo Kanade.
ases into an attentional neural translation model. In 1998. Neural network-based face detection. IEEE
Proceedings of the Conference of the North Amer- Trans. Pattern Anal. Mach. Intell. 20(1):23–38.
ican Chapter of the Association for Computational
Linguistics Human Language Technologies. pages David E. Rumelhart, Geoffrey E. Hinton, and Ronald J.
876–885. Williams. 1986. Learning representations by back-
propagating errors. Nature 323:533–536.
Çaglar Gülçehre, Orhan Firat, Kelvin Xu, Kyunghyun
Cho, Loı̈c Barrault, Huei-Chi Lin, Fethi Bougares, Holger Schwenk. 2008. Investigations on large-scale
Holger Schwenk, and Yoshua Bengio. 2015. On us- lightly-supervised training for statistical machine
ing monolingual corpora in neural machine transla- translation. In Proceedings of IWSLT. pages 182–
tion. CoRR abs/1503.03535. 189.
Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, Rico Sennrich, Barry Haddow, and Alexandra Birch.
Tie-Yan Liu, and Wei-Ying Ma. 2016. Dual learn- 2016. Improving neural machine translation mod-
ing for machine translation. In Proceedings of the els with monolingual data. In Proceedings of the
30th Conference on Neural Information Processing 54th Annual Meeting of Association for Computa-
Systems. tional Linguistics.
Vu Cong Duy Hoang, Philipp Koehn, Gholamreza Richard S. Sutton, David A. Mcallester, Satinder P.
Haffari, and Trevor Cohn. 2018. Iterative back- Singh, and Yishay Mansour. 2000. Policy gradient
translation for neural machine translation. In Pro- methods for reinforcement learning with function
ceedings of the 2nd Workshop on Neural Machine approximation. In Advances in Neural Information
Translation and Generation. pages 18–24. Processing Systems. volume 12, pages 1057–1063.
Aram Ter-Sarkisov, Holger Schwenk, Loı̈c Barrault,
Sepp Hochreiter and Jürgen Schmidhuber. 1997.
and Fethi Bougares. 2015. Incremental adaptation
Long short-term memory. Neural computation
strategies for neural network language models. In
9(8):1735–1780.
Proceedings of the 3rd Workshop on Continuous
Sébastien Jean, Kyunghyun Cho, Roland Memisevic, Vector Space Models and their Compositionality.
and Yoshua Bengio. 2015. On using very large tar- pages 48–56.
get vocabulary for neural machine translation. arXiv
Jø̈rg Tiedemann. 2012. Parallel data, tools and inter-
preprint arXiv:1412.2007 .
faces in opus. In Proceedings of the 8th Interna-
Thang Luong, Ilya Sutskever, Quoc Le, Oriol Vinyals, tional Conference on Language Resources and Eval-
and Wojciech Zaremba. 2015. Addressing the rare uation.
word problem in neural machine translation. In Pro- Nicola Ueffing, Gholamreza Haffari, and Anoop
ceedings of the 53rd Annual Meeting of the Associ- Sarkar. 2008. On using monolingual corpora in sta-
ation for Computational Linguistics and the 7th In- tistical machine translation. Journal of Machine
ternational Joint Conference on Natural Language Translation .
Processing. pages 11–19.
24
Enhancing Phrase-Based Statistical Machine Translation by Learning
Phrase Representations Using Long Short-Term Memory Network
25
Proceedings of Recent Advances in Natural Language Processing, pages 25–32,
Varna, Bulgaria, Sep 2–4 2019.
Recurrent Neural Networks (RNNs) (Rumel- LSTM-based encoder-decoder, thus yielding im-
hart et al., 1986) are a class of artificial neural net- provements in quality. Sentences with the highest
work that has recently resurfaced in the field of scores are selected as the translation output.
MT (Schwenk, 2012). Unlike feed-forward net- The rest of this paper is organized as follows:
works, RNNs leverage recurrent connections that Section 2 discusses the previous related work.
enable the network to refer to prior states and, Sections 3 and 4 describe our PBSMT framework
thus, to process arbitrary sequences of input. The and LSTM encoder-decoder integration, respec-
cornerstone of RNNs is their ability to connect tively. Section 5 presents the key elements of our
previous information to the present task. For ex- approach, bringing together PBSMT and LSTM
ample, given a LM that predicts the next word encoder-decoder for phrase scoring. The experi-
based on previous words, no further context is mental results are covered in Section 6. Finally,
needed to predict the last word in the clouds are Section 7 presents conclusions and future work.
in the sky.
When the gap between the relevant information 2 Related Work
and the place that it is needed is small, RNNs learn Recently, various neural network models have
the next word from past information. But there are been applied to MT. However, few approaches
cases where more context is needed, e.g., in the have made effective use of neural networks to en-
prediction of the last word in I grew up in Spain hance the translation quality of SMT.
and I speak fluent Spanish. The word Spain sug- Sundermeyer et al. (2014) designed a neural TM
gests that the last word is probably the name of a that uses LSTM-based RNNs and Bidirectional
language, but to narrow down that language, ac- RNNs, wherein the target word is conditioned not
cess to a larger context is needed. It is entirely only on the history but also on the future source
possible for the gap between the relevant informa- context. The result was a fully formed source sen-
tion and the point where it is needed to become in- tence for predicting target words.
definitely large. Unfortunately, as that gap grows,
Feed-forward neural LMs, first proposed by
RNNs are increasingly unable to learn to connect
Bengio et al. (2003), were a breakthrough in lan-
the information.
guage modeling. Mikolov et al. (2011) proposed
Long Short-Term Memory (LSTM) networks the use of recurrent neural network in language
(Hochreiter and Schmidhuber, 1997) are an ex- modeling, thus enabling a much longer context
tension of RNNs, capable of learning such long- history for predicting the next word. Experimen-
term dependencies. RNNs are a chain of repeat- tal results showed that the RNN-based LM signifi-
ing modules of neural network and, in their sim- cantly outperforms the standard feed-forward LM.
plest form, the repeating module has a single layer. Schwenk (2012) proposed a feed-forward neu-
LSTMs also have this chain-like structure, but the ral network to score phrase pairs. They em-
repeating modules have four interacting layers. ployed a feed-forward neural network with fixed-
This paper presents a PBSMT model based on size phrasal inputs consisting of seven words, and
the Moses decoder (Koehn et al., 2007) with a LM with zero padding for shorter phrases. The sys-
that is enriched by an external dataset. Scoring of tem also had fixed-size phrasal output consisting
the phrase table generated by Moses is achieved of seven words. Similarly, Devlin et al. (2014)
through a LSTM encoder-decoder, and the result utilized a feed-forward neural network to gener-
is then evaluated in an English-to-Spanish trans- ate translations, but they simultaneously predicted
lation task. Specifically, the model is trained to one word in a target phrase. The use of feed-
learn the translation probabilities between English forward neural networks demands the use of fixed-
phrases and their corresponding Spanish ones. size phrases to work properly.
The trained model is then used as a part of a Zou et al. (2013) also proposed bilingual learn-
classical PBSMT system, with each phrase pair ing of word and phrase embeddings, which were
scored in the phrase table. Our evaluation proves used to compute the distance between phrase
that this approach enhances the translation perfor- pairs. The result was an additional annotation to
mance. Although Moses itself is able to score score the phrase pairs of an SMT system.
phrases as a part of the coding process, our ap- Chandar et al. (2014) trained a feed-forward
proach includes the scoring of phrases using the neural networks to learn the mapping of an in-
26
put phrase to its corresponding output phrase us- correspond to a logarithmic scaling of the proba-
ing a bag-of-words approach. This is closely re- bilities of each model. The translation process in-
lated to the model proposed by Schwenk (2012), volves segmenting the source sentence into source
except that their input representation of a phrase phrases x, each of which is translated into a tar-
was a Bag-Of-Words (BOW). A similar encoder- get phrase y, and reordering these target phrases to
decoder approach that used two RNNs was pro- yield the target sentence ŷ. This model is consid-
posed by Socher et al. (2011), but their model was ered superior in comparison to the noisy-channel
restricted to a monolingual setting. model because of the ability to adjust the impor-
More recently, an encoder-decoder model us- tance of individual features, thus controlling each
ing an RNN was proposed by Auli et al. (2013), feature’s influence on the overall output.
where the decoder was conditioned on a represen- In the PBSMT model, the TM is factored into
tation of either a source sentence or a source con- the translation probabilities of matching phrases
text. Kalchbrenner and Blunsom (2013) proposed in the source and target sentences (Ahmadnia
a similar model that uses the concept of an encoder and Serrano, 2015). These are considered ad-
and decoder. They used an n-gram LM for the en- ditional features in the log-linear model and are
coder part and a combination of inverse LM and weighted accordingly to maximize the perfor-
an RNN for the decoder part. The evaluation of mance as measured by Bilingual Evaluation Un-
their model was based on rescoring the K-best list derstudy (BLEU) (Papineni et al., 2001). The neu-
of the phrases from the SMT phrase table. ral LM Bengio et al. (2003) has become a commu-
nity standard for SMT system development, i.e.,
3 From SMT to PBMST neural networks have been used to rescore trans-
Our enhancement to SMT takes a noisy chan- lation hypotheses (k-best lists). Recently, how-
nel model as a starting point, where translation is ever, there has been an emerging interest in train-
modeled by decoding a source text, thereby elim- ing neural networks to score the translated phrase
inating the noise (e.g., adjusting lexical and syn- pairs using a source-sentence representation as an
tactic divergences) to uncover the intended trans- additional input (Zou et al., 2013). We adopt this
lation. However, as in our prior work (Ahmad- approach for our own PBSMT enrichment, as fur-
nia et al., 2017), we adopt a more general, log- ther detailed below.
linear variant to accommodate an unlimited num-
4 Integration of LSTM Encoder-Decoder
ber of features and to provide a more general
framework for controlling each feature’s influence Following Ahmadnia et al. (2018), we enhance
on the overall output. Standard probabilities are NMT performance by estimating the conditional
scaled to their logarithmic counterparts that are probability P (yiI |xJj ) where (xj , ..., xJ ) is a
then added together, rather than multiplying, fol- source sequence and (yi , ..., yI ) is its correspond-
lowing standard logarithmic rules. The log-linear ing target sequence whose length I may dif-
model is derived via direct modelling of the poste- fer from J. The LSTM computes this con-
rior probability P (y1I |xJ1 ): ditional probability by first obtaining the fixed-
ŷ = arg max P (y1I |xJ1 ) (1) dimensional representation ν of the source se-
y1I quence given by the last hidden state of the LSTM,
and then computing the probability of the target
where x is a source sentence.
sequence with a standard LSTM as the neural LM
The PBSMT model is an example of the noisy
formulation whose initial hidden state is set to the
channel approach, where the translation hypothe-
representation ν of the source sequence. This is
sis y is presented as the target sentence (given x as
specified as follows:
a source sentence), and the log-linear combination
of feature functions is maximized: I
Y
( M ) P (yiI |xJj ) = P (yi |ν, y1i−1 ) (3)
X
J I
ŷ = arg max λm hm (x1 , y1 ) (2) i=1
y1I m=1 where the P (yi |ν, y1i−1 ) distribution is represented
In the log-linear model of Equation (2), λm cor- with a softmax over all words in the vocabulary.
responds to the weighting coefficients of the log- To compose the model, both the encoder and
linear combination. Feature functions hm (xJ1 , y1I ) decoder are implemented employing RNNs, i.e.,
27
the encoder converts source words into a sequence where αij is a weight for the j th source vector at
of vectors, and the decoder generates target words time step t to generate yi :
one-by-one based on the conditional probability
shown in Equation (3). Specifically, the encoder exp (score (ht , hj ))
αij = PJ (10)
takes a sequence of source words as inputs and re- j 0 =1 exp (score (ht , hj 0 ))
→
−
turns forward hidden vectors hj (1 ≤ j ≤ J) of
the forward-RNN: The score function above is calculated as follows:2
→
− −−→
hj = f (hj−1 , x) (4) score(ht , hj ) = hTt hj (11)
←
−
Similarly, backward hidden vectors hj (1 ≤ j ≤ Given training data with K bilingual sen-
J) of the backward-RNN are obtained, in reverse tences, we train the model by maximizing the log-
order. likelihood as follows:
←− ←−−
hj = f (hj−1 , x) (5) K X
J
X
These forward and backward vectors are con- L(θ) = log P (yik |y<i
k
, xk ) (12)
catenated to make source vectors hj (1 ≤ j ≤ J) k=1 i=1
based on Equations (4) and (5):
h→ 5 Phrase Scoring by LSTM
− ← −i
hj = hj ; hj (6)
The centerpiece of our PBSMT enhancements is
The decoder takes source vectors as inputs and the inclusion of two stages: (1) training of a LSTM
returns target words, starting with the initial hid- encoder-decoder on a phrase table; and (2) subse-
den vector hJ .1 Target words are generated in a quent use of training output scores as additional
recurrent manner using the decoder’s hidden state features in the log-linear model when tuning the
and an output context. The conditional output SMT decoder.
probability of a target language word yi is defined During LSTM encoder-decoder training, the
as follows: frequencies of each phrase pair in the original cor-
pora are ignored. This measure is taken to re-
Pθ (yt |y<t , X) = sof tmax (Ws h̃t ) (7) duce the computational expense of randomly se-
where Ws is a parameter matrix in the output layer lecting phrase pairs from a large phrase table ac-
and h̃t is a vector: cording to the frequencies. Additionally, this mea-
sure ensures that the LSTM encoder-decoder does
h̃t = tanh(Wc [ct ; ht ]) (8) not learn to rank each phrase pair according to its
frequency of occurrence.
where Wc is a parameter matrix and ht =
Regarding the latter point, one reason behind
g(ht−1 , yt−1 ). Here, g is an RNN function that
this choice is that the existing translation proba-
takes its previous state vector and previous output
bility in the phrase table already reflects the fre-
word as input and updates its state vector. ct is a
quency of phrase pair occurrence in the origi-
context vector to retrieve source inputs in the form
nal corpus. With a fixed capacity of the LSTM
of a weighted sum of the source vectors hj , first
encoder-decoder, we need to ensure that most of
taking as input the hidden state ht at the top layer
the capacity of the model is focused on learning
of a stacking LSTM.
linguistic regularities, i.e., distinguishing between
The goal of the approach above is to derive a
translations, or learning the manifold3 of transla-
context vector ct that captures relevant source in-
tions. Once the LSTM encoder-decoder is trained,
formation, thus enabling the prediction of the cur-
we add a new score for each phrase pair to the ex-
rent target word yt . While a variety of models may
isting phrase table. This allows the new scores to
be used to derive a range of different context vec-
enter into the existing tuning algorithm with mini-
tors ct , the same subsequent steps are taken. Equa-
mal additional overhead in computation.
tion (8) defines our choice for ct :
An alternative to what is described above is the
S
X replacement of the existing phrase table with the
ct = αij hj (9)
2
j=1 The decoder puts more attention (weights) on source
vectors close to the state vector.
1 3
Concatenated source vector at the end. Region of probability concentration.
28
LSTM encoder-decoder. In such an approach, the produces the best-scored phrases as output.
problem would be recast in the form of the fol- For the training phase of SMT, we apply the fol-
lowing implementation: given a source phrase, the lowing steps:
LSTM encoder-decoder generates a list of target
phrases (Schwenk, 2012). However, in this al- • Tokenization: Insert spaces and punctuation
ternative, an expensive sampling procedure must between words.
be performed repeatedly. In our approach, the
• True-casing: Convert initial words in each
only phrase pairs that are rescored are those in the
sentence to their most probable casing, to re-
phrase table.
duce data sparsity.
6 Experiments and Results • Cleaning: Remove both long and empty sen-
Numerous large resources are available for build- tences, as they may cause misalignment is-
ing an English-Spanish SMT system, many of sues within the training pipeline.
which have become community standards, used in
The LM is built with the target language (in
translation tasks in annual workshops and confer-
our case-study, Spanish is the target language) to
ences on SMT hosted by ACL, NAACL, EACL,
ensure fluent and well-formed output. KenLM
and EMNLP (SMT 2006-2015 and WMT 2016-
(Heafield, 2011), which comes bundled with the
2019). Bilingual datasets include Europarl, News-
Moses toolkit, is used for building our LM. Also
Commentary, UN, and two crawled corpora.4 For
to train the TM, GIZA++ (Och and Ney, 2003) is
our purposes, we have trained the Spanish LM us-
used for word alignment. Finally, the phrases are
ing about 700M words of crawled newspaper ma-
extracted and scored as well. The generated phrase
terial.5
table is later used to translate test sentences that
We select a subset of 350M words for language
are compared to the results of the LSTM encoder-
modeling as well as a subset of 300M words for
decoder.
training the LSTM encoder-decoder. We use the
The following steps are applied to build the
test set newstest2011 and newstest2012 for data
NMT system with the LSTM Encoder-Decoder:
selection and weight tuning with Minimum Er-
ror rate Training (MERT) (Och, 2003), and new- • Vocabulary building: Generate vocabulary
stest2013 as our test set. Each set has more than corpus for both source and target sides.
70K words and a single reference translation.
For training the neural networks, including our • Corpus shuffle: Shuffle the vocabulary cor-
LSTM encoder-decoder, we limited the source and pus of both source and target languages.
target vocabulary to the most frequent 10K words
for both English and Spanish. This covers ap- • Dictionary building: Create dictionary by
proximately 90% of the dataset. All the Out-Of- leveraging an alignment file to replace the
Vocabulary (OOV) words were mapped to a spe- <UNK> words.
cial token (<UNK>). The files mentioned above are fed to the LSTM
The baseline PBSMT system is built on top of encoder for the training phase of the NMT system.
Moses (Koehn et al., 2007) with default settings. The number of epochs is set to 1000. LSTM cells
Moses is an SMT decoder that enables automatic are used with the Adam optimizer (Kingma and
training of TMs for any language pair using a large Ba, 2014), also 0.001 and 512 are as error rate and
parallel corpus. Once the model is trained, an ef- batch size, respectively. The validation split of 0.1
ficient search algorithm finds the highest proba- is used for the training as well. After the training
bility translation among the exponential number step, the LSTM decoder will be used to translate
of candidates. Training the Moses decoder yields given sentences.
a phrase model as well as a TM which, together, In order to analyze the improvement of the
support translation between source and target lan- performance of LSTM encoder-decoder over the
guages. Moses scores each phrase in the phrase SMT system analyzing the scores of the phrase
table with respect to a given source sentence and pairs, we did the same as (Cho et al., 2014); se-
4
These two corpora are quite noisy. lecting those phrase pairs that are scored higher
5
Word counts refer to Spanish words, after tokenization. by the LSTM encoder-decoder compared to the
29
SMT system with respect to a given source sen- compared the resulting translation quality of the
tence. The scoring of the phrases provided by the LSTM against PBSMT and a NMT baseline, and
LSTM encoder-decoder is similar to the scoring demonstrated a BLEU score increase of up to 3.14
of the phrases provided by the phrase table of the and 1.12, respectively. We also noticed that SMT
SMT system as the quality of the translation is ap- works well for long sentences while NMT works
proximately the same. well for short sentences.
We employ BLEU (Papineni et al., 2001) as the Since NMT systems usually have to apply a
evaluation metric. BLEU is calculated for individ- certain-sized vocabulary to avoid time-consuming
ual translated segments by comparing them with a training and decoding, such systems suffer from
dataset of reference translations. The scores, be- OOV issues. Furthermore, NMT lacks a mech-
tween 0 and 100 are averaged over the whole eval- anism to guarantee/control translation of all the
uation dataset to reach an estimate of the transla- source words and favors short translations, result-
tion overall quality. Table 1 shows the results on ing in fluent but inadequate translations.
both development set as well as test set. Issues outlined above are fodder for future
work. For example, the incorporation of SMT fea-
Model BLEU-dev BLEU-test tures into the NMT model within the log-linear
PBSMT 30.84 28.51 framework may be a future next step to address
NMT-baseline 31.95 30.53 the training/decoding issue above.
LSTM 33.49 31.65
Acknowledgments
Table 1: BLEU scores computed on the develop-
We would like to express our sincere gratitude to
ment and test sets using PBSMT, NMT-baseline
Dr. Michael W. Mislove (Tulane University, USA)
and LSTM systems.
for all his support. We also would like to ac-
LSTM encoder-decoder scores indicate in over- knowledge the financial support received from the
all improvement in translation performance in School of Science and Engineering, Tulane Uni-
terms of BLEU scores. As seen in Table 1, our versity (USA). In addition, we are grateful to the
LSTM encoder-decoder outperforms the NMT- RANLP anonymous reviewers for all their helpful
baseline by 1.12 BLEU points while it outper- comments and discussions.
forms the PBSMT by 3.14 BLEU points. Our
LSTM encoder-decoder is able to score a pair of References
sequences (in terms of a conditional probability)
Benyamin Ahmadnia, Parisa Kordjamshidi, and Gho-
or to generate a target sequence given a source se- lamreza Haffari. 2018. Neural machine transla-
quence. tion advised by statistical machine translation: The
We evaluated the proposed model with the SMT case of farsi-spanish bilingually low-resource sce-
task, using the LSTM encoder-decoder to score nario. In Proceedings of the 17th IEEE Interna-
tional Conference on Machine Learning and Appli-
each phrase pair in the phrase table. Qualita-
cations. pages 1209–1213.
tively, we showed that this model captures linguis-
tic regularities in the phrase pairs and also that Benyamin Ahmadnia and Javier Serrano. 2015. Hier-
the LSTM encoder-decoder proposes well-formed archical phrase-based translation model vs. classical
phrase-based translation model for spanish-english
target phrases. We also found that the LSTM statistical machine translation system. In Proceed-
encoder-decoder contribution is orthogonal to the ings of the 31st Conference of the Spanish Society
existing use of neural networks in the SMT sys- for Natural Language Processing.
tem. We conclude that further performance im- Benyamin Ahmadnia, Javier Serrano, and Gholamreza
provements are likely if we were to use the LSTM Haffari. 2017. Persian-spanish low-resource statisti-
encoder-decoder and the neural LM together. cal machine translation through english as pivot lan-
guage. In Proceedings of Recent Advances in Natu-
7 Conclusions and Future Work ral Language Processing. pages 24–30.
In this paper, we described a PBSMT implementa- Michael Auli, Michel Galley, Chris Quirk, and Geof-
frey Zweig. 2013. Joint language and translation
tion within which the phrase table is generated by modeling with recurrent neural networks. In Pro-
using an LSTM encoder-decoder, and sentences ceedings of the conference on Empirical Methods in
with the highest scores are selected as output. We Natural Language Processing. pages 1044–1054.
30
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- Philipp Koehn. 2009. Statistical Machine Translation.
gio. 2015. Neural machine translation by jointly Cambridge University Press.
learning to align and translate. In Proceedings of
the International Conference on Learning Represen- Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris
tations. Callison-Burch, Marcello Federico, Nicola Bertoldi,
Brooke Cowan, Wade Shen, Christine Moran, and
Yoshua Bengio, Réjean Ducharme, Pascal Vincent, Richard Zens. 2007. Moses: Open source toolkit for
and Christian Janvin. 2003. A neural probabilis- statistical machine translation. In Proceedings of the
tic language model. Machine Learning Research 45th annual meeting of the Association for Compu-
3(19):1137–1155. tational Linguistics. pages 177–180.
Sarath Chandar, Stanislas Lauly, Hugo Larochelle, Philipp Koehn, Franz Josef Och, and Daniel Marcu.
Mitesh Khapra, Balaraman Ravindran, Vikas 2003. Statistical phrase-based translation. In Pro-
Raykar, and Amrita Saha. 2014. An autoencoder ceedings of the Conference of the North American
approach to learning bilingual word representations. Chapter of the Association for Computational Lin-
In Proceedings of Advances in Neural Information guistics on Human Language Technology. pages 48–
Processing Systems. pages 1853–1861. 54.
Kehai Chen, Rui Wang, Masao Utiyama, and Eiichiro Daniel Marcu and William Wong. 2002. A phrase-
Sumita. 2019. Neural machine translation with re- based, joint probability model for statistical machine
ordering embeddings. In Proceedings of the 57th translation. In Proceedings of the conference on Em-
Annual Meeting of the Association for Computa- pirical Methods in Natural Language Processing.
tional Linguistics. pages 1787–1799. pages 133–139.
Kyunghyun Cho, Bart Van merrienboer, Çaglar Tomas Mikolov, Stefan Kombrink, Luks Burget, Jan
Gülçehre, Dzmitry Bahdanau, Fethi Bougares, Hol- Cernock, and Sanjeev Khudanpur. 2011. Exten-
ger Schwenk, and Yoshua Bengio. 2014. Learning sions of recurrent neural network language model.
phrase representations using rnn encoder-decoder In Proceedings of IEEE International Conference
for statistical machine translation. In Proceedings on Acoustics, Speech and Signal Processing. pages
of the conference on Empirical Methods in Natural 5528–5531.
Language Processing. pages 1724–1734.
Franz Josef Och. 2003. Minimum error rate training in
Jacob Devlin, Rabih Zbib, Zhongqiang Huang, Thomas statistical machine translation. In Proceedings of the
Lamar, Richard Schwartz, and John Makhoul. 2014. 41st Annual Meeting of the Association for Compu-
Fast and robust neural network joint models for sta- tational Linguistics. pages 160–167.
tistical machine translation. In Proceedings of the
52nd Annual Meeting of the Association for Compu- Franz Josef Och and Hermann Ney. 2003. A systematic
tational Linguistics. pages 1370–1380. comparison of various statistical alignment models.
Computational Linguistics 29(1):19–51.
Jinhua Du and Andy Way. 2017. Pre-reordering for
neural machine translation: Helpful or harmful? Kishore Papineni, Salim Roukos, Todd Ward, and Wei-
Prague Bull. Math. Linguistics 108:171–182. Jing Zhu. 2001. Bleu: A method for automatic eval-
uation of machine translation. In Proceedings of the
Kenneth Heafield. 2011. Kenlm: Faster and smaller 40th Annual Meeting on Association for Computa-
language model queries. In Proceedings of the Sixth tional Linguistics. pages 311–318.
Workshop on Statistical Machine Translation.
David E. Rumelhart, Geoffrey E. Hinton, and Ronald J.
Sepp Hochreiter and Jürgen Schmidhuber. 1997. Williams. 1986. Learning representations by back-
Long short-term memory. Neural computation propagating errors. Nature 323:533–536.
9(8):1735–1780.
Holger Schwenk. 2012. Continuous space translation
Nal Kalchbrenner and Phil Blunsom. 2013. Recur- models for phrase-based statistical machine transla-
rent continuous translation models. In Proceedings tion. In Proceedings of the 24th International Con-
of the conference on Empirical Methods in Natural ference on Computational Linguistics. pages 1071–
Language Processing. pages 1700–1709. 1080.
Shin Kanouchi, Katsuhito Sudoh, and Mamoru Ko- Richard Socher, Eric H. Huang, Jeffrey Pennington,
machi. 2016. Neural reordering model considering Andrew Y. Ng, and Christopher D. Manning. 2011.
phrase translation and word alignment for phrase- Dynamic pooling and unfolding recursive autoen-
based translation. In Proceedings of the 3rd Work- coders for paraphrase detection. In Proceedings
shop on Asian Translation 2016. pages 94–103. of Advances in Neural Information Processing Sys-
tems. pages 801–809.
Diederik P. Kingma and Jimmy Ba. 2014. Adam: A
method for stochastic optimization. arXiv preprint Martin Sundermeyer, Tamer Alkhouli, Joern Wuebker,
arXiv:1412.6980 . and Hermann Ney. 2014. Translation modeling with
31
bidirectional recurrent neural networks. In Proceed-
ings of the Conference on Empirical Methods in Nat-
ural Language Processing. pages 14–25.
Xing Wang, Zhaopeng Tu, Deyi Xiong, and Min
Zhang. 2017. Translating phrases in neural machine
translation. In Proceedings of the Conference on
Empirical Methods in Natural Language Process-
ing. pages 1421–1431.
Will Y. Zou, Richard Socher, Daniel M. Cer, and
Christopher D. Manning. 2013. Bilingual word em-
beddings for phrase-based machine translation. In
Proceedings of the conference on Empirical Meth-
ods in Natural Language Processing. pages 1393–
1398.
32
Automatic Propbank Generation for Turkish
33
Proceedings of Recent Advances in Natural Language Processing, pages 33–41,
Varna, Bulgaria, Sep 2–4 2019.
Section 5, we conclude with the results. ies constitute a base for Turkish proposition bank,
but their size is limited and construction of these
2 PropBank proposition banks consumed a lot of time.
2.1 English PropBank
3 Automatic PropBank Generation
PropBank is the bank of propositions where Studies
predicate-argument information of the corpora is
annotated and semantic roles or arguments that PropBanks are also generated automatically for
each verb can take are posited. It is constituted resource-scarce languages by using parallel cor-
on the Penn Treebank (Marcus et al., 1993) Wall pus. In this section, proposition bank studies for
Street Journal [WSJ]. The primary goal is to la- automatic generation are presented. Zhuang and
bel syntactic elements in a sentence with specific Zong (2010) proposed performing SRL on parallel
argument roles to standardize labels for the simi- corpus of different languages and merging the re-
lar arguments such as the window in John broke sult via a joint inference model can improve SRL
the window and the window broke. PropBank results for both input languages. In the study an
uses conceptual labels for arguments from Arg0 to English and Chinese parallel corpus is used. First
Arg5. Only Arg0 and Arg1 indicate the same roles each predicate is processed by monolingual SRL
across different verbs where Arg0 means agent systems separately for producing argument candi-
or causer and Arg1 is the patient or theme. The dates. After the candidates formed, a Joint Infer-
rest of the argument roles can vary across differ- ence model selects the candidate that is reasonable
ent verbs. They can be instrument, start point, end to the both languages. Also, a log-linear model is
point, beneficiary, or attribute. formulated to evaluate the consistency. This ap-
Moreover, PropBank uses ArgM’s as modifier la- proach increased F1 scores 1.52 and 1.74 respec-
bels where the role is not specific to the verb group tively for Chinese and English.
and generalizes over the corpora such as location, Van der Plas et al. (2011) presents cross-lingual
temporal, purpose, or cause etc. arguments. The semantic transfer from English to French. En-
first version of English PropBank, named as The glish syntactic-semantic annotations were trans-
Original PropBank, is constructed for only ver- ferred using word alignments to French language.
bal predicates whereas the latest version includes French semantic annotations gathered from the
all syntactic realizations of event and state se- first step were then trained with a French joint
mantics by focusing different expressions in form syntactic-semantic parser along with the French
of nouns, adjectives and multi-word expressions syntactic annotations trained separately. Joint
to represent complete event relations within and syntactic-semantic parser is used for learning the
across sentences. relation between semantic and syntactic structure
of the target language and reduces the errors aris-
2.2 PropBank Studies for Turkish ing from the first step. This approach reaches 4%
There have been different attempts to construct lower than the upper bound for predicates and 9%
Turkish PropBank in the literature. Şahin (2016a; for arguments.
2016b), Şahin and Adalı (2017) report semantic Kozhevnikov (2013) shows SRL model transfer
role annotation of arguments in the Turkish de- from one language to another can be achieved
pendency treebank. They construct PropBank by by using shared feature representation. Shared
using ITU-METU-Sabancı Treebank (IMST). In feature representation for language pairs is con-
these studies, frame files of Turkish PropBank are structed based on syntactic and lexical informa-
constructed and extended by utilizing crowdsourc- tion. Afterwards, a semantic role labeling model
ing. 20,060 semantic roles are annotated in 5,635 is trained for source language and then used for
sentences. The size of the resource is stated as a the target language. As a result SRL model of the
drawback in the study. Recently, Ak et al. (2018) target language is generated. Process only requires
construct another Turkish Proposition Bank using a source language model and parallel data to con-
translated sentences of English PropBank. So far, struct target SRL model. Approach is applied for
9,560 of 17,000 translated sentences are annotated English, French, Czech and Chinese languages.
with semantic roles. Also, framesets are created In the next study, Van der Plas (2014) improves the
for 1,330 verbs and 1,914 verb senses. These stud- labeling results with respect to the previous work
34
(Van der Plas et al., 2011) by building separate relations with verbal predicates. In the newer ver-
models for arguments and predicates. Also, prob- sions adjective and noun relations are also anno-
lems of transferring semantic annotations using tated. Since we compare projection results with
parallel corpus is examined in the paper. Token-to- manually annotated corpus (Ak et al., 2018) which
token basis annotation transfer, translation shifts, only contains verbal relations, we use the ini-
and alignment errors in the previous work is re- tial version of the English PropBank. We down-
placed with a global approach that aggregates in- loaded this dataset and imported annotations for
formation at corpus level. Instead of using En- the selected sentences. After this step 6,060 sen-
glish semantic annotations of roles and predicate tences among 9,558 were enhanced with the En-
together with French PoS tags to generate French glish annotations. Below in Figure 1, a sample
semantic annotations, English annotations of pred- sentence is presented. English annotations are in-
icates and roles used separately to generate one serted inside “englishPropbank” tags right after
predicate and one role semantic annotations sep- Turkish annotations which reside in “propbank”
arately. tags. Some of the words have only English an-
Akbik et al. propose a two stage approach (Ak- notation, because there is no word translated in
bik et al., 2015). In the first stage only filtered the Turkish sentence for this node. As an exam-
semantic annotation is projected. Since high con- ple, “their” in Figure 1 has annotations in the en-
fidence semantic labels projected, resulting target glishPropbank tag but there is no equivalent trans-
semantic labels will be high in precision and low lation in Turkish, presented as “*NONE*”, so
in recall. In the next stage, completed target lan- propbank tag does not exist. English tags have
guage sentences sampled and a classifier is trained predicate information that annotation belongs to.
to add new labels to boost recall and preserve pre- “Müşterilerinin” (customers) in the same exam-
cision. Proposed system is applied on 7 different ple has “ARG0$like 01#ARG1$think 01” in the
languages from 3 different language family. These englishPropbank tag which means there exists at
languages are Chinese, Arabic, French, German, least two words whose root is in verb form. Here
Hindi, Russian, and Spanish. the word is annotated with respect to “like” and
“think” separately. We have separated multiple
4 Methods annotations with “#” sign and in each annotation
predicate label and role is distinguished by “$”
Among the studies for Turkish proposition bank, sign. In the Turkish annotation, WordNet id of the
Ak et al. (2018) is constructed on parallel English predicate was used instead of predicate label.
- Turkish sentences from the Original English
PropBank. We have used the corpus provided in 4.1.2 Transfering Annotations to Automatic
this study to automatically generate proposition Turkish PropBank Using Parallel
bank. Sentences
After importing English annotations, it is neces-
4.1 Automatic Turkish PropBank Using sary to determine predicate(s) of the Turkish sen-
Parallel Sentence Trees tences. Morphological structures of the words are
Penn Treebank structure offers advantages for examined to detect predicate candidates. Words
building fully tagged data set in accordance with were morphologically and semantically analyzed
syntactic labels, morphological labels and parallel in translated Penn TreeBank. We have used “mor-
sentences. We used this structure to add English phologicalAnalysis” tag to check the morpholog-
PropBank labels for each word in the corpus. In ical structure of the words. In Figure 1, sample
this manner, we exploited this parallel dataset to morphological structure is displayed.
transfer English PropBank annotations to an auto- The word which has a verb root and verb ac-
matic Turkish PropBank. cording to last inflectional group is treated as the
predicate of the sentence. Once we found a word
4.1.1 English PropBank Labels suitable for these conditions, we gathered English
Original English PropBank corpus (Palmer et al., PropBank annotation. If it is also labeled as predi-
2004) is accessible through Linguistic Data Con- cate in English proposition bank, we got the predi-
sortium (LDC). This resource is the initial version cate label, e.g. like 01, to find annotations with re-
of the English PropBank and it only includes the spect to this predicate. We searched for the found
35
Figure 1: Part of a sentence tree : English PropBank annotations reside in “englishPropBank” tags.
predicate label in the annotations and transfered PropBank labels transferred from English annota-
annotations matching with the predicate label. If tions but no annotation exists in hand annotated
we could not find a predicate in Turkish sentence proposition bank. Annotations to be compared is
or the corresponding English label did not contain not valid so we did not include this set in the eval-
Predicate role annotation, we skipped to the next uation.
predicate candidate.
Transferred Untransferred
During the transfer, a mapping was needed due Correct 19,373 Not H.A. 8,837
to the difference between English and Turkish (Ak Incorrect 6,441
Undetermined 5,999 H.A. 4,129
et al., 2018) argument labeling. English PropBank
Total 31,813 Total 12,966
corpus has “-” sign in ArgM’s like ARGM-TMP
and also some of the arguments from Arg1 to Arg5 Table 1: Results of the comparison between auto-
are labeled with the prepositions such as ARG1- matic proposition bank and hand annotated (H.A.)
AT, ARG2-BY etc. We processed these differ- proposition bank.
ences and then transferred labels into the “prop-
bank” tags. After analyzing Turkish sentences
we found out some sentences have more than one When we remove undetermined 5,999 words in
predicate, so we continued to search for another the comparison; 19,373 annotations from 25,814
predicate in the sentence and ran the same proce- annotations are correct, which gives us ∼75% ac-
dure for each predicate candidate. curacy for transferred and comparable set. These
5,999 annotations may be hand-annotated and re-
4.1.3 Experiments compared for validity of the transferred annota-
Annotations gathered from the English sentence tions.
were compared with the Turkish hand-annotated In Table 2, we present occurrences of erroneous
proposition bank (Ak et al., 2018). Comparisons annotation transfers. Only top ten occurrences are
were done at the word level by checking the an- presented. Arg0-Arg1 transfers are the most oc-
notations for each corpus. Among the 6,060 sen- curred incorrect transfers 1,843 among 6,441 in-
tences enhanced with English PropBank roles, correct annotations. Second most occurred error
848 sentences did not have a predicate in Turk- is in Arg1-Arg2 labels. Errors in Arg0-Arg1 and
ish proposition bank. Therefore, in 5,212 sen- Arg1-Arg2 labels forms ∼44% of the transfer er-
tences, 44,779 word annotations were compared. rors.
31,813 annotations were transferred from English On the other hand, when we look at the all
to Turkish. Results of the comparison are pre- word results, 12,966 roles were not transferred.
sented in Table 1. 19,373 words annotated with If we take these untransferred instances as incor-
PropBank roles correctly . 6,441 annotations are rect; 19,373 annotations out of 38,780 annotation
incorrect, PropBank tags are different in both cor- are true and the accuracy drops to ∼50%. How-
pus. 5,999 annotations are undetermined, valid ever, 8,837 of untransferred annotation are not an-
36
Different Arguments # of Occurrence tree formatted English sentence to extract English
ARG0-ARG1 1,843
ARG1-ARG2 961 propbank annotations. However, since the target
ARG2-ARGMEXT 462 sentence do not have tree structure definition we
ARG1-PREDICATE 255 used other word alignment methods to determine
ARG0-ARG2 229
ARG4-ARGMEXT 226 annotation projection.
ARG1-ARGMPNC 220
ARG1-ARGMMNR 186 4.2.2 Semantic Alignment Using WordNet
ARG1-ARGMTMP 160
ARG1-ARGMLOC 148 In order to transfer annotations, first we tried to
match predicates of English sentence and Turk-
Table 2: Counts of different argument annotations ish translation. Again we utilize “morphological-
between transferred annotations and hand annota- Analysis” tags to determine predicate candidates
tions. in the phrase sentence. Words which have a verb
root and verb according to last inflectional group
is treated as the predicate candidates of the sen-
notated in the hand-annotated corpus. Only 4,129 tence. Once we found all the words ensuring these
are valid PropBank arguments. In this respect, if conditions, we gathered all English PropBank an-
we count only valid arguments for untransferred notation labels which are tagged as “Predicate” in
annotations, accuracy is ∼65%. ‘englishPropbank” tag. To align predicates in dif-
ferent languages, we tried to exploit WordNet’s
4.2 Automatic Turkish PropBank Using
(Ehsani et al., 2018) interlingual mapping capa-
Parallel Sentence Phrases
bilities. For each predicate in English sentence
In the previous method, annotation projection us- we find Turkish translation by searching English
ing parallel sentence trees is discussed. However, synset id in the WordNet. English synset id is lo-
finding such a resource in a special format is diffi- cated in englishSemantics tags as in the sample
cult especially if you are working with a resource- in Figure 1. If there exists any translation in the
scarce language. Most of the time creating a for- WordNet, we take Turkish synset id and search it
matted parallel resource like tree structured sen- in the predicate candidates found for phrase sen-
tences complicates translation procedure. In this tence. Whenever translation found, we align pred-
section, automatic generation with translated sen- icates and try to transfer annotation with respect
tences without tree structure will be examined. to aligned English label. For annotation transfer of
other arguments we again align words using Word-
4.2.1 Phrase Sentence Structure Net’s interlingual mapping. An example WordNet
For the phrase sentences, English sentences re- record is presented in Figure 3.
translated without tree structure. Prior the an- First results gathered with only WordNet map-
notation projection, linguists in the team anno- ping were very low. True annotation count is
tated phrase sentences and populated “propbank” 2,195 among 29,168 annotations tagged manually
and “shallowParse” tags so that we check the cor- which yields 7.53%. However, transferred false
rectness after the annotation transfer. 6,511 sen- annotation count is only 342. System heavily re-
tences among 9,557 phrase sentences have pred- lies on semantic annotations for both English and
icate according to hand annotations for newly Turkish words where some of the words failed to
translated sentences. However, only 5,259 sen- have semantic annotation. We look deeper into
tences have English PropBank annotation, so we dataset provided by Ak et al. (2018), 11,006 En-
take this set to transfer annotations. As you re- glish words do not have semantic annotation so we
member, the same number in the previous section failed to match these words with Turkish counter-
was 5,212. Here translation and annotation differ- parts.
ences change the processed sentence count. Some words are not annotated semantically
Tag structure of Penn Treebank is preserved such as, proper nouns, time, date, numbers, or-
to simplify morphologic and semantic analysis dinal numbers, percentiles, fractional numbers,
requirements during the annotation transfer. In number intervals, and reel numbers. Most of
Figure 2, sample phrase sentence can be seen. these words are same in Turkish translation so
Unlikely Figure 1, syntactic tags which indicate we matched English and Turkish words by string
tree structure are not included. We used original match. For example if a sentence contains proper
37
Figure 2: Part of a phrase sentence : Translated words in Turkish tags. Helper tags gives additional
information for each word.
38
(1) [The less-rigorous Senate version]ARG0 English Word Turkish Word Probability
[would]ARGM-MOD [defer]Predicate [the deductibility]ARG1 Reserve Reserve 0.72270917
for [roughly five years.]ARG2-for Reserve Rezerv 0.15328414
Reserve mevduat 0.03056293
(2) [Daha az sıkı türden bir Senato versiyonu]Özne - Subject Reserve Bankası’nın 0.02731664
[aşağı yukarı beş yıl için]Zarf Tümleci - Adverbial Clause Reserve kuruluşlarındaki 0.01375332
[düşülebilirliği]Nesne - Object [ertelerdi.]Yüklem - Predicate Reserve komisyonları 0.01375332
Reserve Bankasının 0.00611259
(3) [Daha az sıkı türden bir]NONE [Senato]ARG0 Reserve kuruluşlarında 0.00458444
[versiyonu]NONE [aşağı yukarı]ARG2 [beş]NONE Reserve komisyon 0.00458444
[yıl]ARG2 [için düşülebilirliği]NONE [ertelerdi.]PREDICATE Reserve Federe 0.00458444
alignment models offer solution to our word align- Table 4: Word alignment probabilities for English
ment problem. IBM Models are mainly used for word ”Reserve” calculated by IBM Model 2.
statistical machine translation to train a transla-
tion model and an alignment model. IBM Model 1 words tagged with “PREDICATE” tag in English
(Brown et al., 1993) is the primary word alignment sentence are stored into a map which includes
model offered by IBM. It is widely used for solv- predicate label from the “englishPropbank” tag
ing word alignments while working with parallel e.g. “like 01” and English word from the “en-
corpora. It is a generative probabilistic model that glish” tag e.g. “like”. Then we search alignments
calculates probabilities for each word alignment for each found English predicate. Here we ob-
from source sentence to target sentence. It takes a served that aligned Turkish words may not occur
corpus of paired sentences from two languages as in the phrase sentence as they found in the align-
training data. These paired sentences are possible ment table. Words may include additional suffixes,
translation of the sentences from source language so we use Finite State Machine(FSM) morpholog-
to target. With this training corpus, parameters of ical analyzer available in our NLP Toolkit of Ak
the model estimated using EM (expectation maxi- et al. (2018) to extract roots of the aligned Turkish
mization). IBM Model 2 has an additional model words. Since we have several possible morpho-
for alignment and introduce alignment distortion logical parse for each aligned word, we created
parameters. We decided to use IBM model 1 & an array for possible roots. In parallel, we found
2 to establish word alignments instead of Word- predicate candidates from the phrase sentence as
Net’s interligual mapping. We input sentence pairs we stated in the previous methods. Then we tried
and gather alignment probabilities for each En- to match aligned words and possible roots with the
glish word to Turkish equivalent. 244,024 word found predicate candidates. If there exists a predi-
pairs are taken as output where for each English cate candidate that matches with the aligned word
word, 10 most probable Turkish words are listed. or one of its roots in the array, we tagged the can-
Alignment probabilities for word “Reserve” is pre- didate as “PREDICATE” and update map as pred-
sented in Table 3 and 4 for IBM Model 1 and 2 icate label and synset id of Turkish predicate.
respectively. After finishing predicate discovery, we transfer
After gathering alignment data, we transfer an- annotations for found predicates. To do that we
notations to phrase sentences from English Prop- look for the annotations with respect to the predi-
Bank labels in the tree structured sentences. All cate labels in the map. For each record in map we
39
took the predicate label and corresponding Turk- We use these rules to tag any untagged word.
ish synset id. When we found an annotation with After applying these rules annotation transfer re-
this predicate label, first we extract the argument sult is as shown in Table 5 and 6. Results show
and try to find aligned word for the processed En- that rules applied slightly change the correct an-
glish word. For the alignment again we find the notations. For model 1 rules output much more
most probable word from the table and use FSM correct annotation than the incorrect ones whereas
morphological analyzer to extract possible roots. in model 2 the number of correct and incorrect an-
Then for each word we search Turkish sentence to notations gathered are nearly same. However, pre-
match words with aligned word or possible roots cision for model 1 is improved to 59.44% and for
extracted. If matched Turkish words do not have model 2 precision become 59.86%.
argument annotation, we transfer argument with
IBM Model 1 + Reinforce + Rules
the synset id found in the map record. Transferred Untransferred
As we discuss in the previous annotation trans- Correct 17,340 Not H.A. 1,151
fer procedure 4.2.2, some of the English words Incorrect 9,664
Undetermined 14,384 H.A. 2,170
such as proper nouns, time, date, numbers, ordinal Total 41,388 Total 3,321
numbers, percentiles, fractional numbers, number
intervals, and reel numbers stay same or take ad- Table 5: Results for IBM Model 1 alignment.
ditional suffixes in Turkish translation. So we in-
clude the same method used for matching these
words. In a case words are not aligned with the IBM Model 2 + Reinforce + Rules
Transfered Untransfered
information from alignment table, and a valid an-
Correct 17,464 Not H.A. 1,078
notation present in English word, we search exact Incorrect 9,635
string match or any word starts with the root of Undetermined 14,457 H.A. 2,075
Total 41,556 Total 3,153
English word in the Turkish sentence.
We run our procedure with IBM Model 1 & 2 Table 6: Results for IBM Model 2 alignment.
separately. We add reinforce step previously used
in Section 4.2.2. Unlikely previous attempts, af-
ter examining language structure we decided to 5 Conclusion
add rules to tag any untagged words after anno-
tation transfer. We observed argument types affect We proposed methods to generate automatic
noun inflections, for some argument types the last Turkish proposition bank by transferring cross-
word in constituent boundary is taking certain suf- language semantic information. Using the paral-
fixes. So first we find untagged word and select lelism with English proposition bank gives us an
the last word in its constituent boundary. Since opportunity to create a proposition bank in a short
we run reinforce step beforehand, only untagged time with less effort. We currently have 64% ac-
constituents exists in the sentence. In this respect, curacy with the hand-annotated proposition bank
we set the following rules to determine argument (Ak et al., 2018) for parallel sentence trees. When
annotation for untransfered words; we consider only transferred annotations, accu-
• For nouns and proper nouns: racy is rising to ∼75%. We also present annotation
projection to phrase sentences using WordNet and
– Have no suffix then ARG0
– Last morpheme tag is “ACCUSATIVE” (-(y)H, IBM alignment models. WordNet alignment heav-
-nH) or “DATIVE” (-(y)A, -nA) then ARG1 ily relies on semantic annotations, correct anno-
– Last morpheme tag is “LOCATIVE” (-DA, - tations transferred after this method is ∼14.59%.
nDA) or “ABLATIVE” (-DAn, -nDAn ) then
ARGMLOC However, 4,255 correct argument roles are trans-
– Last morpheme tag is “INSTRUMENTAL” (- ferred among 5,457 arguments which means 79%
(y)lA) then ARG2 of the transferred roles are correct. To increase
• For all word types annotation transfer for phrase sentences, we have
– Morphological parse contains date, time then also proposed alignment with IBM Model 1 and
ARGMTMP 2. Both models yields ∼60% correct annotations.
– Morphological parse contains cardinal number, Annotations transferred with these methods can
fraction, percent,
range, real number, ordinal number then provide a basis for proposition bank creation in
ARGMEXT resource-scarce languages. Annotations may then
40
be checked quickly by the annotators and proposi- P. Kingsbury and M. Palmer. 2003. Propbank: The next
tion bank reach the final state. level of treebank. In Proceedings of Treebanks and
Lexical Theories. Växjö, Sweden.
M. Kozhevnikov and I. Titov. 2013. Cross-lingual
References transfer of semantic role labeling models. In
ACL (1). The Association for Computer Linguistics,
Meyers A., R. Reeves, C. Macleod, R. Szekely, pages 1190–1200.
V. Zielinska, B. Young, and R. Grishman. 2004.
The nombank project: An interim report. In HLT- M. P. Marcus, M. A. Marcinkiewicz, and B. Santorini.
NAACL 2004 Workshop: Frontiers in Corpus Anno- 1993. Building a large annotated corpus of en-
tation. Association for Computational Linguistics, glish: The penn treebank. Computational linguistics
Boston, Massachusetts, USA, pages 24–31. 19(2):313–330.
K. Ak, O. T. Yıldız, V. Esgel, and C. Toprak. 2018. M. Palmer, D. Gildea, and P. Kingsbury. 2005. The
Construction of a Turkish proposition bank. Turk- proposition bank: An annotated corpus of semantic
ish Journal of Electrical Engineering and Computer roles. Comput. Linguist. 31(1):71–106.
Science 26:570 – 581.
M. Palmer, P. Kingsbury, O. Babko-Malaya, S. Cotton,
A. Akbik, L. Chiticariu, M. Danilevsky, Y. Li, and B. Snyder. 2004. Proposition bank i. Philadel-
S. Vaithyanathan, and H. Zhu. 2015. Generating phia: Linguistic Data Consortium. LDC2004T14.
high quality proposition banks for multilingual se-
mantic role labeling. In ACL (1). The Association L. Van der Plas, M. Apidianaki, and C. Chen. 2014.
for Computer Linguistics, pages 397–407. Global methods for cross-lingual semantic role and
predicate labelling. In COLING. ACL, pages 1279–
C. Bonial, J. Bonn, K. Conger, J. D. Hwang, and 1290.
M. Palmer. 2014. Propbank: Semantics of new
predicate types. In Proceedings of the Ninth In- L. Van der Plas, P. Merlo, and J. Henderson. 2011.
ternational Conference on Language Resources and Scaling up automatic cross-lingual semantic role an-
Evaluation (LREC-2014). European Language Re- notation. In ACL (Short Papers). The Association
sources Association (ELRA). for Computer Linguistics, pages 299–304.
P. F. Brown, V. J. D. Pietra, S. A. D. Pietra, and R. L. T. Zhuang and C. Zong. 2010. Joint inference for bilin-
Mercer. 1993. The mathematics of statistical ma- gual semantic role labeling. In Proceedings of the
chine translation: Parameter estimation. Comput. 2010 Conference on Empirical Methods in Natu-
Linguist. 19(2):263–311. ral Language Processing. Association for Computa-
tional Linguistics, Stroudsburg, PA, USA, EMNLP
G. G. Şahin. 2016a. Framing of verbs for turkish prop- ’10, pages 304–314.
bank. In TurCLing 2016 in conj. with 17th Interna-
tional Conference on Intelligent Text Processing and
Computational Linguistics (CICLING 2016).
41
Multilingual Sentence-Level Bias Detection in Wikipedia
42
Proceedings of Recent Advances in Natural Language Processing, pages 42–51,
Varna, Bulgaria, Sep 2–4 2019.
it is believed, science says); uncertainty mark- is too costly for large multilingual datasets.
ers, known as hedges (e.g., very, much, a bit, of- Hube and Fetahu (2018) learn to detect bias in
ten, approximately), editorializing (e.g., without a Wikipedia on a manually annotated corpus of sen-
doubt, arguably, however) and more. When an tences from the inherently biased Conservapedia,
article is considered biased, an editor can flag it with a precision of 0.74. When tested on an unla-
by adding a tag such as {{POV}} to its source, beled dataset extracted from Wikipedia however,
which displays a disputed neutrality warning ban- the classifier obtains a precision of 0.66 for the
ner on the page. These explicit guidelines (and sentences classified with a certainty over 0.8.
the editors who apply them) help reduce biased Recasens et al. (2013) first propose a heuristic to
language in Wikipedia over time through a con- automatically build a labeled corpus with biased
tinuous process of collaborative content revision sentences. Out of all revisions of NPOV-tagged
(Pavalanathan et al., 2018). Still, new instances of articles, they identify the bias-driven edits based
bias are introduced just as often as old ones are on the comments the editors left at commit. Al-
overlooked because of humans’ inherent difficulty though reliable, this method yields a fairly small
with subtle expressions of point-of-view partiality. set of examples for English (2,235 sentences) and
Recasens et al. (2013) showed that when presented none for smaller Wikipedias, first because of its
with a biased sentence from Wikipedia, annotators dependence on revision comments (which are op-
manage to correctly identify the loaded word in tional), and second, because it limits the examples
only 37% of the cases. to bias-driven edits containing five or fewer words.
43
0.68 on a corpus of 2,143 biased sentences from Tag Count Ratio
news articles, vectorized using tf-idf and classified weasel 201,092 0.5748
with a Mutlinomial Naive Bayes, and an accuracy weasel-inline 89,352 0.2554
of 0.77 for a CNN and 0.78 with a RNN. Finally, weasel words 21,755 0.0622
Hube and Fetahu (2018) achieve an F1 measure of weasel word 16,991 0.0486
0.70 using Random Forest on 686 manually anno- weasel section 3,954 0.0113
tated sentences from Conservapedia. weasel-section 3,743 0.0107
weasel inline 2,631 0.0075
3 Dataset Description
weaselinline 2,213 0.0063
We propose a procedure to semi-automatically de- weasel-words 2,176 0.0062
rive a labeled corpus of biased sentences from a weasel-word 2,102 0.0060
Wikipedia dump in any language, which, for this weaselword 1,967 0.0056
paper, we applied to the April 2019 dumps2 for weasel-name 956 0.0027
Bulgarian, French and English. weaselwords 503 0.0014
weasel section 225 0.0007
3.1 Tagset Curation weasel words 124 0.0004
First, we manually compile a list of NPOV- weasel word 80 0.0002
related tags for each of the target languages us-
ing the names of relevant Wikipedia maintenance Table 1: “Weasel” tag variation in English
templates3 ({{POV}}, {{NPOV}}, {{neutral
point of view}}, {{peacock}}, etc.). we extract the tag together with the pair of adja-
Most tags, however, vary in spelling, not only cent revisions, where the older one is tagged as
based on the context (e.g., inline or at the be- biased and the newer one is not. We opted for
ginning of an article), but also because of the this diachronic retrieval method, rather than re-
open and collaborative nature of Wikipedia. Ta- lying on the repertoire of articles in Wikipedia’s
ble 1 shows the sixteen most frequent “weasel” tag “NPOV dispute” section (Herzig et al., 2011; Re-
variations, only five of which (in bold) are docu- casens et al., 2013) since the latter only features
mented on Wikipedia. While the official tag is the currently tagged articles, while our method digs
most frequently used, the unofficial variations ac- NPOV violations from revision histories.
count for almost 35% of the most frequent ways to
tag a page containing weasel words. 3.3 Processing and Filtering
While it may be effortless for human editors
to interpret the meaning of these variations, it Each of these revision pairs undergoes a clean-
is not trivial to automatically identify all NPOV- ing process using regular expressions to strip as
related ones. Simply extracting all the tags starting much of the Wikipedia markup, links, and page
with the official form of “weasel” yields unrelated references as possible, while preserving visible
tags such as “weasel, back-striped” (an animal) or text and essential punctuation. At this point, we
“weasel, ben” (a punk singer). For that reason, we proceed to tokenize the text and split it into sen-
automatically compiled exhaustive tag frequency tences using the rule-based tokenizer and senten-
lists in each language, and then manually selected cizer methods of spaCy (Honnibal and Montani,
the relevant variations of each. 2017), whose 2.1.3 version supports 51 languages.
Finally, we replace all numbers with a special to-
3.2 Revision Extraction ken (numtkn), strip all remaining punctuation,
We look for occurrences of the selected tags across and convert everything to lowercase.
all revisions of each page, going forward from the Our algorithm also extracts revision pairs where
oldest one. When a biased revision is found, we the second member was the subject of a redirect or
follow its evolution until the POV tag disappears, vandalism, which we filter out. We then compare
at which point we assume the problematic con- the revisions to obtain the lists of deleted and in-
tent has been either rewritten or edited out. Next, serted sentences for each pair. In about 20% of
2 the cases, the difference consists in simply delet-
https://dumps.wikimedia.org
3
For English, see https://en.wikipedia.org/ ing the NPOV tag, which we believe is an artifact
wiki/Category:Neutrality_templates of editorial wars (Sumi et al., 2011; Yasseri et al.,
44
2012), given the contentiousness of most NPOV- 4.1 Protocol
flagged topics. Another 20% of the revision differ- For each language, we distribute the tagged/
ences we set aside are punctuation or case-related. untagged revision pairs into four bins, based on
We further clear the dataset from outliers the number of sentences that were removed in the
(mostly acts of vandalism) by removing those edit (bin 1: 1 or 2 sentences removed, bin 2: 3–
with more than 400 edited sentences. Finally, 6, bin 3: 7–15, bin 4: 16 or more; these val-
we exclude revision pairs with minor differ- ues were determined empirically to yield balanced
ences (character-based Levenshtein distance of 1), bins in terms of revision pairs). Each annotator la-
which are spelling corrections rather than bias res- beled 296 randomly picked sentences for a given
olution. Table 2 gives the number of initial, final language, distributed equally across the four bins.
and excluded revisions per language. 72 of these sentences (24%) were shared by all
annotators working on the same language, while
Revision pairs BG FR EN the remaining 224 were labeled by a single anno-
initial number 1,021 46,331 197,953 tator (cf. Table 4), thus allowing us to annotate
tag removal –257 –10,255 –61,397 more sentences while maintaining enough over-
punct./case –194 –5,967 –44,345 lap to measure inter-annotator agreement (IAA).
redir./vandalism –56 –1,524 –17,154 The Bulgarian sample was annotated by two native
deletions only –33 –2,740 –11,331 speakers, English by three with near-native profi-
insertions only –28 –2,819 –2,938 ciency, and French by two natives.
spelling –3 –136 –400
outliers –2 –153 –609 Lang All Ann1 Ann2 Ann3 Total
BG 72 224 224 — 520
Total pairs 448 22,737 59,779
FR 72 224 224 — 520
Table 2: Number of revision pairs per language EN 72 224 224 224 744
45
is thus not trivial but reasonable agreement can be (2) a. (before) its support is low only in the cholla
expected. province which has for nearly numtkn years sup-
ported kim dae jung a well known leftist politician
Bin BG EN FR avg std born in that province who also served as president
1 0.34 0.51 0.47 0.44 0.07 of south korea numtkn numtkn
2 0.64 0.45 0.45 0.52 0.09 b. (after) its support is low only in the jeolla province
3 0.63 0.45 0.38 0.48 0.11 which has for nearly numtkn years supported kim
4 0.63 0.52 0.34 0.50 0.12 dae jung a well known progressive politician born
in that province who also served as president of
avg 0.56 0.48 0.41 0.48 0.06
south korea numtkn numtkn
std 0.13 0.03 0.05 0.03 0.10
(3) a. (before) from the numtkn th century confucianism
Table 5: Positives in annotations was losing its influence on vietnamese society
monetary economy began to develop but unfortu-
Bin BG EN FR avg std nately in negative ways
1 0.32 0.55 0.67 0.51 0.15 b. (after) from the numtkn th century confucianism
2 0.22 0.58 0.44 0.41 0.15 was losing its influence on vietnamese society and
3 0.32 0.31 0.61 0.41 0.14 a monetary economy began to develop
4 – 0.23 0.39 0.68 0.28 0.38
5 Expressions of Bias
avg 0.16 0.46 0.60 0.41 0.18
std 0.23 0.11 0.10 0.08 0.21 The manual annotation also highlighted the vari-
ety of bias expression. Previously, Recasens et al.
Table 6: Inter-annotator agreement (Fleiss’ κ)
(2013) had identified two major classes: episte-
About half of the annotated sentences turn out mological and framing bias (subjective intensi-
to be neutral. Below, we discuss the sources of the fiers and one-sided terms), where they considered
noise we have observed in our dataset (including the first one to group more implicit expressions
the added sentences). such as factive and assertive verbs, entailment and
hedges. Based on their work and Wikipedia’s
4.3 Sources of Noise Manual of Style, we present biased examples from
We identified two types of noise: pipeline-related our corpus4 and discuss them in terms of the overt/
and human-related. Pipeline-related noise is either covert nature of the biased statement, its length
noise introduced at the pre-processing phase (e.g., (one or more words), and its level of ambiguity.
due to inconsistent sentence segmentation) or
Subjective intensifiers are mostly expressed
noise that remains despite our filtering and clean-
through single-word verbal and nominal modifiers
ing efforts (e.g., NPOV-unrelated edits longer than
(adverbs and adjectives) as in (4) and (5), but may
one character, differences resulting from the in-
also take the form of superlatives or quantifiers.
troduction of an infobox, differences consisting in
They explicitly undermine tone neutrality by in-
changing the spelling of numbers).
troducing overstatements and exaggerations (6).
Human editor-related noise comes from the data
itself and stems from the behaviour of Wikipedia’s (4) a. (before) some prominent liberals including scott
editors. It includes edits which introduce bias (of- reid were strongly critical of volpe s response
ten intentionally, as in (1) below), vandalism, cor- b. (after) some prominent liberals including scott reid
rections of factual mistakes unrelated to bias, re- criticized volpe s response
placing bias with another bias (cf. (2)), and collat-
eral edits, i.e., neutral sentences neighbouring bi- (5) (before) he is truly one of the greatest americans
ased ones indirectly targeted by a large-scope edit (6) a. (before) this is an absurd statement because the
(cf. (3)). Below are some examples. cavalry of any age is designed first and foremost to
run over the enemy and separate them as to make
(1) a. (before) cardinal health inc is a holding company
b. (after) cardinal health is a healthcare company ded- 4
Examples are taken from the English evaluation subsets,
icated to making healthcare safer and more pro- where sentences are in lowercase, stripped of punctuation and
ductive numbers are replaced by numtkn.
46
them far more vulnerable to being overwhelmed and security zone
overrun b. (after) in the first operation litani in numtkn the
b. (after) this is wrong because the cavalry of any age israel defense forces and south lebanon army oc-
is designed first and foremost to run over the enemy cupied a narrow strip of land described as the se-
and separate them as to make them far more vulner- curity zone
able to being overwhelmed and overrun
To state an opinion as a fact may be done with
Clichés and jargon tend to be non-ambiguous the use of an adverb (12) or an omission (13).
but introduce low-frequency words in the corpus,
as a result of being discouraged by Wikipedia. (12) a. (before) in fact the need for fast and secure fund
transfers is growing and in the next year instant
(7) (before) x force was concocted by illustrator rob liefeld
payments will quickly become the new normal for
who started penciling the new mutants comic book in
electronic fund transfers
numtkn
b. (after) it is predicted that in the next year instant
Describing or analyzing rather than reporting payments will become the standard for electronic
events is a form of partiality harder to model, as it fund transfers
may not necessarily contain explicitly proscribed
(13) a. (before) in numtkn the journal won the praise of
vocabulary.
fascist leaders
(8) (before) he was a former club rugby and an opening b. (after) there are some authors who retain that
batsman in club cricket but did not have the ability to the journal won the praise of fascist leaders
make it all the way to the top level these two sports have
become his particular area of expertise however he is Intentional vagueness or the omission of fac-
very knowledgable on all sports that are played tual information (14), is arguably the hardest type
of bias expression to detect not only for machines,
(9) (before) however the most important consequence of the which are expected to recognize the lack of data
battle was that president lincoln was able to sieze upon
as an informative feature, but also for humans,
the victory claim it as a strategic victory for the north
since filling factual gaps requires a fair amount of
and release his emancipation proclamation
domain-specific knowledge.
Active voice may be used in cases like (10) to
stress the agency of a participant in a situation, (14) a. (before) as of numtkn it is the ethnic minority party
alongside a positively loaded support verb. in romania with representation in the romanian
parliament
(10) a. (before) the united states department of justice in- b. (after) as of numtkn it is the ethnic minority party
dicted the company but amway secured an ac- in romania with representation in the romanian
quittal parliament and is part of the governing coalition
b. (after) the united states department of justice in- along with the justice and truth alliance and the
dicted the company but amway were acquitted conservatives
47
6.1 Embeddings • C: 1.0e–3, 1.0e–2, 1.0e–1, 1.0e0, 1.0e+1,
We used fastText’s classification function (Joulin 1.0e+2 and 1.0e+3.
et al., 2017), which implements a multinomial lo- • Solver: sag, saga.
gistic regression algorithm on top of pretrained Using the training and development sets to run
word embeddings. It uses word and character level the grid search optimization on all three lan-
embeddings to predict the class value of an in- guages, the average F1 measure was used to see
stance. The parameter optimization was done by which parameter values offered the best average
altering values for epoch (5, 10, 25), learning rate performance across the board. The selected values
(0.1, 0.01, 0.05), word n-grams (1 to 5), minimum were then used to run the same algorithm once on
count (1–5), embedding dimensions (100, 300), each language’s training and test sets.
loss function (softmax, ns, hs), minimum charac-
ter level n-gram size (2, 3), using pretrained vec- 7 Results and Discussion
tors or not, and learning rate update rate (50, 100).
When applying fastText’s pretrained vectors,5 Table 7 shows the results for the experiments
we obtained comparable results for English and detailed in §6 for the SGD, fastText and logis-
French without any significant gain, and with tic regression (LR) algorithms. For each perfor-
lower performance on Bulgarian. Thus, the fi- mance measure, dataset section, algorithm and
nal model chosen for its overall best performance language, we provide results with respect to the
across all three languages was trained without the biased class. The highest performance obtained
use of an additional language model. The best per- on the test dataset of each language is in bold.
forming values were then tried out on the test set. For the LR algorithm, the best performances
were obtained using a C value of 0.001 with the
6.2 Bag-of-Words Vectorization saga solver using a unigram model of 100 fea-
We also experimented with classic bag-of-word tures without inverse document frequency (idf)
vectorization with the stochastic gradient descent reweighting. The best parameters for the SGD
(SGD) (LeCun et al., 1998) and logistic regres- used a model of unigrams to trigrams, with an α
sion (Hosmer and Lemeshow, 2000) algorithms. of 0.001 and idf reweighting. For fastText, the best
Each algorithm was run with the same settings performing parameter set used the default values7
on all three datasets to get the best average over- and a minimum of 5 occurrences per token.
all performances for precision, recall and F1 mea- Overall, the similar results between the devel-
sure. Parameter optimization was done using a opment and test sets for each algorithm confirm
grid search. Stop word lists were used for each that they did not overfit. Furthermore, all three
language, which is the only language-specific as- measures have relatively low variance across lan-
pect of the experiment. guages, except for recall with SGD, which is con-
The optimization for SGD ran 72 permutations siderably lower for Bulgarian (also impacting F1 )
with the following parameters: than for the other two languages.
• Bag-of-word n-gram size: unigrams only, un- We observe that FastText’s vectorization and
igrams and bigrams, unigrams to trigrams. classification methods deliver higher precision
• Bag-of-word size: 100, 150, 300, 500, 1,000 upon larger datasets, but SGD and LR assure a
and 3,000. much higher recall regardless of the number of ex-
• Use idf reweighting or not. amples.
• α value: 0.01, 0.001. While relatively better, the SGD performance
All the other parameters were set to their default level on the test set leaves room for improvement.
values.6 For logistic regression, 504 permutations This is likely due to the noise level in the sentences
were tested using the following settings: labeled as biased, which count many non-biased
examples (see §4.2). The results are equally likely
• Same BOW n-gram size and BOW size and affected by the lexical and contextual ambiguity of
value type as SGD. the biased expressions, as discussed in §5. How-
5
Available for 157 languages, pretrained on Common ever, we do observe comparable best performance
Crawl and Wikipedia (Grave et al., 2018) https://
7
fasttext.cc/docs/en/crawl-vectors.html For version 0.8.3 of https://github.com/
6
Version 0.21.2 of the sklearn toolkit. facebookresearch/fastText
48
Measure Lang. Dev-SGD Test-SGD Dev-fastText Test-fastText Dev-LR Test-LR
Precision BG 0.5387 0.5886 0.5324 0.5330 0.5182 0.5032
FR 0.5059 0.5087 0.5533 0.5520 0.5151 0.5161
EN 0.5112 0.5083 0.5656 0.5634 0.5230 0.5224
Recall BG 0.4318 0.5049 0.4752 0.4937 0.6219 0.6303
FR 0.8877 0.8363 0.5724 0.5721 0.6751 0.6739
EN 0.8357 0.8277 0.5686 0.5718 0.5344 0.5354
F1 BG 0.4794 0.5435 0.5022 0.5126 0.5653 0.5596
FR 0.6444 0.6146 0.5627 0.5619 0.5844 0.5845
EN 0.6334 0.6291 0.5671 0.5676 0.5286 0.5288
Table 7: Results for each language, dataset and classification method for the biased class
across corpora of varying size and languages from indicates that the task is not trivial even for hu-
different families. mans.
On the test set, our best overall average F1 Using our corpora, we tested three classification
measure ranged between 0.56 and 0.62. This is algorithms: bag-of-word vectorization with SGD,
lower than Vincze (2013)’s 0.71 or Hube and Fe- fastText, and logistic regression.
tahu (2018)’s 0.70, but our approach uses a large In future work, we would like to improve the
corpus, automatically derived from Wikipedia in quality of the dataset by addressing issues uncov-
any language with minimal language-specific in- ered during the human evaluation, such as incoher-
put, applied to sentence-level bias detection, while ent sentence segmentation, enumerations, minor
Vincze (2013) used a monolingual, dictionary- edits and remaining noise. Another conceivable
based approach, and Hube and Fetahu (2018) re- optimization is to segment the dataset into two or
lied on language-specific resources to extract mul- more subsets according to the main forms of bias
tiple lexical and grammatical features. Our re- expression (e.g., explicit vs implicit). It would al-
sults set the baseline for sentence-level bias de- low to explore and evaluate different forms of bias
tection across the three languages of this corpus. separately, which in turn might motivate differen-
Higher performance for a specific language may tial classification techniques. Finally, populating
be achieved by a reconfiguration of the parameters the negative examples class with sentences from
or by the introduction of additional features. Wikipedia’s Featured Articles (in line with Bhos-
ale et al. 2013) might help reduce class ambiguity
8 Conclusion and Future Work by reinforcing the contrast between neutral ency-
clopedic tone and expressions of bias.
We presented a semi-automatic method to extract
biased sentences from Wikipedia in Bulgarian,
Acknowledgments
French and English. As this method does not
rely on language-specific features, apart from the This work has been supported by the Ministère de
NPOV tag list and a stop word list, it can be easily l’Économie et de l’Innovation du Québec (MEI).
applied to Wikipedia archives in other languages. We would like to thank the annotators for their
It relies on the tags added by human editors in the help with the quality evaluation process and the
articles that they considered biased. We retrieve anonymous reviewers for their insightful com-
the last tagged revision and the untagged revision ments.
following it and regard them respectively as biased
and unbiased. By comparing the revisions, we get
the lists of removed and added sentences. References
We manually annotated 1,784 of the removed
sentences, for all three languages combined, and Khalid Al Khatib, Hinrich Schütze, and Cathleen Kant-
ner. 2012. Automatic detection of point of view dif-
found that only about half of them were actually ferences in Wikipedia. In Proceedings of COLING
biased. An average Fleiss’ κ of 0.41 (0.46 if ig- 2012. The COLING 2012 Organizing Committee,
noring an outlier), consistent with the literature, Mumbai, India, pages 33–50.
49
Stefano Baccianella, Andrea Esuli, and Fabrizio Sebas- Matthew Honnibal and Ines Montani. 2017. spacy 2:
tiani. 2010. SentiWordNet 3.0: An enhanced lexi- Natural language understanding with bloom embed-
cal resource for sentiment analysis and opinion min- dings, convolutional neural networks and incremen-
ing. In Proceedings of the Seventh conference on tal parsing. To appear .
International Language Resources and Evaluation
(LREC’10). European Languages Resources Asso- David W. Hosmer and Stanley Lemeshow. 2000. Ap-
ciation (ELRA), Valletta, Malta. plied logistic regression. John Wiley and Sons.
Ricardo Baeza-Yates. 2018. Bias on the web. Commu- Christoph Hube and Besnik Fetahu. 2018. Detecting
nications of the ACM 61(6):54–61. biased statements in wikipedia. In Companion Pro-
ceedings of the The Web Conference 2018. Interna-
Ramy Baly, Georgi Karadzhov, Dimitar Alexandrov, tional World Wide Web Conferences Steering Com-
James Glass, and Preslav Nakov. 2018. Predict- mittee, pages 1779–1786.
ing factuality of reporting and bias of news media
sources. EMNLP-2018 . Mohit Iyyer, Peter Enns, Jordan Boyd-Graber, and
Philip Resnik. 2014. Political ideology detection us-
Martha Bellows. 2018. Exploration of Classifying Sen- ing recursive neural networks. In Proceedings of the
tence Bias in News Articles with Machine Learning 52nd Annual Meeting of the Association for Compu-
Models. Ph.D. thesis, University of Rhode Island. tational Linguistics (Volume 1: Long Papers). vol-
ume 1, pages 1113–1122.
Shruti Bhosale, Heath Vinicombe, and Raymond
Mooney. 2013. Detecting promotional content in Armand Joulin, Edouard Grave, Piotr Bojanowski, and
wikipedia. In Proceedings of the 2013 Conference Tomas Mikolov. 2017. Bag of tricks for efficient text
on Empirical Methods in Natural Language Pro- classification. In Proceedings of the 15th Confer-
cessing. pages 1851–1857. ence of the European Chapter of the Association for
Computational Linguistics: Volume 2, Short Papers.
Robert M Entman. 2007. Framing bias: Media in the Association for Computational Linguistics, Valen-
distribution of power. J. Commun. 57(1):163–173. cia, Spain, pages 427–431.
Richárd Farkas, Veronika Vincze, György Móra, János Sicong Kuang and Brian D Davison. 2016. Semantic
Csirik, and György Szarvas. 2010. The conll-2010 and context-aware linguistic model for bias detec-
shared task: learning to detect hedges and their tion. In Proc. of the Natural Language Processing
scope in natural language text. In Proceedings of meets Journalism IJCAI-16 Workshop. pages 57–62.
the Fourteenth Conference on Computational Nat-
ural Language Learning—Shared Task. Association Yann LeCun, Léon Bottou, Genevieve B. Orr, and
for Computational Linguistics, pages 1–12. Klaus-Robert Müller. 1998. Efficient backprop. In
Neural Networks: Tricks of the Trade, This Book is
Viola Ganter and Michael Strube. 2009. Finding an Outgrowth of a 1996 NIPS Workshop. Springer-
hedges by chasing weasels: Hedge detection using Verlag, London, UK, UK, pages 9–50.
wikipedia tags and shallow linguistic features. In
Proceedings of the ACL-IJCNLP 2009 Conference Chenghua Lin, Yulan He, and Richard Everson.
Short Papers. pages 173–176. 2011. Sentence subjectivity detection with weakly-
supervised learning. In Proceedings of 5th Interna-
Njagi Dennis Gitari, Zuping Zhang, Hanyurwimfura tional Joint Conference on Natural Language Pro-
Damien, and Jun Long. 2015. A Lexicon-based Ap- cessing. pages 1153–1161.
proach for Hate Speech Detection. International
Journal of Multimedia and Ubiquitous Engineering Gabriel Murray and Giuseppe Carenini. 2009. Detect-
10(4):215–230. ing subjectivity in multiparty speech. In Tenth An-
nual Conference of the International Speech Com-
Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Ar- munication Association.
mand Joulin, and Tomas Mikolov. 2018. Learning
word vectors for 157 languages. In Proceedings Cathy O’Neil. 2016. Weapons of Math Destruction:
of the International Conference on Language Re- How Big Data Increases Inequality and Threatens
sources and Evaluation (LREC 2018). Democracy. Crown.
Livnat Herzig, Alex Nunes, and Batia Snir. 2011. An Umashanthi Pavalanathan, Xiaochuang Han, and Ja-
annotation scheme for automated bias detection in cob Eisenstein. 2018. Mind your POV: Conver-
wikipedia. In Proceedings of the 5th Linguistic An- gence of articles and editors towards wikipedia’s
notation Workshop. Association for Computational neutrality norm. Proc. ACM Hum. -Comput. Inter-
Linguistics, Stroudsburg, PA, USA, LAW V ’11, act. 2(CSCW):137:1–137:23.
pages 47–55.
Evaggelia Pitoura, Panayiotis Tsaparas, Giorgos
Nicholas P Hirning, Andy Chen, and Shreya Shankar. Flouris, Irini Fundulaki, Panagiotis Papadakos,
2017. Detecting and identifying bias-heavy sen- Serge Abiteboul, and Gerhard Weikum. 2018. On
tences in news articles. Technical report, Stanford measuring bias in online information. ACM SIG-
University. MOD Record 46(4):16–21.
50
Marta Recasens, Cristian Danescu-Niculescu-Mizil,
and Dan Jurafsky. 2013. Linguistic models for an-
alyzing and detecting biased language. In Proceed-
ings of the 51st Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Pa-
pers). volume 1, pages 1650–1659.
Ellen Riloff and Janyce Wiebe. 2003. Learning extrac-
tion patterns for subjective expressions. In Proceed-
ings of the 2003 conference on Empirical methods in
natural language processing.
Lee D Ross, Teresa M Amabile, and Julia L Steinmetz.
1977. Social roles, social control, and biases in
social-perception processes. J. Pers. Soc. Psychol.
35(7):485.
Nicole Shadowen. 2019. Ethics and bias in ma-
chine learning: A technical study of what makes
us “good”. In Newton Lee, editor, The Transhu-
manism Handbook, Springer International Publish-
ing, Cham, pages 247–261.
R Sumi, T Yasseri, A Rung, A Kornai, and J Kertesz.
2011. Edit wars in wikipedia. In 2011 IEEE Third
International Conference on Privacy, Security, Risk
and Trust and 2011 IEEE Third International Con-
ference on Social Computing. ieeexplore.ieee.org,
pages 724–727.
A Tversky and D Kahneman. 1974. Judgment un-
der uncertainty: Heuristics and biases. Science
185(4157):1124–1131.
Veronika Vincze. 2013. Weasels, hedges and peacocks:
Discourse-level uncertainty in wikipedia articles. In
Proceedings of the Sixth International Joint Confer-
ence on Natural Language Processing. pages 383–
391.
Janyce Wiebe and Ellen Riloff. 2005. Creating sub-
jective and objective sentence classifiers from unan-
notated texts. In International conference on intel-
ligent text processing and computational linguistics.
Springer, pages 486–497.
51
Supervised Morphological Segmentation Using Rich Annotated Lexicon
Ebrahim Ansari,†‡ Zdeněk Žabokrtský,†
Mohammad Mahmoudi,‡ Hamid Haghdoost‡ and Jonáš Vidra†
†
Institute of Formal and Applied Linguistics,
Faculty of Mathematics and Physics, Charles University
‡
Department of Computer Science and Information Technology,
Institute for Advanced Studies in Basic Sciences
{ansari,m.mahmodi,hamid.h}@iasbs.ac.ir
{zabokrtsky,vidra}@ufal.mff.cuni.cz
52
Proceedings of Recent Advances in Natural Language Processing, pages 52–61,
Varna, Bulgaria, Sep 2–4 2019.
segmented training data is available, then the en- niques into their CRF-based model via a feature
tirely unsupervised systems tend not to be com- set augmentation. (Ruokolainen et al., 2014)
petitive. Furthermore, unsupervised segmenta-
Long Short Term Memory (LSTM) networks
tion still has considerable weaknesses, including
(Hochreiter and Schmidhuber, 1997) have recently
over-segmentation of roots and erroneous segmen-
achieved great success in sequence learning tasks,
tation of affixes (Wang et al., 2016). To deal
including outstanding results on sequential tasks
with those limitations, recent works show a grow-
such as machine translation (Sutskever et al.,
ing interest in semi-supervised and supervised
2014). Wang et al. (2016) proposed three types
approaches (Kohonen et al., 2010; Ruokolainen
of window-based LSTM neural network models
et al., 2013, 2014; Sirts and Goldwater, 2013;
named Window LSTM, Multi-window LSTMs
Wang et al., 2016; Kann and Schütze, 2016; Kann
and Bidirectional Multi-Window LSTMs, in or-
et al., 2018; Cotterell and Schütze, 2017; Grönroos
der to automatically learn sequence structures and
et al., 2019) which employ annotated morpheme
predict morphological segmentations of words in a
boundaries in the training phase.
raw text. They used only word boundary informa-
In our work we designed and evaluated various tion without any need for extra feature engineer-
machine learning models and trained them using ing in the training phase. The authors compared
only the annotated lexicon in a supervised manner. their models with selected supervised models as
Our models do not leverage the unannotated data well as with an LSTM architecture (Wang et al.,
nor context information and only use the primary 2016), and similarly to the work of Ruokolainen
hand-annotated segmentation lexicons. et al. (2013), their architecture is based on the
Experimental results show that our Bi-LSTM whole text and context information instead of us-
model perform slightly better than other models in ing only the lexicon. Cotterell and Schütze (2017)
boundary prediction for our hand-segmented Per- increased the segmentation accuracy by employ-
sian lexicon, while KNN (K-Nearest Neighbors al- ing semantic coherence information in their mod-
gorithm) performs better when the whole word ac- els. They used RNN (Recurrent Neural Network)
curacy is under question. to design a composition model. They also found
The paper is organized as follows: Section 2 that using RNN with dependency vector has the
addresses the related work on morphological seg- best results on vector approximation (Cotterell and
mentation. Section 3 describes the methodology Schütze, 2017).
and machine learning models used in this work.
Section 4 introduces our hand-segmented Persian Recently, using encoder-decoder models Bah-
lexicon as well as related preprocessing phases. danau et al. (2014) (attention-based models) made
Section 5 presents the experiment results com- some great successes in machine translation sys-
pared to some other baseline systems and finally tems. Kann and Schütze (2016) used an encoder-
Section 6 concludes the paper. decoder model which encodes the input as a se-
quence of morphological tags of source and tar-
2 Related Work gets and feeds the model by sequence of letters of
a source form. They select the final answer using
Supervised morphological segmentation, i.e. us- a majority voting amongst their five different en-
ing a lexicon (or a corpus) with annotated mor- sembled RNN encoder-decoder models. Kann and
pheme boundaries in the training phase, has at- Schütze (2016), proposed a seq2seq (sequence-to-
tracted increasing attention in recent years. One sequence network) architecture for the word seg-
of the most recent successful research directions mentation task. They used a bi-directional RNN
on supervised morphological segmentation is the to encode the input word (i.e. sequence of charac-
work of (Ruokolainen et al., 2013), whose au- ters) and concatenated forward and backward hid-
thors employ CRF (Conditional Random Fields), a den states yielded from two GRUs and pass the re-
popular discriminative log-linear model to predict sult vector to decoder part. The decoder is a single
morpheme boundaries given their local sub-string GRU which uses segmentation symbols for train-
contexts instead of learning a morpheme lexicon. ing. She introduced two multi-task training ap-
(Ruokolainen et al., 2014) extended their work proaches as well as data augmentations to improve
to semi-supervised learning version by exploiting the quality of the presented model. She shows that
some available unsupervised segmentation tech- neural seq2seq models perform on par with or bet-
53
ter than other strong baselines for polysynthetic the continuation letter, respectively. For exam-
languages in a minimal-resource setting. Their ple for word “goes”, the encoded segmentation
suggested neural seq2seq models constitute the is “LLBL”, which shows that there is a segmen-
state of the art for morphological segmentation in tation boundary in front of the third letter (“e”).
high-resource settings and for (mostly) European While in our model we consider only morpholog-
languages (Kann et al., 2018). ically segmented lexicon and we do not employ
The main studied language in our work is Per- any other information like corpus contexts or lists
sian, which belongs to morphologically rich lan- of unannotated words, this encoding is sufficient
guages and which is powerful and versatile in and make the specification of boundary location
word building. Having many affixes to form new easy.
words (over a hundred), and the ability to build af- In the case of presence of a semi-space let-
fixes and especially prefixes from nouns, the Per- ter (a feature specific for the Persian written lan-
sian language is considered as an agglutinative guage), the semi-space letter is always considered
language since it also frequently uses derivational as a boundary letter. An experiments focused on
agglutination to form new words from nouns, ad- this feature is described in Subsection 5.2.3, which
jectives, and verbal stems. Hesabi (1988) claimed shows that our models could perform better when
that Persian can derive more than 226 million this information exists in the annotated lexicon.
words (Hesabi, 1988).
3.1 Classification-Based Segmentation
To the best of our knowledge, the research
Models
on morphology of the Persian language is very
limited. Rasooli et al. (2013) claimed that per- In the first setup, we convert the segmentation task
forming morphological segmentation in the pre- (the task of segmenting a word into a sequence
processing phase of statistical machine transla- morphemes) simply to a set of independent bi-
tion could improve the quality of translations for nary decisions capturing the presence or absence
morphology rich and complex languages. Al- of a segmentation boundary in front of each let-
though they segmented very low portion of Per- ter in the word. For this task, we use various
sian words (only some Persian verbs), the qual- standard off-the-shelf classifiers available in the
ity of their machine translation system increases Scikit-learn toolkit (Pedregosa et al., 2011).
by 1.9 points of BLEU score. Arabsorkhi and So far, we provide the classifiers only with fea-
Shamsfard (2006) proposed a Minimum Descrip- tures that are extractable from the word alone.
tion Length (MDL) based algorithm with some More specifically, we use only character-based
improvements for discovering the morphemes of features. These character-based features include
Persian language through automatic analysis of letters and letter sequences (and their combina-
corpora. tions) before and after under the character under
question, which is subsequently assigned one out
3 Our Machine Learning Models of two classes: “B” for boundary characters, and
“L” which stands for continuation characters. The
In this work we decided to evaluate selected ma- main task of these methods is then to train a classi-
chine learning models including those feature- fication model to classify all characters in the word
based machine learning approaches in which the into those two classes, given binary features based
task of word segmentation is reformulated as on surrounding characters. For example, for the
a classification task, as well as various deep- fifth character of word “hopeless”, some of our
learning (DL for short) neural network models. features could be: “e”, “le”, and “ope”. The classi-
Because of huge number of learned parame- fication predictions are performed independently.
ters in DL, having enough training data is critical.
The fact that we decided to create a large hand- 3.2 Deep Neural Network Based Models
annotated dataset for Persian allows evaluating the Besides the classification-based segmentation
effect of the training data size on a relatively wide models, we designed and evaluated five DL mod-
scale, as described in Section 4. els based on GRU, LSTM, Bi-LSTM, seq2seq and
We convert all segmentations into a simple Bi-LSTM with the attention mechanism, respec-
string format in which letters “B” and “L” en- tively. The first three models are illustrated in Fig-
code the presence of the boundary letter and ures 1 and 2. The presented seq2seq model, is
54
similar to the model described in (Grönroos et al., et al., 2011) as well as the Czech dataset used in
2019). The last presented model is an attention our experiments are described.
based model, which is shown in Figure 3. In this
model, we use Bi-LSTM as encoder and LSTM as 4.1 Persian Hand-Annotated Morphological
attention layer, and finally, outputs of encoder and Segmentation Dataset
attention layers are added together. We extracted our primary word list from three dif-
ferent corpora. The first corpus contains sentences
extracted from the Persian Wikipedia (Karimi
et al., 2018). The second one is popular Persian
mono-lingual corpus BijanKhan (Bijankhan et al.,
2011), and the last one is Persian-NER1 (Poostchi
et al., 2018).
For all introduced corpora, using Hazm tool-
set (Persian preprocessing and tokenization tools)2
and the stemming tool presented by Taghi-Zadeh
et al. (2015), we extracted and normalized all
Figure 1: The schema of the LSTM/GRU models sentences and in the final steps using our rule-
used in this experiments. based stemmer and a Persian lemma collection, all
words are lemmatized and stemmed. Finally all
semi-spaces are automatically detected and fixed.
Words with more than 10 occurrences in the cor-
pora were selected for manual annotation. We
decided to send all 80K words to our 16 anno-
tators in the way that each word is checked and
annotated by two independent persons. Annota-
tors decided about the lemma of a word under
question, segmentation parts, plurality, ambiguity
(whether a word has more than one meaning) or
they might delete the word if they think is not a
proper Persian word. Moreover, some segmenta-
Figure 2: The schema of the Bi-LSTM model used
tions predicted by our automatic segmentator with
in this experiments.
high confidence score were offered to our annota-
tors. We removed almost 30K words which were
selected to be deleted by both annotators. And re-
maining 50K words sent for inter-annotation com-
parison part. In this step, all disagreements were
checked and corrected by the authors of this paper
and finally all words were quickly reviewed by two
Persian linguists. The whole process took around
six weeks. In order to use a hand-annotated lex-
icon in our work, we extracted the segmentation
part from the dataset and converted it to our binary
model which is described in Section 3.
Figure 3: The schema of the Bi-LSTM with the The total number of words we used in our Per-
attention mechanism model used in this experi- sian dataset is 40K. The dataset is publicly avail-
ments. able in the LINDAT/CLARIN repository (Ansari
et al., 2019).
4 Morphological Segmentation Lexicons
In this section, the rich Persian hand-annotated 1
https://github.com/HaniehP/
dataset and the existing Finnish datasets from PersianNER
the Morpho-Challenge shared task 2010 (Virpioja 2
https://github.com/sobhe/hazm
55
4.2 Existing Finnish and Czech Segmentation such accuracy in our experiments in addition to
Datasets our lexicon evaluation. For this new experiment,
We downloaded the Finnish segmentation dataset we selected a part of a mono-lingual text and after
from the Morpho-Challenge shared task 20103 removing all presented words in the text from our
(Virpioja et al., 2011) and converted them into our training lexicon, the remaining segmented words
binary format. The Finnish dataset contains 2000 are considered as the training set and finally ac-
segmented words. While comparing to our hand- curacy of word segmentation of words in test sen-
annotated Persian dataset these datasets are small, tences is reported separately.
we used them to see the efficiency of our presented
5.1 Baselines
models when the size of training dataset is limited.
The Czech dataset results from a prototype seg- We used two baseline systems which we selected
mentation annotation of Czech words. A sample to compare our models with. The first base-
of 1000 lemmas were selected randomly from De- line is an unsupervised version of MORFESSOR,
riNet, which is a lexical database focus on deriva- which is introduced and implemented by Creutz
tion in Czech (Žabokrtský et al., 2016). The et al. (2007). The second baseline is FlatCat
lemmas were manually segmented by two inde- (Grönroos et al., 2014), which is a well-known
pendent annotators, and all annotation differences semi-supervised version of MORFESSOR that
were resolved subsequently during a third pass uses the Hidden Markov Model for segmentation.
through the data. The annotation resulted in 4.6 In addition to the annotated data, semi-supervised
morphemes per word, partially as a result of the MORFESSOR (i.e. FlatCat) uses a set of 100,000
fact that the lemmas were sampled uniformly, re- word types following their frequency in the cor-
gardless of their corpus frequency, and thus the se- pus as their unannotated training dataset. For both
lection is biased towards longer words. baselines, the best performing model is selected
and compared with our neural network based mod-
5 Experimental Results els.
To partition our dataset (Persian, Czech and 5.2 Results and Discussion
Finnish) into training, development and test sets
As described in Section 4, we designed vari-
a commonly used method is used (Ruokolainen
ous models for the morphological segmentation
et al., 2013), which involves sorting words accord-
task. In the following subsections, different ex-
ing to their frequency and assigning every eighth
periments done in this work are reviewed. In all
term starting from the first one into the test set
tables, the column entitled by W% indicates the
and every eighth term from the second into the de-
proportion of perfectly segmented words. The
velopment set, while moving the remaining terms
column entitled by Ch% indicated the accuracy
into the training set.
of characters which are classified as boundary or
In order to evaluate the effect of the training
non-boundary. Finally, P%, R%, and F% indi-
data size, we randomly select the first 1/64, 1/32,
cate precision, recall and F-measure score respec-
1/16, 1/8, 3/8, 1/4, 1/2, 3/4 and all amount of data
tively for the morpheme boundary detection, natu-
from the training set to carry out experiments with
rally excluding the trivial final position characters
different training sizes. In all experiments, we re-
from our evaluation.
port three evaluation measures: the number of cor-
rectly predicted morpheme boundaries (in terms of 5.2.1 Comparison of Different Models
precision, recall, and f-measure), the percentage Table 1 shows the evaluation results of mor-
of correct binary predictions on all characters, and phological segmentation using our Persian hand-
the percentage of correctly segmented words. annotated dataset if the whole training data is
As described in Section 2, some previous works used. For each model, only results of the best-
reported accuracy in terms of the number of cor- performing hyperparameter configuration are re-
rect predictions (boundary and word) in a running ported. As is shown in Table 1, our Bi-LSTM
text, instead of considering unique words sampled model performs slightly better than the rest in
from a lexicon. Hence we decided to also report boundary prediction, however, the classification
3
http://morpho.aalto.fi/events/ models are surprisingly almost on the par with our
morphochallenge2010/datasets.shtml complex DL model. Considering word accuracy,
56
Model P% / R% / F% W% Ch% Model P% / R% / F% W% Ch%
LSTM 90.09 / 87.55 / 88.80 64.10 93.20 LSTM 99.67 / 29.08 / 44.98 03.58 81.57
GRU 85.43 / 84.50 / 84.96 58.35 91.44 GRU 99.99 / 28.01 / 45.01 03.59 81.60
Bi-LSTM 92.50 / 88.65 / 90.53 66.51 94.37 Bi-LSTM 86.96 / 32.82 / 47.66 04.88 81.30
Seq2Seq 88.04 / 84.04 / 86.09 59.10 91.65 Bi-LSTM with Attention 81.50 / 44.18 / 57.30 05.53 78.26
Bi-LSTM with Attention 92.57 / 85.85 / 89.08 65.30 93.52 SVC, Kernel: linear 78.39 / 76.83 / 77.31 38.11 91.16
SVC, Kernel: linear 85.86 / 82.20 / 83.94 73.08 94.45 SVC, Kernel: poly, Degree: 2 89.00 / 77.62 / 82.23 47.55 93.63
SVC, Kernel: poly, Degree: 2 89.57 / 85.86 / 87.67 78.72 95.72 SVC, Kernel: rbf 90.06 / 74.83 / 81.74 45.92 93.34
SVC, Kernel: rbf 89.71 / 84.42 / 86.99 77.61 95.52 SVC, Kernel: poly, Degree: 5 91.35 / 64.71 / 75.75 35.83 91.75
SVC, Kernel: poly, Degree: 5 89.77 / 83.91 / 86.74 77.17 95.45 SVC, Kernel: poly, Degree: 3 89.70 / 76.56 /82.61 46.57 93.58
SVC, Kernel: poly, Degree: 3 89.58 / 85.89 / 87.70 78.70 95.73 Logistic Regression, Solver: sag 82.43 / 69.37 / 75.34 31.92 90.95
Logistic Regression, Solver: sag 87.55 / 79.66 / 83.42 72.60 94.39 Logistic Regression, Solver: liblinear 82.43 / 69.37 / 75.34 31.92 90.95
Logistic Regression, Solver: liblinear 87.55 / 79.60 / 83.42 72.63 94.39 Logistic Regression, Solver: lbfgs 82.43 / 69.37 / 75.34 31.92 90.95
Logistic Regression, Solver: lbfgs 87.49 / 79.78 / 83.46 72.64 94.39 KNeighbors, Neighbors: 5 82.56 / 71.23 / 76.48 33.55 91.27
KNeighbors, Neighbors: 5 86.22 / 82.47 / 84.30 73.12 94.56 KNeighbors, Neighbors: 10 82.56 / 71.23 / 76.48 33.55 91.27
KNeighbors, Neighbors: 10 86.22 / 82.47 / 84.03 73.12 94.56 KNeighbors, Neighbors: 30 82.56 / 71.23 / 76.48 33.55 91.27
KNeighbors, Neighbors: 30 90.23 / 86.69 / 88.42 78.64 95.73 Ada Boost, Estimators: 100 76.45 / 38.48 / 51.19 16.28 85.38
Ada Boost, Estimators: 100 83.34 / 64.10 / 72.46 58.21 90.83 Decision Tree 79.58 / 76.29 / 77.90 39.41 91.38
Decision Tree 88.25 / 87.05 / 87.65 76.83 95.38 Random Forest, Estimators: 10 87.41 / 68.44 / 76.77 37.45 91.75
Random Forest, Estimators: 10 89.75 / 84.87 / 87.15 76.08 95.30 Random Forest, Estimators: 100 88.08 / 72.83 / 79.73 44.29 92.62
Random Forest, Estimators: 100 89.93 / 85.92 / 87.88 77.37 95.54 Bernoulli Naive Bayes 64.27 / 76.43 / 69.82 26.38 86.84
Bernoulli Naive Bayes 78.38 / 88.31 / 83.05 66.71 93.21 Perceptron MaxIteration: 50 73.22 / 75.36 / 74.27 31.92 89.60
Perceptron MaxIteration: 50 83.98 / 74.45 / 78.93 65.07 92.52 Unsupervised MORFESSOR 25.85 / 89.87 / 40.15 00.32 30.53
Unsupervised MORFESSOR 69.58 / 81.10 / 74.90 29.01 83.28 Supervised MORFESSOR 70.48 / 79.67 / 74.79 31.49 87.68
Supervised MORFESSOR 82.13 / 92.94 / 87.20 59.56 91.60
57
this also explains the inferior performance of our kind of space in Persian words and consequently
seq2seq model compared to the Bi-LSTM model our hand-annotated dataset contains these semi-
implemented for this work. spaces correctly. While we wanted to test the ef-
Finally, Table 5 shows results of selected mod- fect of having this prior knowledge in the lexicon,
els when the segmentation is done on all words we evaluated our models in two different forms. In
occurring in a corpus instead of a segmented lexi- the first case, we used our hand annotated dataset
con. In this experiments we expected those words as is. In the second case, we removed all semi-
with more frequency has higher effect on results spaces from the lexicon. Table 6 shows a compar-
comparing with less frequent words. ison for deploying our models on these two dif-
ferent datasets and as could be seen in this table,
5.2.2 Effect of Training Data Size
having the accurate dataset which is created by our
In order to evaluate the effect of the training data preprocessing strategy could improve results dras-
size on our DL models, different amount of train- tically.
ing data are selected from and feed to our models.
Figure 4 and Figure 5 demonstrate an experiment 6 Conclusion
in which the baseline line is the results of unsu-
pervised version of MORFESSOR for similar test The main task of this work is to evaluate different
dataset. Only four best performing feature-based supervised models to find the best segmentation
models in addition to two DL-based models are of a word when only a segmented lexicon without
selected to be shown here. As this figure shows, any extra information is available in the training
after having more than 10K training instances, in- phase. In recent years, recurrent neural networks
creasing the training data further does not have a (RNN) attracted a growing interest in morpholog-
substantial effect any more. ical analysis, that is why we decided to design
and evaluate various neural network based mod-
5.2.3 Semi-Space Feature for Persian Words els (LSTM, Bi-LSTM, GRU, and attention based
An important feature of the Persian and Arabic models) as well as some machine learning classi-
languages is the existence of semi-space. For ex- fication models including SVM, Random Forest,
ample word “( ”کتابهاbooks) is a combination of Logistic Regression and others for our morpho-
word “ ”کتابand “”ها, in which the former is Per- logical segmentation task. While a critical point
sian translation of word “book” and the latter is in any DL model is the training data size, we de-
morpheme for a plural form. We can say these cided to create a rich hand annotated Persian lex-
semi-space signs segment words into smaller mor- icon which is the only segmented corpus for Per-
phemes. However, in formal writing and in all Per- sian words. Using this lexicon we evaluated our
sian normal corpora, this space is neglected fre- presented models as well as the effect of train-
quently and it could make a lot of problems in ing data size on results. Moreover, we evaluated
Persian and Arabic morphological segmentation and tested our models on some limited datasets for
task. For example both forms for the previous Czech and Finnish languages. Experimental re-
example, “ ”کتابهاand “ ”کتابها, are considered sults show our Bi-LSTM model performs slightly
correct in Persian text and have the same mean- better in boundary prediction, however the results
ing. To deal with this problem and in order to of classification-based approaches overcome the
improve the quality of our segmentation dataset, DL models in percentage of completely correctly
we implemented a preprocessor to distinguish this segmented words.
Acknowledgments
Model P% / R% / F% W% Ch%
LSTM
Bi-LSTM
94.42 / 92.93 / 93.67
95.97 / 93.69 / 94.89
78.14
78.37
95.13
95.79
The research was supported by OP RDE project
SVC, Kernel: poly, Degree: 3
KNeighbors, Neighbors: 30
93.88 / 92.11 / 92.99
94.50 / 92.77 / 93.63
89.85
89.91
97.02
96.93
No. CZ.02.2.69/0.0/0.0/16 027/0008495, Interna-
Random Forest, Estimators: 100 94.32 / 91.99 / 93.10 88.64 96.66 tional Mobility of Researchers at Charles Univer-
sity, and by grant No. 19-14534S of the Grant
Table 5: Experiment results when a model is used Agency of the Czech Republic. It has been us-
to predict boundaries of Persian words of a small ing language resources developed, stored, and dis-
corpus instead of lexicon words. Only five best tributed by the LINDAT/CLARIN project of the
performing models are shown. Ministry of Education, Youth and Sports of the
58
Figure 4: The effect of Persian training data size on boundary detection F-measure.
Figure 5: The effect of Persian training data size on whole-word segmentation accuracy.
with semi without semi
Model
P% / R% / F% W% Ch% P% / R% / F% W% Ch%
LSTM 90.09 / 87.57 / 88.80 64.10 93.20 91.15 / 74.76 / 82.15 51.42 89.53
Bi-LSTM 92.50 / 88.65 / 90.53 66.51 94.37 89.19 / 77.18 / 82.75 52.58 89.64
SVC, Kernel: linear 85.86 / 82.20 / 83.94 73.08 94.45 81.67 / 77.75 / 79.66 68.17 92.62
SVC, Kernel: poly, Degree: 2 89.57 / 85.86 / 87.67 78.72 95.72 86.52 / 82.96 / 84.71 75.66 94.43
SVC, Kernel: rbf 89.71 / 84.42 / 86.99 77.61 95.52 86.34 / 80.96 / 83.56 74.39 94.08
SVC, Kernel: poly, Degree: 5 89.77 / 83.91 / 86.74 77.17 95.45 86.11 / 80.11 / 83.00 73.00 93.90
SVC, Kernel: poly, Degree: 3 89.58 / 85.89 / 87.70 78.70 95.73 86.30 / 83.02 / 84.63 75.30 94.39
Logistic Regression, Solver: sag 87.55 / 79.66 / 83.42 72.60 94.39 83.83 / 75.75 / 79.58 68.61 92.77
Logistic Regression, Solver: liblinear 87.55 / 79.60 / 83.42 72.63 94.39 83.84 / 75.75 / 79.59 68.63 92.78
Logistic Regression, Solver: lbfgs 87.49 / 79.78 / 83.46 72.64 94.39 83.74 / 75.59 / 79.46 68.47 92.73
KNeighbors, Neighbors: 5 82.47 / 86.22 / 84.30 73.12 94.56 82.19 / 76.34 / 79.15 67.36 92.52
KNeighbors, Neighbors: 10 86.22 / 82.47 / 84.30 73.12 94.56 82.19 / 76.34 / 79.15 67.36 95.52
KNeighbors, Neighbors: 30 90.23 / 86.69 / 88.42 78.64 95.73 82.19 / 76.34 / 79.15 67.36 92.52
Ada Boost, Estimators: 100 83.34 / 64.10 / 72.46 58.21 90.83 75.17 / 51.87 / 61.39 52.95 87.87
Decision Tree 88.25 / 87.05 / 87.65 76.83 95.38 88.24 / 86.05 / 87.13 75.92 95.21
Random Forest, Estimators: 10 89.75 / 84.87 / 87.15 76.08 95.30 85.04 / 78.17 / 81.46 70.83 93.38
Random Forest, Estimators: 100 89.93 / 85.92 / 87.88 77.37 95.54 85.21 / 79.66 / 82.34 71.95 93.65
Bernoulli Naive Bayes 78.38 / 88.31 / 83.05 66.71 93.21 75.63 / 84.91 / 80.00 62.01 92.01
Perceptron MaxIteration: 50 83.98 / 74.45 / 78.93 65.07 92.52 75.41 / 77.28 / 76.34 62.51 90.05
Unsupervised MORFESSOR 69.58 / 81.10 / 74.90 29.01 83.28 71.16 / 81.88 / 76.14 30.33 83.48
Supervised MORFESSOR 82.13 / 92.94 / 87.20 59.56 91.60 81.60 / 92.24 / 86.60 58.84 90.80
Table 6: The effect of considering semi-space on training data when all training data are used.
59
Morphologically Segmented Lexicon 0.5. LIN- tics for Uralic Languages. Association for Compu-
DAT/CLARIN digital library at the Institute of tational Linguistics, Tartu, Estonia, pages 15–26.
Formal and Applied Linguistics (ÚFAL), Faculty https://www.aclweb.org/anthology/W19-0302.
of Mathematics and Physics, Charles University,
https://hdl.handle.net/11234/1-3011. Stig-Arne Grönroos, Sami Virpioja, Peter Smit,
and Mikko Kurimo. 2014. Morfessor Flat-
Mohsen Arabsorkhi and Mehrnoush Shamsfard. 2006. Cat: An HMM-Based Method for Unsuper-
Unsupervised Discovery of Persian Morphemes. vised and Semi-Supervised Learning of Mor-
In Proceedings of the Eleventh Conference of phology. In Proceedings of COLING 2014,
the European Chapter of the Association for the 25th International Conference on Compu-
Computational Linguistics: Posters Demonstra- tational Linguistics: Technical Papers. Dublin
tions. Association for Computational Linguistics, City University and Association for Computational
Stroudsburg, PA, USA, EACL ’06, pages 175–178. Linguistics, Dublin, Ireland, pages 1177–1185.
http://dl.acm.org/citation.cfm?id=1608974.1609002. https://www.aclweb.org/anthology/C14-1111.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Zellig S. Harris. 1970. From Phoneme to Mor-
Bengio. 2014. Neural machine translation by pheme, Springer Netherlands, Dordrecht, pages 32–
jointly learning to align and translate. Cite 67. https://doi.org/10.1007/978-94-017-6059-1 2.
arxiv:1409.0473Comment: Accepted at ICLR 2015
as oral presentation. http://arxiv.org/abs/1409.0473.
Mahmoud Hesabi. 1988. Persian Affixes and Verbs,
Mahmood Bijankhan, Javad Sheykhzadegan, Mo- volume 1. Javidan.
hammad Bahrani, and Masood Ghayoomi. 2011.
Lessons from building a Persian written corpus: Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long
Peykare. Language Resources and Evaluation short-term memory. Neural Comput. 9(8):1735–
45(2):143–164. 1780. https://doi.org/10.1162/neco.1997.9.8.1735.
Kris Cao and Marek Rei. 2016. A joint model Katharina Kann, Jesus Manuel Mager Hois,
for word embedding and word morphology. In Ivan Vladimir Meza Ruiz, and Hinrich Schütze.
Proceedings of the 1st Workshop on Representa- 2018. Fortification of neural morphological
tion Learning for NLP. Association for Computa- segmentation models for polysynthetic minimal-
tional Linguistics, Berlin, Germany, pages 18–26. resource languages. In Proceedings of the 2018
https://doi.org/10.18653/v1/W16-1603. Conference of the North American Chapter of
the Association for Computational Linguistics:
Ryan Cotterell and Hinrich Schütze. 2017. Joint Human Language Technologies, Volume 1 (Long
semantic synthesis and morphological analysis Papers). Association for Computational Lin-
of the derived word. CoRR abs/1701.00946. guistics, New Orleans, Louisiana, pages 47–57.
http://arxiv.org/abs/1701.00946. https://doi.org/10.18653/v1/N18-1005.
Mathias Creutz, Teemu Hirsimäki, Mikko Ku- Katharina Kann and Hinrich Schütze. 2016. MED:
rimo, Antti Puurula, Janne Pylkkönen, Vesa The LMU system for the SIGMORPHON 2016
Siivola, Matti Varjokallio, Ebru Arisoy, Murat shared task on morphological reinflection. In Pro-
Saraçlar, and Andreas Stolcke. 2007. Morph- ceedings of the 14th SIGMORPHON Workshop
based speech recognition and modeling of out- on Computational Research in Phonetics, Phonol-
of-vocabulary words across languages. ACM ogy, and Morphology. Association for Computa-
Trans. Speech Lang. Process. 5(1):3:1–3:29. tional Linguistics, Berlin, Germany, pages 62–70.
https://doi.org/10.1145/1322391.1322394. https://doi.org/10.18653/v1/W16-2010.
Mathias Creutz and Krista Lagus. 2002. Un-
Akbar Karimi, Ebrahim Ansari, and Bahram Sadeghi
supervised discovery of morphemes. In Pro-
Bigham. 2018. Extracting an English-Persian Paral-
ceedings of the ACL-02 Workshop on Mor-
lel Corpus from Comparable Corpora. In Proceed-
phological and Phonological Learning. Associa-
ings of the Eleventh International Conference on
tion for Computational Linguistics, pages 21–30.
Language Resources and Evaluation, LREC 2018,
https://doi.org/10.3115/1118647.1118650.
Miyazaki, Japan, May 7-12, 2018..
John Goldsmith. 2001. Unsupervised learn-
ing of the morphology of a natural lan- Oskar Kohonen, Sami Virpioja, Laura Leppänen, and
guage. Comput. Linguist. 27(2):153–198. Krista Lagus. 2010. Semi-supervised extensions to
https://doi.org/10.1162/089120101750300490. Morfessor Baseline.
Stig-Arne Grönroos, Sami Virpioja, and Mikko Ku- Karthik Narasimhan, Regina Barzilay, and Tommi
rimo. 2019. North Sámi morphological seg- Jaakkola. 2015. An unsupervised method for uncov-
mentation with low-resource semi-supervised se- ering morphological chains. Transactions of the As-
quence labeling. In Proceedings of the Fifth In- sociation for Computational Linguistics 3:157–167.
ternational Workshop on Computational Linguis- https://doi.org/10.1162/tacl a 00130.
60
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, Hossein Taghi-Zadeh, Mohammad Hadi Sadreddini,
B. Thirion, O. Grisel, M. Blondel, P. Pretten- Mohammad Hasan Diyanati, and Amir Hos-
hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas- sein Rasekh. 2015. A new hybrid stem-
sos, D. Cournapeau, M. Brucher, M. Perrot, and ming method for Persian language. Digi-
E. Duchesnay. 2011. Scikit-learn: Machine learning tal Scholarship in the Humanities 32(1):209–221.
in Python. Journal of Machine Learning Research https://doi.org/10.1093/llc/fqv053.
12:2825–2830.
Sami Virpioja, Ville T. Turunen, Sebastian Spiegler,
Hoifung Poon, Colin Cherry, and Kristina Toutanova. Oskar Kohonen, and Mikko Kurimo. 2011. Empir-
2009. Unsupervised morphological segmentation ical comparison of evaluation methods for unsuper-
with log-linear models. In Proceedings of Hu- vised learning of morphology. TRAITEMENT AU-
man Language Technologies: The 2009 Annual TOMATIQUE DES LANGUES 52(2):45–90.
Conference of the North American Chapter of
Linlin Wang, Zhu Cao, Yu Xia, and Gerard de Melo.
the Association for Computational Linguistics.
2016. Morphological Segmentation with Window
Association for Computational Linguistics, Strouds-
LSTM Neural Networks. In Dale Schuurmans and
burg, PA, USA, NAACL ’09, pages 209–217.
Michael P. Wellman, editors, AAAI. AAAI Press,
http://dl.acm.org/citation.cfm?id=1620754.1620785.
pages 2842–2848.
Hanieh Poostchi, Ehsan Zare Borzeshi, and Massimo Zdeněk Žabokrtský, Magda Ševčı́ková, Milan Straka,
Piccardi. 2018. BiLSTM-CRF for Persian Named- Jonáš Vidra, and Adéla Limburská. 2016. Merging
Entity Recognition ArmanPersoNERCorpus: the data resources for inflectional and derivational mor-
First Entity-Annotated Persian Dataset. In Proceed- phology in Czech. In Nicoletta Calzolari, Khalid
ings of the Eleventh International Conference on Choukri, Thierry Declerck, Marko Grobelnik, Bente
Language Resources and Evaluation, LREC 2018, Maegaard, Joseph Mariani, Asunción Moreno, Jan
Miyazaki, Japan, May 7-12, 2018.. Odijk, and Stelios Piperidis, editors, Proceedings
of the 10th International Conference on Language
Mohammad Sadegh Rasooli, Ahmed El Kholy, and Resources and Evaluation (LREC 2016). European
Nizar Habash. 2013. Orthographic and Morpho- Language Resources Association, Paris, France,
logical Processing for Persian-to-English Statisti- pages 1307–1314.
cal Machine Translation. In Proceedings of the
Sixth International Joint Conference on Natural
Language Processing. Asian Federation of Natural
Language Processing, Nagoya, Japan, pages 1047–
1051. https://www.aclweb.org/anthology/I13-1144.
61
Combining Lexical Substitutes in Neural Word Sense Induction
62
Proceedings of Recent Advances in Natural Language Processing, pages 62–70,
Varna, Bulgaria, Sep 2–4 2019.
account all information we have about a partic- The approach presented in this paper is also
ular ambiguous word occurrence for better sub- an instance of vector clustering methods. More
stitutes generation. We compare several methods specifically, it exploits contextual substitutes for
for combining distributions. Substitutes retrieved the ambiguous word to differentiate between its
from the combined distribution perform much bet- senses. Baskaya et al. (2013) proposed using sub-
ter for WSI achieving the a state-of-the-art on the stitute vectors for WSI, and their system AI-KU
SemEval 2013 dataset for English as well two was one of the best-performing systems at Se-
datasets for Russian. mEval 2013. Alagić et al. (2018) proposed another
approach which leverages lexical substitutes for
2 Related Work unsupervised word sense induction. They perform
clustering of contexts using the affinity propaga-
The first methods to word sense induction were tion algorithm (Dueck and Frey, 2007). The sim-
proposed already in the late 90s (Pedersen and ilarity between instances is measured using three
Bruce, 1997; Schütze, 1998; Lin, 1998) with sev- different measures based on cosine similarities be-
eral competitions being organized to systemat- tween pre-trained word embeddings by Mikolov
ically evaluate various methods, including Se- et al. (2013). One measure relies on an average of
mEval 2007 task 2 (Agirre and Soroa, 2007), Se- embeddings of context words. Another one relies
mEval 2010 task 14 (Manandhar et al., 2010) on an average of embeddings of lexical substitutes
and SemEval 2013 task 13 (Jurgens and Kla- (also combination of both measures is tested). Fi-
paftis, 2013) for the English language, and RUSSE nally, Amrami and Goldberg (2018) proposed us-
2018 (Panchenko et al., 2018) for the Russian lan- ing neural language models and dynamic symmet-
guage.1 Navigli (2012) provides a survey of word ric patterns establishing a new best result on this
sense induction and related approaches. Methods dataset. Their approach is described in details in
for word sense induction can be broadly classified Section 3 as a starting point for our method.
into three groups: context clustering approaches,
word (ego-network) clustering, and latent variable 2.2 Word/Graph Clustering Methods
models. We discuss these approaches below. Also, This group of methods cluster word ego-networks
note that methods for learning word sense embed- consisting of a single node (ego) together with
ding (Camacho-Collados and Pilehvar, 2018) can the nodes they are connected to (alters) and all
be used to induce vector representations of senses the edges among those alters. Nodes of an ego-
from text. network can be words semantically similar to the
target word or context features relevant to the
2.1 Context/Vector Clustering Methods target. This line of work starts from the sem-
This methods from this group represent a word in- inal work of (Widdows and Dorow, 2002) who
stance by a vector that characterizes its context, used graph-based methods for unsupervised lexi-
where the definition of context can vary greatly. cal acquisition. In this work, senses of the word
These vectors are subsequently clustered. were defined as connected components in a graph
Early approaches, such as (Pedersen and Bruce, which excludes the ego. Véronis (2004), Biemann
1997; Schütze, 1998; Reisinger and Mooney, (2006), and Hope and Keller (2013) further devel-
2010) used sparse vector representations. Later oped this idea by performing clustering of nodes
approaches dense vector representations were instead of the simple search for connected com-
adopted, e.g. Arefyev et al. (2018) and Kutuzov ponents. Pelevina et al. (2016) proposed to trans-
(2018) used weighted word embeddings (Mikolov form word embeddings to sense embeddings using
et al., 2013) pre-trained on a large corpus to rep- graph clustering (Biemann, 2006). The obtained
resent context of an ambiguous target word. An- sense embeddings were used to solve the WSI task
war et al. (2019) used contextualized (Peters et al., based on similarity computations between the con-
2018) and non-contextualized (Mikolov et al., text and the induced sense.
2013) word embeddings to cluster occurrences of 2.3 Latent Variable Methods
ambiguous occurrences of verbs according to their
Methods from this group, define a generative pro-
semantic frames.
cess of the documents which include word senses
1
https://russe.nlpub.org as a latent variable and then perform estimation
63
of the model from unlabeled textual data. For in- is done S times resulting in S representatives of
stance, Lau et al. (2013) relies on the Hierarchical the original example consisting of 2L substitutes
Dirichlet Process (HDP) (Teh et al., 2006). Latent each. Then TF-IDF BoW vectors for all represen-
topics discovered in the training instances, spe- tatives of all examples of a particular ambiguous
cific to every word, are interpreted as word senses. word are built. Finally, agglomerative clustering
Since the HDP is generative, also new instances is performed on the obtained representations with
can be assigned a sense topic. Latent variable a fixed number of clusters. To provide more infor-
model of Bartunov et al. (2016) is a Bayesian ex- mation to the LMs the target word can be included
tension of Skip-gram (Mikolov et al., 2013) that in the context using the technique called dynamic
automatically learns the number of word senses; patterns. For example, given the sentence These
it relies on the stick-breaking process. Amplayo apples are sold everywhere instead of ’These ’ the
et al. (2019) propose another graphical model forward LM receives ’These apples and ’ and in-
which tackles the sense granularity problem, set- stead of ’ are sold everywhere’ the backward LM
ting new state-of-the-art results for the SemEval receives ’ and apples are sold everywhere’. The
2010/2013 WSI datasets. underscore denotes the position for which the lan-
guage model predicts possible words.
3 Bayesian Fusion of Lexical Substitutes Thus, lexical substitutes are obtained indepen-
from Bidirectional Language Models dently from the forward and the backward LM
and then united. For soft clustering required by
In this section, we describe the method of word
the SemEval-2013 dataset, the probability distri-
sense induction proposed by Amrami and Gold-
bution over clusters for each example is estimated
berg (2018), which is based on lexical substitutes
from the number of representatives of this example
generated given left and right context separately
put in each cluster. For the RUSSE (the Russian
and then united together. Then we propose sev-
WSI) datasets we further convert soft clustering
eral methods to build a combined distribution in-
into hard clustering by selecting the most proba-
corporating information from left and right con-
ble cluster for each example.
text as well as the similarity to the target word
for better substitutes generation. For qualitative The second baseline (named base) simplifies
comparison, Table 1 lists lexical substitutes gener- the original method by skipping sampling and us-
ated by different methods for several randomly se- ing S = 1 representative consisting of the union of
lected sentences from the TWSI dataset (Biemann, the top K predictions from each LM. While being
2012). For readability, we select either the top 10 simpler and deterministic, this modification also
predictions from the combined distributions or the delivers better results on RUSSE. Additionally,
union of the top 5 predictions from the forward we have found that baselines with dynamic pat-
and the backward distributions. The actual num- terns translated into Russian perform worse than
ber of substitutes may be smaller due to duplicates their counterparts without patterns (original-no-
appearing after lemmatization of substitutes. pat and base-no-pat) on RUSSE. This is in line
with the ablations study from Amrami and Gold-
3.1 Baselines: No Fusion (Union of berg (2018) who found that the patterns are use-
Substitutes) ful for verbs and adjectives but almost useless for
nouns which the RUSSE datasets consist of. Inter-
We base our approach on the method by Amrami estingly, our best models perform better without
and Goldberg (2018) (named original hereafter), dynamic patterns on all datasets.
which has achieved state-of-the-art results on the
SemEval-2013 dataset for English WSI. Suppose
3.2 Fusion at the Level of LM Distributions
c is the target ambiguous word and l, r are its left
and right contexts. First, the method employs pre- During preliminary experiments, we have found
trained forward and backward ELMo LMs (Peters that uniting substitutes retrieved from the forward
et al., 2018) to estimate probabilities for each word and the backward LM independently results in
w to be a substitute for c given only the left context lots of substitutes not related to the target word
Pf wd (w|l) or only the right context Pbwd (w|r). sense. For instance, consider the first example
Second, from the top K most probable words of in Table 1 where the ambiguous word is the last
each distribution L substitutes are sampled. This word of the sentence. The backward LM simply
64
base-no-pat base BComb-LMs BComb-3
It offers courses at the Undergraduate and Post Graduate levels in various subjects.
sept, industry, feb, univer- offer, course, teach, subject, profession, subject, indus- field, occupation, lan-
sity, discipline, nov, dec, style, topic, background, try, university, discipline, guage, discipline, sector,
language, field, oct size, include, provide sector, guise, language, guise, profession, subject,
field, department department, industry
Wakeboards with a three - stage rocker push more water in front of the wakeboard, making
the ride slower but riders are able to jump higher off the water.
slightly trip perfect be- faster landing climb bend rid jump incline slope climb dive incline climb trek slope
come journey climb trek bit harder speed walk bike bend trek journey crawl
speed
The couple were married on the bride’s family estate at Ballyhooly, Cork, Ireland; after-
wards the couple set up home at Caddington Hall.
tree bear residence holiday marry mansion be live castle honeymoon croft ranch farm ranch residence wed-
wedding vacation live farm farm cottage move divorce vineyard homestead resi- ding croft cottage home-
cottage dence farmhouse wedding stead
farm cottage
Table 1: Examples of generated lexical substitutes: baselines and our models. Contexts are from the
TWSI dataset. Ambiguous word is underlined, substitutes intersecting with human-generated are bold.
Here base is the baseline approach of Amrami and Goldberg (2018) and base-no-pat is its simplified
version without patterns, while BComb-LMs and BComb-3 are our models described in Section 3.
predicts all words which can appear before dot re- we estimate fused distribution as follows:
sulting mostly infrequent abbreviations. Dynamic
P (l, r|w)P (w)
patterns help a little, but there is still no context P (w|l, r) =
available for the backward LM to disambiguate P (l, r)
the target word. To solve this problem we pro- P (l|w)P (r|w)P (w)
= (2)
pose combining distributions from the forward and P (l, r)
the backward LM first and then taking the top K P (w|l)P (w|r)
∝ .
words from this combined distribution. We exper- P (w)
iment with the following combinations.
The numerator is estimated as Pf wd Pbwd , but pre-
trained ELMo LMs don’t contain frequencies of
3.2.1 Average (avg) the words in the vocabulary, so we cannot directly
estimate the denominator. Instead we approximate
This straightforward method of fusion of two dis-
it with Zipf distribution (the vocabulary is sorted
tributions computes an average of forward and
by frequency):
backward distributions (no information about the
target word is used): 1
P (w) ∝ , (3)
(k + rank(w))s
1
P (w|l, r) = (P (w|l) + P (w|r)) where k and s are hyperparameters: the first is
2 (1) needed to perform adjustment for frequent words
1
= (Pf wd + Pbwd ). while the second defines how quickly word fre-
2
quency drops as its rank grows.
65
about the target word using dynamic patterns, but ambiguous words, including 20 nouns, 20 verbs,
here we propose a probabilistic approach of fusion and 10 adjectives. It provides 20-100 contexts
of forward and backward distribution with the in- per word, and 4,664 contexts in total, which were
formation about the target word. Namely, we esti- gathered from the Open American National Cor-
mate similarity using a scaled dot product of out- pus and annotated with senses from WordNet. We
put embeddings from ELMo: used this dataset as the test set and tuned all hy-
embTw embc perparameters except for the number of clusters on
P (w|c) ∝ exp( ), (4) the TWSI dataset (Biemann, 2012).
T emperature
where T emperature is a hyperparameter which Evaluation metrics. Performance is measured
allows scaling this distribution to fit to the LM dis- with two cluster comparison measures: Fuzzy
tributions. Similarly to BComb-LMs and suppos- NMI (fNMI) and Fuzzy B-Cubed (fB3 ) as defined
ing the target word is independent from the con- in (Jurgens and Klapaftis, 2013).
text given possible substitutes (which can be inter-
preted as fixing a particular sense of the target): 4.1.2 Discussion of Results
Figure 1 shows a geometric average (AVG) be-
P (w|l)P (w|r)P (w|c)
P (w|l, c, r) ∝ . (5) tween Fuzzy Normalized Mutual Information
P 2 (w)
(fNMI) and Fuzzy B-Cubed F1 (fB3 ) depending
on the number of clusters. Following Amrami
and Goldberg (2018), Table 2 reports the results
for the number of clusters equal to 7 which is
the average number of senses in SemEval-2013.
BComb-3 shows the best results closely followed
by BComb-LMs, while the avg combination meth-
ods performs worse but still outperforms baseline
methods.
66
Model fNMI fB3 AVG
One sense for all 0.000 0.623 0.000
One sense per instance 0.071 0.000 0.000
Best competition results (Jurgens and Klapaftis, 2013)
AI-KU 0.065 0.390 0.159
Unimelb 0.060 0.483 0.170
Best after-competition results
(Amrami and Goldberg, 2018) 0.113 0.575 0.254
(Amplayo et al., 2019) 0.096 0.622 0.244
This paper
avg 0.120 0.562 0.260
BComb-LMs 0.139 0.566 0.280
BComb-3 0.135 0.586 0.281
Table 2: SemEval 2013 task 13: comparison to the previous best results. Following Amrami and
Goldberg (2018) the number of clusters is 7, other hyperparameters are selected on the TWSI dataset.
Figure 2: RUSSE-2018 development sets: ARI with respect to the number of clusters per word. Hyper-
parameters are selected on the TWSI dataset.
Evaluation metrics. Performance is measured ters BComb methods are indeed better than Avg.
using Adjusted Rand Index (ARI) (Hubert and We report results for (i) a fixed number of clusters
Arabie, 1985). (selected on the development sets) and for (ii) in-
dividual number of clusters for each word selected
4.2.2 Discussion of Results by maximizing the silhouette score of clustering4 .
Figure 2 shows results on the development set us- Using individual number of clusters consistently
ing the same hyperparameters used for SemEval- improves results for all our methods.
2013. Despite being selected on an English WSI 4.3 Experiment 3: TWSI Lexical
dataset, they perform surprisingly well. Similarly Substitution
to SemEval-2013, on active-dict BComb methods
outperform Avg by a large margin. However, on 4.3.1 Experimental Setup
bts-rnc dataset, Avg seems to be the best perform- In the third experiment, we evaluated the quality of
ing method which we attribute to suboptimal hy- lexical substitutes generated by our methods com-
perparameters. For our final submissions to the paring them with human-generated ones from the
leaderboard reported in Table 3 we selected hy- 4
https://scikit-learn.org/
perparameters on the development set correspond- stable/modules/clustering.html#
ing to each dataset and with these hyperparame- silhouette-coefficient
67
bts-rnc active-dict
Model
Test Test
avg 0.355 / 0.436 0.254 / 0.255
BComb-LMs 0.464 / 0.502 0.304 / 0.331
BComb-3 0.455 / 0.473 0.300 / 0.332
post compet’n best results 0.348 0.307
competition 1st best result 0.351 0.264
competition 2nd best result 0.281 0.236
Table 3: RUSSE 2018 test sets: comparison to the previous best results. The number of clusters is
selected on corresponding development sets (like other hyperparameters) / using silhouette score.
68
sociation for Computational Linguistics, pages 7– Chris Biemann. 2012. Turk bootstrap word sense in-
12. ventory 2.0: A large-scale resource for lexical sub-
stitution. In LREC. pages 4038–4042.
Domagoj Alagić, Jan Šnajder, and Sebastian Padó.
2018. Leveraging lexical substitutes for unsuper- Jose Camacho-Collados and Mohammad Taher Pile-
vised word sense induction. In Thirty-Second AAAI hvar. 2018. From word to sense embeddings: A sur-
Conference on Artificial Intelligence. vey on vector representations of meaning. Journal
of Artificial Intelligence Research 63:743–788.
Reinald Kim Amplayo, Seung won Hwang, and Min
Song. 2019. Autosense model for word sense in- Delbert Dueck and Brendan J Frey. 2007. Non-metric
duction. In AAAI. affinity propagation for unsupervised image catego-
rization. In 2007 IEEE 11th International Confer-
Asaf Amrami and Yoav Goldberg. 2018. Word ence on Computer Vision. IEEE, pages 1–8.
sense induction with neural biLM and symmet- David Hope and Bill Keller. 2013. UoS: A
ric patterns. In Proceedings of the 2018 Con- Graph-Based System for Graded Word Sense In-
ference on Empirical Methods in Natural Lan- duction. In Second Joint Conference on Lex-
guage Processing. Association for Computational ical and Computational Semantics (*SEM), Vol-
Linguistics, Brussels, Belgium, pages 4860–4867. ume 2: Proceedings of the Seventh Interna-
https://www.aclweb.org/anthology/D18-1523. tional Workshop on Semantic Evaluation (SemEval
2013). Atlanta, Georgia, USA, 1, pages 689–694.
Saba Anwar, Dmitry Ustalov, Nikolay Arefyev, Si- http://www.aclweb.org/anthology/S13-2113.
mone Paolo Ponzetto, Chris Biemann, and Alexan-
der Panchenko. 2019. HHMM at SemEval-2019 Lawrence Hubert and Phipps Arabie. 1985. Compar-
task 2: Unsupervised frame induction using con- ing partitions. Journal of classification 2(1):193–
textualized word embeddings. In Proceedings 218.
of the 13th International Workshop on Semantic
Evaluation. Association for Computational Linguis- David Jurgens and Ioannis Klapaftis. 2013. SemEval-
tics, Minneapolis, Minnesota, USA, pages 125–129. 2013 task 13: Word sense induction for graded
https://www.aclweb.org/anthology/S19-2018. and non-graded senses. In Second Joint Con-
ference on Lexical and Computational Semantics
Valentina Apresjan. 2011. Active dictionary of the rus- (*SEM), Volume 2: Proceedings of the Seventh In-
sian language: theory and practice. Meaning-Text ternational Workshop on Semantic Evaluation (Se-
Theory 2011:13–24. mEval 2013). Association for Computational Lin-
guistics, Atlanta, Georgia, USA, pages 290–299.
Nikolay Arefyev, Pavel Ermolaev, and Panchenko https://www.aclweb.org/anthology/S13-2049.
Alexander. 2018. Russian word sense induction by
clustering averaged word embeddings. In Proceed- Andrey Kutuzov. 2018. Russian word sense induc-
ings of the 24th International Conference on Com- tion by clustering averaged word embeddings. In
putational Linguistics and Intellectual Technologies Proceedings of the 24th International Conference
(Dialogue2018). RGGU, Moscow, Russia. on Computational Linguistics and Intellectual Tech-
nologies (Dialogue2018). RGGU, Moscow, Russia.
Sergey Bartunov, Dmitry Kondrashkin, Anton Osokin,
and Dmitry Vetrov. 2016. Breaking sticks and am- Jey Han Lau, Paul Cook, and Timothy Baldwin. 2013.
biguities with adaptive skip-gram. In Proceedings unimelb: Topic Modelling-based Word Sense Induc-
of the International Conference on Artificial Intelli- tion. In Second Joint Conference on Lexical and
gence and Statistics (AISTATS). Computational Semantics (*SEM): SemEval 2013.
Atlanta, Georgia, USA, volume 2, pages 307–311.
Osman Baskaya, Enis Sert, Volkan Cirik, and Deniz http://www.aclweb.org/anthology/S13-2051.
Yuret. 2013. AI-KU: Using substitute vectors Dekang Lin. 1998. An information-theoretic definition
and co-occurrence modeling for word sense induc- of similarity. In Proceedings of ICML. Madison,
tion and disambiguation. In Second Joint Con- WI, USA, volume 98, pages 296–304.
ference on Lexical and Computational Semantics
(*SEM), Volume 2: Proceedings of the Seventh In- Suresh Manandhar, Ioannis Klapaftis, Dmitriy Dli-
ternational Workshop on Semantic Evaluation (Se- gach, and Sameer Pradhan. 2010. SemEval-2010
mEval 2013). Association for Computational Lin- task 14: Word sense induction &disambiguation.
guistics, Atlanta, Georgia, USA, pages 300–306. In Proceedings of the 5th International Workshop
https://www.aclweb.org/anthology/S13-2050. on Semantic Evaluation. Association for Computa-
tional Linguistics, Uppsala, Sweden, pages 63–68.
Chris Biemann. 2006. Chinese whispers: an efficient https://www.aclweb.org/anthology/S10-1011.
graph clustering algorithm and its application to nat-
ural language processing problems. In Proceedings Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Cor-
of the first workshop on graph based methods for rado, and Jeff Dean. 2013. Distributed representa-
natural language processing. Association for Com- tions of words and phrases and their composition-
putational Linguistics, pages 73–80. ality. In Advances in Neural Information Process-
69
ing Systems 26. Curran Associates, Inc., Lake Tahoe,
NV, USA, pages 3111–3119.
Roberto Navigli. 2012. A quick tour of word sense
disambiguation, induction and related approaches.
In International Conference on Current Trends in
Theory and Practice of Computer Science. Springer,
pages 115–129.
Alexander Panchenko, Anastasiya Lopukhina, Dmitry
Ustalov, Konstantin Lopukhin, Nikolay Arefyev,
Alexey Leontyev, and Natalia Loukachevitch. 2018.
Russe’2018: a shared task on word sense induction
for the russian language. Computational Linguistics
and Intellectual Technologies pages 547–564.
Ted Pedersen and Rebecca Bruce. 1997. Distinguish-
ing word senses in untagged text. In Proceedings
of the Second Conference on Empirical Methods
in Natural Language Processing. Providence, RI,
pages 197–207.
Maria Pelevina, Nikolay Arefiev, Chris Biemann, and
Alexander Panchenko. 2016. Making sense of word
embeddings. In Proceedings of the 1st Workshop on
Representation Learning for NLP. Association for
Computational Linguistics, Berlin, Germany, pages
174–183. https://doi.org/10.18653/v1/W16-1620.
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt
Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. 2018. Deep contextualized word rep-
resentations. In Proceedings of the 2018 Con-
ference of the North American Chapter of the
Association for Computational Linguistics: Hu-
man Language Technologies, Volume 1 (Long
Papers). Association for Computational Linguis-
tics, New Orleans, Louisiana, pages 2227–2237.
https://doi.org/10.18653/v1/N18-1202.
Joseph Reisinger and Raymond J. Mooney. 2010.
Multi-prototype vector-space models of word mean-
ing. In Human Language Technologies: The
2010 Annual Conference of the North American
Chapter of the Association for Computational Lin-
guistics. Los Angeles, California, pages 109–117.
http://www.aclweb.org/anthology/N10-1013.
Hinrich Schütze. 1998. Automatic word sense discrim-
ination. Computational linguistics 24(1):97–123.
Yee Whye Teh, Michael I Jordan, Matthew J
Beal, and David M Blei. 2006. Hierarchi-
cal Dirichlet Processes. Journal of the Amer-
ican Statistical Association 101(476):1566–1581.
https://doi.org/10.1198/016214506000000302.
Jean Véronis. 2004. Hyperlex: lexical cartography
for information retrieval. Computer Speech & Lan-
guage 18(3):223–252.
Dominic Widdows and Beate Dorow. 2002. A graph
model for unsupervised lexical acquisition. In Pro-
ceedings of the 19th international conference on
Computational linguistics. Taipei, Taiwan, pages 1–
7. https://doi.org/10.3115/1072228.1072342.
70
Detecting Clitics Related Orthographic Errors in Turkish
71
Proceedings of Recent Advances in Natural Language Processing, pages 71–76,
Varna, Bulgaria, Sep 2–4 2019.
model employs a conditional random field (CRF) Our model focuses on the Turkish clitic “de/da”
for choosing the best prediction based on score that means “also, as well, too” and must always be
vectors that are provided by a multilayered bidi- written separately. It is commonly confused with
rectional LSTM. Words in input sentences are re- the locative suffix “de/da” that means “at” or “in”
placed with word embeddings trained with differ- as explained in Section 1.
ent algorithms. The model is tested with various
combinations of these pretrained embeddings on 2.2 The CoNLL Sentence Representation
a synthetically constructed dataset, where the best In 2003 a data format was introduced for the
scores were obtained when all three embeddings CoNLL-2003 shared task: Language-independent
were used that yielded an F1-Measure of 86.67%. named entity recognition (Kim Sang and De Meul-
It was also tested on a manually created more chal- der, 2003). In this format, each word is on a sep-
lenging dataset. arate line with an empty line after each sentence.
The main contributions of this work are: The first item of a line is a word, the second is a
part-of-speech (POS) tag, the third is a syntactic
• state-of-the-art spelling corrector that handles
chunk tag, and the fourth is the named entity tag.
the “de/da” misspellings in Turkish,
To represent sequences of meaningful words, the
• a comparative analysis of alternative word em- chunks and entities use B-TYPE to indicate the
bedding models for spell checking Turkish beginning and I-TYPE to indicate being inside the
sentences, phrase. The TYPE refers to the type of the entity
(i.e., person). Numerous datasets for NLP tasks
• a dataset of Turkish sentences with difficult to
utilize this format for interoperability. A word with
detect “de/da” errors, and
tag “O” (outside) is considered as not being a part
• a demo website for spellchecking sentences of a phrase. The CoNLL format is often used for
including “de/da” cases. publishing datasets. We use a variant of this for-
mat for representing correct and incorrect sentence
The remainder of the paper is organized as follows:
samples as detailed in Section 4.1.
Section 2 presents background information needed
to follow this work, Section 3 discusses the state-of- 2.3 Word Embeddings
the-art and current solutions to spelling corrections
Word embeddings are the vector representations of
in Turkish, Section 4 discusses the model and ex-
different sets of words. They are one of the most
periments, Section 5 presents an evaluation of the
widely utilized methods used for language repre-
proposed model, Section 6 reflects on observations
sentation. Word embeddings are capable of captur-
and provides insights about the future work, and
ing the semantic and syntactic similarity between
finally concluding remarks are given in Section 7.
words. In this work the word embeddings that
2 Background are used are GloVe (Pennington et al., 2014), fast-
Text (Grave et al., 2018) and Word2Vec (Mikolov
2.1 Clitic, Conjunction and Locative Suffix et al., 2013).
A clitic is a morpheme that is syntactically inde- Global Vectors for Word Representation
pendent but phonologically dependent and attached (GloVe) (Pennington et al., 2014) is an unsu-
to a host. It has the syntactic characteristics of a pervised learning algorithm to acquire word
word, but depends phonologically on another word vectors form words. It works on word to word
or phrase. global co-occurrence matrices and is successful
A conjunction is a word that syntactically con- in capturing semantic information. It combines
nects other words or larger constituents while also global matrix factorization and local context
expressing a semantic relationship between them. window methods to create word embeddings.
Some conjunction examples from English include FastText (Grave et al., 2018) is an open-source,
and, or, but and if. The clitic “de/da” can be given lightweight library for very fast text classification
as an example conjunction in Turkish. introduced by Facebook in 2016. FastText is pro-
The locative suffix indicates the locative case, posed as an extension of Word2Vec that trains
which is the grammatical case that conveys a loca- models given labeled texts, performs predictions,
tion. In Turkish, the locative case is specified by and evaluates models. It is a hierarchical classifier
the suffix “-de/-da”. where labels are represented in a binary tree that
72
Benim ben+Pron+Pers+A1sg+Pnon+Gen 3 Related Work
de de+Conj
aklım akıl+Noun+A3sg+P1sg+Nom Zemberek is a collection of natural language pro-
sende sen+Pron+Pers+A2sg+Pnon+Loc cessing tools for Turkish and is capable of various
kaldı kal+Verb+Pos+Past+A3sg tasks including morphological analysis, tokeniza-
tion and sentence boundary detection and basic
Table 1: The morphological analysis of a Turkish spell checker. It is also used as the spell checker for
sentence (My mind also remains with you) with LibreOffice. However, it is not capable of detecting
both the clitic and the affix forms of “de”. the misspelling of the clitic "de/da" as it does not
make a semantic analysis on the sentence.(Akın
and Akın, 2007)
facilities much faster model training without loss ITU Turkish Natural Language Processing
of accuracy. FastText breaks words into n-grams Pipeline can make syntactic and morphological
creating sub-words that are fed to the model to ob- analysis of raw Turkish sentences, although it is
tain the embeddings of each word. The tri-grams not capable of making a semantic analysis and thus
of the word selam are sel, ela, and lam. In this fails to classify and correct spellings of Turkish
way information about patterns within words are "de/da" clitic (Eryiğit, 2014).
captured, which enables out of vocabulary words The spelling correctors for Turkish do not satis-
to be processed. factorily correct misspellings of the “de/da” clitic
as they are limited to the morphological analysis of
Word2Vec models generate word embeddings
words which is insufficient for accurately classify
with a two-layer neural network that creates a set
them. Google, Microsoft Office, and LibreOffice
of feature vectors for words in a corpus.
all have different spell checkers for Turkish but
none of them present satisfactory results in the case
of handling the “de/da” clitics in Turkish. Their
2.4 Turkish Language
accuracy is significantly lower compared to our
model as will be detailed in Section 5.
Turkish is an agglutinative language, where com-
plex words are derived by stringing together mor- 4 Experiments and Results
phemes. In agglutinative languages a sequence of
affixes are attached to the end of the words. Ta- 4.1 Data
ble 1 shows the morphological analysis of the sen- To train the model, sentences with both correct and
tence (using the ITU NLP pipeline (Eryiğit, 2014)): incorrect spellings of the clitic “de/da” are required.
“Benim de aklım sende kaldı.”, which roughly trans- For this purpose, incorrect sentences have been
lates to “My mind remains with you too” (a manner generated from the correct sentences from a cor-
of expressing that one’s thoughts are with some- pus consisting of approximately 75 million Turkish
one). More literally it translates to “Also, my mind sentences extracted from various websites, novels
has remained with you.” This sentence includes and news sites (Yildiz et al., 2016). Since the cor-
both forms of “de”, which are shown in bold. The pus was extracted from novels and news sites, the
“de” following “Benim” refers to ”also”. The affix sentences are assumed to include only a few or
“de” within “sende” is locative and means at you no orthographic errors. Thus, the spellings of the
(in English this is expressed as with you). “de/da” cases are considered to be correct when
The morphological analysis of Turkish sentences written separately, attached as a locative suffix, or
can get very complex. It is rather difficult for used as a conjunction. Note that some words sim-
non native speakers to learn the ordering of affixes ply end with “de/da” and these suffixes are not due
and to distinguish among the clitics. Even native to locative morphemes (i.e., ‘ziyade’ meaning plen-
speakers may have trouble distinguishing the in- tiful). However, such cases are few and considered
tended meaning and will need to clarify the con- negligible.
text. These complexities present significant chal- To generate incorrectly spelled forms of “de/da”
lenges to building language supporting tools for samples, two simple actions are performed: (1)
Turkish. Although, machine learning approaches append the separately written “de/da” to its pre-
show promise. ceding word and (2) separate the “de/da” suffixes
73
Train Dev Test lem since it achieved the state-of-the-art results for
Sentences 15,203 3,729 2,070 named entity recognition, part-of-speech tagging
Tokens 383,066 94,232 51,226 and chunking tasks.
Table 2: The number of sentences and tokens for 4.3 Experimental Setup
the training, development, and test dataset used in The initial task was to train the model with Turkish
training our models. word embeddings. For this task, GloVe was used
with the dimension size of 300 and window size of
from the words that contain them. For example, 15. The pretrained word vectors for Turkish were
for the sentence “Kedi de gördüm” (meaning “I obtained from the model trained on Common Crawl
also saw a cat”), the sentence “Kedide gördüm” and Wikipedia using fastText (Grave et al., 2018).
(meaning “I saw it at the cat”) is generated by con- The pretrained Word2Vec vectors are for Turkish
catenation. Both are syntactically correct sentences with dimension size 300 (Güngör and Yıldız, 2017).
but have very different meanings. The sentence The models were trained using Continuous Bag of
“Evde kalıyorum” meaning “I am staying at home” Words (CBOW), with position-weights, dimension
which uses the locative suffix “de/da” correctly, the size of 300, character n-grams of length 5, and a
sentence “Ev de kalıyorum” is generated. The re- window size of 5 and 10 negatives.
sulting sentence is an incorrectly separated “de/da”, Parameter optimization was performed to
which translates to “I am staying also home”, which achieve the best F1 scores. During hyperparam-
doesn’t make sense. eter optimization, the training was performed for
The generated sentences are tagged in a manner 10 epochs using fastText embeddings for all possi-
like the CoNLL NER tags (Section 2.2). We tag in- ble configurations for the following criteria:
correctly spelled terms with “B - ERR” and all others
• batch size: [8, 16, 32, 64]
with “O” (other), such as:
Correct sentence Incorrect sentence • RNN layer size: [1, 2, 3, 4]
Onlar O Onlarda B-ERR
da O 'Sende O • learning rate: [0.05, 0.1, 0.15, 0.2]
'Sende O kalsın O
• hidden size: [16, 32, 64, 128, 256]
kalsın O , savcılığa O
, savcılığa O verirsin O The hyperparameters with the highest F1 score
verirsin O 'O are: batch size=16, RNN layer size=2, learning
'O dediler O rate=0.2, and hidden size=256. These parameter
dediler O . O values were used to train models with different
. O word embedding configurations for 150 epochs. All
models were trained on a PC with GPU GeForce
The dataset consisting of sentences whose words RTX2080 with 32 GB RAM. A single training took
are tagged with “B-ERR” and “O” are divided into approximately 10 hours to complete.
training, development, and test sets (Table 2).
In addition to the this synthetically constructed 5 Results and Evaluation
dataset, a dataset consisting of 100 Turkish sen-
A total of seven different models were trained with
tences with misspelled forms of “de/da” is formed
the optimal parameters. The embedding types
manually. The sentences in this second dataset is
used were GloVe (Pennington et al., 2014), fast-
created so that they are syntactically correct but
Text (Grave et al., 2018) and Word2Vec. These
semantically challenging to understand1 .
embeddings were also combined by concatenat-
4.2 Model ing them to form a new embedding with a higher
number of dimensions. Furthermore, two baseline
A multilayered bidirectional LSTM and CRF based models were used for comparison purposes. Base-
model (Akbik et al., 2018) that uses pretrained line model baseline1 considers only the separately
embeddings was considered suitable for our prob- written “de/da” as correct, falsely classifying the
1
Both this and the synthetic dataset is shared at https: correctly spelled locative suffix “de/da” as a mis-
//github.com/derlem/kanarya spelling. Baseline model baseline2 considers only
74
Model BL P R F1
G fT W 1 2 (%) (%) (%) Maça iyide başlamıştık aslında ama
+ 10.60 25.67 15.00 olmadı.
+ 59.89 74.32 66.33 Bu adam öyle aslında çokta kötü bir adam
+ 87.09 81.53 84.22 değil.
+ 87.05 79.73 83.23
Ya ders calış yada çık dışarıda oyna.
+ 87.67 79.50 83.39
+ + 90.55 81.98 86.05 Sonunda bizde derin öğrenmeye geçtik.
+ + 89.79 81.83 85.63 Kalemleri ve kitabı ev de kalmış.
+ + 87.59 80.03 83.64
+ + + 91.56 82.28 86.67 Belkide Galatasaray’ı şampiyonluktan
ettik.
Table 3: A comparison of the results our model Onuda yaptığında gidebilirsin.
trained with various combinations of the Glove (G),
fastText (fT) and Word2Vec(W) methods on a syn-
thetically constructed dataset against two baseline
models (BL-1 & BL-2). P, R and F1 refer to the Figure 1: The errors caught by our model with the
precision, recall, and F1 measures. best configuration on challenging sentences.
75
Ours Others Acc Onur Güngör and Eray Yıldız. 2017. Linguistic fea-
G fT W ITU (%) tures in Turkish word representations. In 2017
+ 55 25th Signal Processing and Communications Appli-
cations Conference (SIU). IEEE, pages 1–4.
+ 64
+ 71 Erik Tjong Kim Sang and Fien De Meulder. 2003.
+ + 66 Introduction to the CoNLL-2003 shared task:
language-independent named entity recognition. In
+ + 67 Proceedings of the seventh conference on Natural
+ + 69 language learning at HLT-NAACL 2003-Volume 4.
+ + + 65 Association for Computational Linguistics, pages
142–147.
+ 34
+ 29 Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Cor-
+ 0 rado, and Jeffrey Dean. 2013. Distributed represen-
tations of words and phrases and their composition-
+ 0 ality. In Proceedings of the 26th International Con-
ference on Neural Information Processing Systems -
Table 4: Results of spell checking of semantically Volume 2. USA, pages 3111–3119.
challenging sentences. G, fT, and W refer to Glove,
fastText and Word2Vec respectively. ITU is the Jeffrey Pennington, Richard Socher, and Christopher
Manning. 2014. GloVe: Global vectors for word rep-
ITU NLP Pipeline for Turkish, and the icons , , resentation. In Proceedings of the 2014 conference
and the spellcheckers of Google Docs, Microsoft on empirical methods in natural language process-
Office, and LibreOffice. ing (EMNLP). pages 1532–1543.
References
Alan Akbik, Duncan Blythe, and Roland Vollgraf.
2018. Contextual string embeddings for sequence
labeling. In COLING 2018, 27th International Con-
ference on Computational Linguistics. pages 1638–
1649.
Ahmet Afsin Akın and Mehmet Dündar Akın. 2007.
Zemberek: An open source NLP framework for Tur-
kic languages. Structure 10:1–5.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training of
deep bidirectional transformers for language under-
standing. In Proceedings of the 2019 Conference of
the North American Chapter of the Association for
Computational Linguistics: Human Language Tech-
nologies, Volume 1. pages 4171–4186.
Gülşen Eryiğit. 2014. ITU Turkish NLP web service.
In Proceedings of the Demonstrations at the 14th
Conference of the European Chapter of the Associ-
ation for Computational Linguistics. pages 1–4.
Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Ar-
mand Joulin, and Tomas Mikolov. 2018. Learning
word vectors for 157 languages. In Proceedings
of the International Conference on Language Re-
sources and Evaluation (LREC 2018).
76
Benchmark Dataset for Propaganda Detection in Czech Newspaper Texts
77
Proceedings of Recent Advances in Natural Language Processing, pages 77–83,
Varna, Bulgaria, Sep 2–4 2019.
Figure 1: Numbers of articles with significant attribute values (not null, neutral or missing) in the whole
collection of 7,494 documents. The first (yellow) columns show numbers for the whole collection and the
second (blue) columns show an example of a filtered subset of articles containing the word "Trump".
search in the Sketch Engine corpus manager (Kil- g) Fear mongering: is the text trying to appeal
garriff et al., 2014). As far as we know, this is the to fear, uncertainty or other threat?
first corpus of propaganda text annotated for de-
tailed ensemble of manipulative techniques. The h) Fabulation: does the text contain unsub-
full document texts were thus supplemented with stantiated, overstated or otherwise incorrect
the following attributes (see Figure 1 for represen- claims?
tation of particular attributes in the dataset):
i) Opinion: does the author of the text present
a) Blaming: does the text accuse someone of his or hers personal opinion?
something?
j) Location: what is the main location the text
b) Labelling: the text uses specific labels – talks about?
short and impactful phrases or words – to de-
k) Source: is the text presented as being based
scribe a person or a group.
on a specific source?
c) Argumentation: does the text present facts
l) Russia: is the topic related to Russia?
or arguments (logical, emotional, etc.) to
support the main claim? m) Expert: is the text or opinion in the text pre-
sented as being supported by an expert?
d) Emotions: What is the main emotion the text
is trying to evoke in the reader? Anger, hate, n) Attitude to a politician: neutral, negative,
fear. positive for up to 3 mentioned politicians.
e) Demonizing: is the “enemy” and/or his/her o) Topic: migrant crisis, domestic politics, etc.
goals or interests presented in the text as be-
ing evil? p) Genre: report, interview, or commentary.
f) Relativizing: are the presented actions of a q) Focus: foreign, domestic, can’t be deter-
person, group or party being relativized? mined.
78
Figure 2: An example of (a part of) an annotated article with ranges showing demonizing and grievance
as a value of the emotions attribute.
79
Table 2: Classifier Accuracy
Overall sentiment
Fear mongering
Argumentation
Demonizing
Relativizing
Fabulation
Labelling
Emotions
Location
Blaming
Opinion
Source
Russia
Expert
Server
Genre
Focus
Topic
dummy .59 .79 .69 .81 .96 .93 .91 .74 .86 .41 .60 .70 .74 .32 .89 .53 .75 .63
bernoulli nb .67 .78 .59 .74 .87 .85 .84 .75 .84 .56 .63 .73 .63 .53 .91 .72 .72 .80
multinomial nb .67 .79 .70 .81 .96 .93 .91 .74 .86 .52 .60 .71 .74 .54 .89 .86 .75 .72
nearest centroid .66 .71 .62 .63 .74 .80 .75 .71 .75 .58 .60 .55 .67 .56 .80 .66 .65 .73
passive aggressive .70 .79 .72 .77 .96 .94 .92 .78 .84 .74 .67 .79 .80 .69 .95 .85 .73 .92
random forest .69 .81 .74 .81 .96 .93 .92 .77 .87 .67 .68 .80 .80 .63 .92 .85 .76 .88
ridge .72 .82 .75 .81 .96 .94 .92 .79 .89 .75 .70 .80 .81 .71 .96 .87 .78 .91
sgd elasticnet .71 .82 .73 .81 .96 .94 .92 .78 .89 .76 .70 .82 .80 .71 .96 .87 .77 .93
sgd l1 .70 .81 .72 .81 .96 .94 .92 .78 .89 .76 .70 .82 .81 .70 .96 .87 .77 .94
sgd l2 .70 .82 .73 .81 .96 .94 .92 .78 .89 .76 .70 .81 .80 .71 .96 .87 .77 .92
80
Table 4: Classifier prediction accuracy sorted by the weighted F1-score which takes into account im-
balanced attribute classes. The resulting accuracy is compared to the baseline accuracy of the majority
class.
best classifier weighted F1 accuracy baseline difference
Demonizing sgd l2 .85 .96 .96 .00
Genre sgd elasticnet .84 .96 .89 .07
Server sgd l1 .83 .94 .63 .31
Relativizing sgd elasticnet .82 .94 .93 .01
Fear mongering passive .81 .92 .91 .01
aggressive
Opinion sgd l2 .79 .89 .86 .03
Focus ridge .77 .87 .53 .34
Labelling ridge .73 .82 .79 .03
Expert ridge .73 .81 .74 .07
Russia sgd l1 .71 .82 .70 .12
Emotions ridge .70 .81 .81 .00
Fabulation ridge .70 .79 .74 .04
Overall sentiment ridge .70 .78 .75 .04
Location sgd l2 .68 .76 .41 .36
Argumentation ridge .65 .75 .69 .06
Blaming ridge .65 .72 .59 .13
Topic sgd elasticnet .64 .71 .32 .39
Source ridge .63 .70 .60 .10
3.3 Evaluation Metrics positive words, only negative words and of their
Each trained classifier predicts the class for a doc- average score computed as a difference between
ument based on its text. By comparing the re- word positivity and negativity. The overall doc-
sults to the dataset gold standard data, each of the ument scores were then computed as a maximum
classifier was evaluated by means of its attribute- positive paragraph score, maximum negative para-
related accuracy, precision, recall, and F1 score. graph score and maximum and minimum of the
The accuracy results are summarized in Table 2 average word score for each paragraph.
and compared with the dummy baseline accuracy Each of the resulting document sentiment scores
in Table 4. were evaluated for a correlation4 with positive and
negative values of the selected attributes annotated
3.4 Correlations of Attributes and Sentiment in the data. The results are presented in Table 5.
Coefficients None of the attributes has proven really strong cor-
relation, but several attributes partly correlate with
The set of article attributes contains several items
the maximum negative sentiment of the document.
which express sentiment values, either to the arti-
Interestingly, there is no correlation in case of the
cle as a whole or to a mentioned politician. We
emotions attribute.
have evaluated the possibility of using the article
sentiment analysis to predict the corresponding at- 4 Conclusion and Future Directions
tribute values for the texts.
The paragraph sentiment analysis results were We have introduced a new benchmark dataset
explicitly expressed as an average score of posi- for propaganda manipulative techniques detection
tivity and negativity of particular words. A list of in Czech newspaper texts. The dataset contains
6,261 words was prepared as projections of Senti- 7,494 documents annotated for the presence of
WordNet (Baccianella et al., 2010) scores via the eight manipulative techniques and 10 document
Czech WordNet (Rambousek et al., 2018; Horák attributes relevant for propaganda detection. The
et al., 2008) database, see Table 3 for examples. 4
Computed as Spearman’s correlation coefficient with
Each paragraph received an average value of only statistical significance.
81
Table 5: Correlations of selected attributes and document sentiment analysis scores. The † symbol
denotes statistically significant values (p < 0.05) of Spearman’s correlation coefficient.
Attribute max positive max negative max average min average
blaming 0.18 † 0.23 † 0.17 † -0.23 †
demonizing 0.11 † 0.13 † 0.11 † -0.12 †
fear mongering 0.16 † 0.18 † 0.16 † -0.18 †
emotions compassion 0.02 -0.00 0.03 -0.00
emotions fear -0.07 † 0.02 -0.07 † -0.02
emotions hate 0.06 † 0.04 0.06 -0.04
emotions grievance -0.00 -0.05 -0.00 0.05
overall sentiment 0.16 † 0.18 † 0.16 † -0.18 †
attitude1 0.04 † 0.04 † 0.04 † -0.04 †
attitude2 0.10 † 0.15 † 0.09 † -0.16 †
attitude3 0.13 † 0.13 † 0.11 † -0.13 †
attitude avg 0.13 † 0.14 † 0.11 † -0.15 †
dataset is currently being expanded with the third passive-aggressive algorithms. Journal of Machine
part of documents from 2018 and it is planned to Learning Research 7(Mar):551–585.
be released for public access after this expansion. Lisa K Fazio, Nadia M Brashier, B Keith Payne, and
We have evaluated the current data with 10 cur- Elizabeth J Marsh. 2015. Knowledge does not pro-
rent classification techniques. Regularized linear tect against illusory truth. Journal of Experimental
regression and Support vector machines are able Psychology 144(5):993–1002.
to classify the data with the best accuracies, even Todd C Helmus, Elizabeth Bodine-Baron, Andrew
though the manipulative techniques need to em- Radin, Madeline Magnuson, Joshua Mendelsohn,
ploy extra features to significantly improve over William Marcellino, Andriy Bega, and Zev Winkel-
man. 2018. Russian Social Media Influence: Under-
the baseline. standing Russian Propaganda in Eastern Europe.
In the currently running experiments, we are Rand Corporation.
preparing new evaluation of the dataset using de-
Edward Herman and Noam Chomsky. 2012. A propa-
tailed stylometric features and distributed seman-
ganda model. Media and cultural studies: Keyworks
tic representations of the texts. pages 204–230. Reproduced from Manufacturing
Content, 1988.
Acknowledgments.
Aleš Horák, Karel Pala, and Adam Rambousek. 2008.
This project was partially supported by the Grant The Global WordNet Grid Software Design. In Pro-
Agency of Masaryk University within the project ceedings of the Fourth Global WordNet Conference,
University of Szegéd. pages 194–199.
MUNI/G/0872/2016.
Adam Kilgarriff, Vı́t Baisa, Jan Bušta, Miloš
Jakubı́ček, Vojtěch Kovář, Jan Michelfeit, Pavel
References Rychlý, and Vı́t Suchomel. 2014. The Sketch En-
gine: ten years on. Lexicography 1(1):7–36.
Stefano Baccianella, Andrea Esuli, and Fabrizio Sebas-
tiani. 2010. Sentiwordnet 3.0: an enhanced lexical Andrew McCallum, Kamal Nigam, et al. 1998. A com-
resource for sentiment analysis and opinion mining. parison of event models for naive bayes text classi-
In Proceedings of LREC 2010. pages 2200–2204. fication. In AAAI-98 workshop on learning for text
categorization. pages 41–48.
Vı́t Baisa, Ondřej Herman, and Aleš Horák. 2017. Ma-
nipulative Propaganda Techniques. In Proceedings Robert M McIntyre and Roger K Blashfield. 1980.
of Recent Advances in Slavonic Natural Language A nearest-centroid technique for evaluating the
Processing, RASLAN 2017. pages 111–118. minimum-variance clustering procedure. Multivari-
ate Behavioral Research 15(2):225–238.
Leo Breiman. 2001. Random forests. Machine learn-
ing 45(1):5–32. Jan Michelfeit, Jan Pomikálek, and Vı́t Suchomel.
2014. Text tokenisation using unitok. In Aleš Horák
Koby Crammer, Ofer Dekel, Joseph Keshet, Shai and Pavel Rychlý, editors, RASLAN 2014. Tribun
Shalev-Shwartz, and Yoram Singer. 2006. Online EU, Brno, Czech Republic, pages 71–75.
82
Christopher Paul and Miriam Matthews. 2016. The
Russian ”firehose of falsehood” propaganda model.
Rand Corporation pages 2–7.
Adam Rambousek, Aleš Horák, and Karel Pala. 2018.
Sustainable long-term WordNet development and
maintenance: Case study of the Czech WordNet.
Cognitive Studies/Études cognitives (18).
Ryan M Rifkin and Ross A Lippert. 2007. Notes on
regularized least squares. Computer Science and
Artificial Intelligence Laboratory Technical Reports
https://dspace.mit.edu/handle/1721.1/37318.
83
Diachronic Analysis of Entities by Exploiting Wikipedia Page revisions
Pierpaolo Basile Annalina Caputo
University of Bari Aldo Moro ADAPT Centre
Dept. of Computer Science Trinity College Dublin
Bari, Italy Dublin, Ireland
pierpaolo.basile@uniba.it annalina.caputo@adaptcentre.ie
84
Proceedings of Recent Advances in Natural Language Processing, pages 84–91,
Varna, Bulgaria, Sep 2–4 2019.
pages. The Wikipedia page history, sometimes a distributed, time-based, word representation for
called revision history or edit history, tracks the the JISC UK Web Domain Dataset (1996-2013)
order in which changes were made to any editable corpus.
Wikipedia page. We believe that this corpus can
Other research efforts have been directed to
help researchers to design approaches for track-
release resources and applications for the visual
ing entities usage over time. This resource can be
analysis and querying of these diachronic collec-
functional to promote new research for dynamic
tions. The Google Ngram viewer (Michel et al.,
embeddings of named entities. We propose some
2011) was released as a tool for allowing users to
preliminary case studies for proving the potential-
query the Google Ngram corpus, a collection of
ity of this resource.
ngram occurrences spanning several years and lan-
The paper is structured as follows: Section 2 re- guages extracted from the Google Book project.
views the state of the art, while Section 3 describes Hellrich and Hahn (2017) proposed a system that
the methodological aspects of our approach. Sec- allows users to explore different corpora via a di-
tion 4 shows some use cases of our resource fol- achronic semantic search. They used the Cor-
lowed by some final remarks. pus of Historical American English, the Deutsches
Textarchiv “German Text Archive”, and the Royal
2 Related Work Society Corpus, in addition to the Google Books
Ngram Corpus.
The diachronic analysis of language via word em-
beddings has been an active area of research dur- Research directed toward the specific problem
ing the past decade that has generated many re- of detecting changes in the context surrounding
sources for several time periods, languages and named entities has attracted limited attention com-
genres. Kim et al. (2014) used Google Ngram pared to the broader area of automatic detection
as a diachronic resource to build word embed- of the semantic shift of words. Some previous
dings via Word2Vec on a random sample of the work on named entities focused on problems re-
10 million 5-grams from the English fiction por- lated to searching (Berberich et al., 2009; Kan-
tion of the corpus. The authors made the resource habua and Nørvåg, 2010a; Zhang et al., 2016).
available, but due to space limitations, they re- Tahmasebi et al. (2012) proposed an interesting
leased the word embeddings only for the 5-year approach to identify the evolution of named en-
time period. A similar approach was proposed tities. Berberich et al. (2009) defined a method
by Grayson et al. (2016), where Word2Vec em- for query reformulations able to paraphrase the
beddings are trained on the Eighteenth-Century user’s information need using terminology preva-
Collections Online corpus (ECCO-TCP) by tak- lent in the past. In this work, the original dataset
ing into account five twenty-year periods for 150 is enriched with annotated phrases extracted from
million words randomly sampled from the “Liter- the text by using Wikipedia page titles. In Kan-
ature and Language” section of the corpus. Hamil- habua and Nørvåg (2010a), Wikipedia internal
ton et al. (2016b) also trained word embeddings on links and redirect pages are exploited for finding
the Google Ngram for detecting semantic changes. synonyms across time by using different snapshot
The authors analysed four different languages, i.e. of Wikipedia. The identified synonyms are used
English, French, German and Chinese, and cre- for query expansion in order to increase the re-
ated a resource which has been successfully used trieval effectiveness. In some respects, this ap-
in subsequent studies (Garg et al., 2017; Hamilton proach is similar to ours. However, it does not
et al., 2016a). A different approach to detect the use page revisions and the relation between con-
semantic shift of words was adopted by Kulkarni cepts, surface forms and contexts. Zhang et al.
et al. (2015). The authors adopt a change point (2016) described an approach to find past simi-
detection algorithm on the time series generated lar terms closest to a given present term. The
by computing the cosine similarity between word goal was to improve the retrieval effectiveness in
embeddings trained on several corpora, such as: archives and collections of past documents. In this
Twitter, Amazon reviews, and the Google Book work, Wikipedia is only functional to the creation
Ngrams. A similar approach is proposed in Basile of the test set, where only the information about
and McGillivray (2018), in which the Temporal the entity lifetime is used (e.g. the time when the
Random Indexing (TRI) is adopted for building name of a country or a company changed). Re-
85
garding named entity evolution, Tahmasebi et al. the changes over time of the temporal relation
(2012) proposed a method to capture the evolution existing between the surface form, the target
of one name into another by using a sliding win- and context. For example, it is possible to track
dow of co-occurrence terms. The corpus used for the change over time of different surface forms
the evaluation is the New York Times Annotated linking to a specific target or to detect the change
Corpus. Lansdall-Welfare et al. (2017) analysed in the target context. All these capabilities open
a collection of historical data spanning 150 years several possibilities to the analysis of entities
of British articles. The authors focus on historical using a diachronic perspective.
and cultural changes that are tracked via a quan-
titative analysis of word frequencies. However,
they expand their methodology to a “semantic”
level by working on named entities extracted from
text. The work proposed in Szymanski (2017)
is the first attempt to highlight the potential of
diachronic word embeddings for solving analogy
tasks involving entities and relationships, although
this work does not seek to capture named entities
in an explicit way. Moreover, Caputo et al. (2018)
applied a method to recognise and linking named
entities in the whole New York Times corpus. The
Temporal Random Indexing is then applied on the
annotated corpus in order to build a semantic vec-
tor representation for entities and tracking signifi-
cant shift in their contexts. An explicit representa-
tion of named entities is also provided in (Bianchi Figure 1: Flowchart of the dataset creation.
et al., 2018) where the authors tackle the problem
of incorporating time in the Knowledge Graph em- Figure 1 depicts the process followed for the
beddings in order to provide a similarity measure creation of our resource. The starting point is the
between entities that accounts for temporal fac- Wikipedia meta history dump which includes all
tors. the page revisions in XML format. The dump
is composed of several XML files containing the
3 Methodology page revisions in Mediawiki syntax. Each XML
file is parsed using the DKPro-JWPL API1 , which
The revision history associated with each is able to produce the accurate Abstract Syntax
Wikipedia page opens the way to different di- Tree (AST) of each page revision. From the
achronic analyses of the highly interconnected AST, we extract all the internal links that refer
concepts represented by its pages. In Wikipedia, to standard2 Wikipedia pages; each internal link
pages are interconnected by internal links man- has a surface form and the name of the linked
ually created by users that consist of a surface Wikipedia page. In addition, we extract the year
form and a target. The target is another Wikipedia from the revision timestamp and the context as
page, and can be regarded as a “conceptual” link the n words around the internal link. The con-
created by the user between the surface form and text is processed using the StandardAnalyzer pro-
a specific concept (the Wikipedia page). The vided by the Apache Lucene API3 . Each extracted
same surface form can link several entities and internal link is saved in a CSV file as a record con-
the same entity can be linked to several surface sisting of: year, pageId, target, surface form, left
forms. Moreover, since a surface form occurs context and right context.
in a specific context, we can define the surface An example of a row in a CSV file is reported
context as a window of n words to the left and below:
to the right of the surface form. Each page has 1
https://github.com/dkpro/dkpro-jwpl
multiple revisions created every time a user edits 2
We remove links to special pages, such as category and
that page, and each revision page is associated user pages.
with a timestamp, so that it is possible to track 3
http://lucene.apache.org/
86
2003 11057 forge forge forging henson 1 associates 1 eric 1
term shaping metal use heat tracking 1 facial 1 technology
hammer basic smithy contains 1 collaborated 1 six 1 starting 1
sometimes called hearth heating worked 1 animation 1
metals commonly iron steel The aggregated format shows that the surface
malleable temperature form Apple Computer was used twice for link-
The row meaning is that in page 11057 in the ing the target Apple Computer, while the surface
year 2003 the target forge is linked by the surface form Apple was used only once. The BoW follows
form forge with the right context forging, term, ... each surface form. In the first aggregation step,
and the left context sometimes, called, .... an aggregated file is created for each segment of
Since the tuple <year, surface form, target> the Wikipedia dump, then in the second aggrega-
can occur multiple times, we aggregate multiple tion step, all the segments are merged in the final
tuple occurrences in a single record. The aggrega- dataset.
tion step is performed several times, one time for In this first version of the dataset we do not take
each dump file plus a final step that aggregates all into account disambiguation pages and redirects.
the records in a single file that represents our final Managing redirects is a very challenging problem
dataset. since they are not consistent over dumps.
In the final file, information is stored as follows: Relying on this final dataset, we built a search
• A row starting with the sequence #T API for easily retrieving all the information related
<TAB>Ti which identifies the beginning of a to the target, the surface form and the context ac-
sequence of rows in the file that are related to cording to a specific time period4 .
the page (concept) Ti (until a new row start- We exploit the meta history dump dated 1st
ing with #T is encountered). Ti represents the February 2019; the first Wikipedia pages are dated
Wikipedia page title; 2001. The original dump size is about 950GB, the
total size of the CSV files is about 30GB, while
• A sequence of rows containing several val- the final dataset obtained by aggregating data from
ues separated by the tabular character in the the CSVs is about 47GB. We set to 10 the dimen-
form: year yk , surface form sj , the number sion of the context window. Since a page can have
of time that the surface sj is used for link- multiple revisions in the same year, in building our
ing Ti in the year yk . Then, we build a Bag- resource we consider only the latest one for each
of-Word (BoW) from the words occurring in year. It is possible to perform a more fine-grained
the context, and in the same row we provide analysis by taking into account more revisions per
the BoW size followed by all the words in year (e.g. a revision for each month). The to-
the BoW represented as a sequence of pairs tal number of distinct targets is about 31M, which
<word, occurrences>. is larger than the effective number of Wikipedia
pages for several reasons: 1) some targets are a
A row in the aggregate format is shown in the
redirect to other targets; 2) some pages have been
following example:
removed or renamed over the years; 3) some tar-
#T Apple Computer
gets are a link to a specific section of a page. In
2018 Apple Computer 2 30 freedos
this release, we do not take into account these is-
1 x 1 supports 1 support 3
sues, which we plan to tackle in a future release.
officially 1 10 1 s 1 programming
The search API can be used for building sev-
1 9 1 scsi 1 bda 1 2005 1 usb
eral applications, such as a RESTful Web Services
2 mac 3 announced 2 storage 2
for remotely querying the data, data analysis for
august 1 31 1 ray 2 advanced 1 os
discovering when named entities or surface forms
3 its 2 interface 2 blu 2 joined
change their usage, and data visualisation.
1 aspi 1 march 1 8.5.1 1 disc 2
It is important to underline that the proposed ap-
mass 2
proach is completely unsupervised and language
2018 Apple 1 21 developed 1
independent since it does not require any NLP pre-
computer 1 years 1 independently
1 group 1 computer’s 1 1987 4
The dataset and the source code are available here
1 he 1 while 1 advanced 1 https://github.com/pippokill/dae
87
processing step. Moreover, the proposed method- 4.2 Contexts of a given target
ology is intrinsically multi-language because it is Another interesting analysis concerns the change
possible to rely on the specific Wikipedia dump of over time of the contexts of a given target. In this
the language under analysis. In addition, it is pos- case, it is possible to compute the displacement
sible to exploit multi-language Wikipedia links for over time of the target concept by computing the
comparing the evolution of named entities across cosine similarity between the context BoWs. For
different languages. each pair of years, we build a BoW vector for the
One limit of our approach is the short time context of the target concept. Then, we generate a
frame taken into account since Wikipedia was time series by computing the cosine similarity be-
launched in 2001. However, our approach is in- tween the BoW of two consecutive years (BoWyi
cremental and the dataset can grow when new and BoWyi+1 ). Figure 2 reports the time series
Wikipedia dumps are available. Moreover, the generated for the concept “Donald Trump”; we
dataset is not only useful for diachronic analysis observe a change point corresponding to a drop in
of entities, but the detection of semantic changes similarity between 2015 and 2016.
over a short period of time can be exploited to
improve the performance of several algorithms,
such as entity linking, relation extraction and on-
tology/knowledge graph population.
4 Use cases
88
2015 2016 2017 2018 2019
Donald Trump Donald Trump pres- Donald Trump Donald Trump Donald Trump
idential campaign,
2016
Protests against Don- Protests against Don- Protests against Don- Donald Trump pres-
ald Trump ald Trump ald Trump idential campaign,
2016
Donald Trump sexual Donald Trump pres- Donald Trump pres- Protests against Don-
misconduct allegations idential campaign, idential campaign, ald Trump
2016 2016
Political positions of Donald Trump sexual Donald Trump sexual Donald Trump sexual
Donald Trump misconduct allegations misconduct allegations misconduct allegations
Stop Trump movement Donald Trump (Last Presidency of Donald Presidency of Donald
Week Tonight) Trump Trump
activity have started to appear, to become the most gets (concepts). Using this approach it is possi-
frequent words in the BoWs since 2004. ble to show how the similarity between two tar-
Another interesting use case is the analysis of gets changes over time. In particular, for each time
the BoWs of the targets linked by the same sur- point we build the BoW of each concept and then
face form. This analysis may highlight changes we compute the similarity between the BoWs. It is
in the way common words are used for referring important to point out that the target BoW is built
to named entities. For example, analysing the us- by taking into account the context around each oc-
age of the surface form “tweet”, we observe that currence of the target in the corpus. In this way,
since 2012 it has been used to refer to the concept if two targets occur in similar contexts their BoWs
“Twitter”, while before 2012 it did not refer to any will be similar. We adopt two strategies:
concept.
point-wise: each BoW is built by taking into ac-
4.3 Similarity between two entities over time count only the target occurrences at time ti ;
The last scenario shows the possibility to compute cumulative: each BoW is built by taking into ac-
the similarity between two entities over time as the count all the target occurrences up to time ti ,
cosine similarity between the target contexts. Fig- including time ti . The idea is to take into ac-
ure 4 reports the time series of similarities between count all the previous history of the target and
three pairs of entities (Apple-Microsoft, Apple- not only the time period under analysis.
IBM, IBM-Microsoft). It is interesting to observe
that the similarity between IBM and Microsoft is Observing the plots in Figure 5, it is possible
higher then the similarity between Apple and the to note that the similarity between United States-
other two entities, although Apple is equally re- U.S. President and United States-Donald Trump is
lated to both of them. constant across time, while we observe an incre-
ment in similarity between U.S. President-Donald
Trump after the year 2018. This increment is
clearly evident in the point-wise analysis (Figure
5a), as we expected. It is important to underline
that in Figure 5a some points are near zero (2009-
2014) this means that the targets do not occur in
similar contexts in that periods and indeed the two
BoWs share just a few words. Figure 5b show a
different trend, since we take into account all the
previous target occurrences before the time ti by
exploiting the cumulative approach.
Figure 4: Comparison between pair of entities. The promising results obtained in this prelimi-
nary case study about BoW similarity suggest that
Finally, the plots in Figure 5 show the cosine it is possible to build effective “time-dependent”
similarity between the BoWs of two different tar- embeddings by using our resource.
89
(a) Plot of the point-wise targets BoW cosine similarity over(b) Plot of the cumulative targets BoW cosine similarity over
time. time.
Figure 5: BoW analysis of pair of targets: plot over time of the cosine similarity between BoWs of two
targets with point-wise (a) and cumulative (b) strategy.
90
: Characterising 19th Century Fiction via Word Em- 55th Annual Meeting of the Association for Compu-
beddings. In 24th Irish Conference on Artificial In- tational Linguistics (Volume 2: Short Papers). As-
telligence and Cognitive Science (AICS’16), Univer- sociation for Computational Linguistics, pages 448–
sity College Dublin, Dublin, Ireland, 20-21 Septem- 453. https://doi.org/10.18653/v1/P17-2071.
ber 2016.
Nina Tahmasebi, Gerhard Gossen, Nattiya Kanhabua,
William L. Hamilton, Jure Leskovec, and Dan Jurafsky. Helge Holzmann, and Thomas Risse. 2012. Neer:
2016a. Cultural shift or linguistic drift? comparing An unsupervised method for named entity evolution
two computational measures of semantic change. In recognition. Proceedings of COLING 2012 pages
Proceedings of the 2016 Conference on Empirical 2553–2568.
Methods in Natural Language Processing. Associ-
ation for Computational Linguistics, pages 2116– Yating Zhang, Adam Jatowt, Sourav S Bhowmick,
2121. https://doi.org/10.18653/v1/D16-1229. and Katsumi Tanaka. 2016. The past is not a for-
eign country: Detecting semantically similar terms
William L Hamilton, Jure Leskovec, and Dan Juraf- across time. IEEE Transactions on Knowledge and
sky. 2016b. Diachronic word embeddings reveal Data Engineering 28(10):2793–2807.
statistical laws of semantic change. arXiv preprint
arXiv:1605.09096 .
Johannes Hellrich and Udo Hahn. 2017. Exploring Di-
achronic Lexical Semantics with J E S EM E. In
Proceedings of ACL 2017, System Demonstrations.
pages 31–36. http://aclweb.org/anthology/P17-
4006.
Nattiya Kanhabua and Kjetil Nørvåg. 2010a. Exploit-
ing time-based synonyms in searching document
archives. In Proceedings of the 10th annual joint
conference on Digital libraries. ACM, pages 79–88.
Nattiya Kanhabua and Kjetil Nørvåg. 2010b. Quest:
Query expansion using synonyms over time. In
José Luis Balcázar, Francesco Bonchi, Aristides
Gionis, and Michèle Sebag, editors, Machine
Learning and Knowledge Discovery in Databases.
Springer Berlin Heidelberg, Berlin, Heidelberg,
pages 595–598.
Yoon Kim, Yi-I Chiu, Kentaro Hanaki, Darshan Hegde,
and Slav Petrov. 2014. Temporal Analysis of Lan-
guage through Neural Language Models. Arxiv
pages 61–65. http://arxiv.org/abs/1405.3515.
Vivek Kulkarni, Rami Al-Rfou, Bryan Perozzi, and
Steven Skiena. 2015. Statistically significant de-
tection of linguistic change. In Proceedings of the
24th International Conference on World Wide Web.
ACM, pages 625–635.
Thomas Lansdall-Welfare, Saatviga Sudhahar, James
Thompson, Justin Lewis, FindMyPast Newspaper
Team, and Nello Cristianini. 2017. Content analy-
sis of 150 years of British periodicals. Proceedings
of the National Academy of Sciences 114(4):E457–
E465.
Jean-Baptiste Michel, Yuan Kui Shen, Aviva Presser
Aiden, Adrian Veres, Matthew K Gray, Joseph P
Pickett, Dale Hoiberg, Dan Clancy, Peter Norvig,
Jon Orwant, et al. 2011. Quantitative analysis of
culture using millions of digitized books. science
331(6014):176–182.
Terrence Szymanski. 2017. Temporal word analo-
gies: Identifying lexical replacement with di-
achronic word embeddings. In Proceedings of the
91
Using a Lexical Semantic Network for the Ontology Building
92
Proceedings of Recent Advances in Natural Language Processing, pages 92–101,
Varna, Bulgaria, Sep 2–4 2019.
didate concepts. They exploit such methods as for- sourcing methods and, in particular, games with a
mal concept analysis (Mondary, 2011) or knowledge purpose (GWAPS) such as JeuxDeMots3 and addi-
based methods (for instance, TextToOnto2 ). ”Lex- tional games . This commons sense network has
ical ontologies” (Abu Helou et al., 2014) are suc- been built since 2007. It is a directed, typed and
cessfully used for ontology building. Numerous ap- weighted graph. At the time of our writing, Rezo-
proaches targeted at high level ontology or informa- JDM contains 2.7 millions of terms that are modeled
tion retrieval ontology based on general knowledge as nodes of the graph and 240 millions of relations
(such as (Marciniak, 2013)) rely on PWN. Others (arcs).
use PWN and domain specific semantic lexicons for
The MLSN (Bebeshina-Clairet, 2019) is a multilin-
forming the concepts (Turcato et al., 2000). Many
gual LSN (it covers French, English, Spanish, and
other ontology learning techniques use distributional
Russian) with an interlingual pivot built for the cui-
semantics to learn lightweight ontologies, for exam-
sine and nutrition domain. This network is inspired
ple, (Wong, 2009).
by the RezoJDM in terms of its model. Structurally
In the framework of corpora based approaches to speaking, the MLSN is a directed, typed, and val-
the ontology building such as described in (Kietz uated graph. It contains k sub-graphs correspond-
et al., 2000), the idea of notable (salient, relevant) ing to each of the k languages it covers and a spe-
element or relevant piece of knowledge (RPK) has cific sub-graph which fulfills the role of the interlin-
been introduced. It corresponds either to the fre- gual pivot. Similar to the RezoJDM, we call terms
quent terms appearing in a corpus and to the tacit the nodes of the MLSN and relations - its typed,
knowledge contained in texts. Such tacit knowl- weighed, and directed arcs. The MLSN nodes may
edge corresponds to the semantic relationships (sub- correspond to one of the following types : lexi-
somption relationship and other specialized relation- cal items (garlic), interlingual items (pertaining
ships). Their presence in texts may take the form of to the interlingual pivot and also called covering
”indices”. In contrast, the explicit elements may re- terms), relational items (i.e. relationship reifica-
veal the presence of concepts. The main drawback tions such as salad[r has part]garlic), and category
such definition of RPKs is that it relies on the fea- items parts of speech or other morpho-syntactic fea-
tures defined or recorded for a particular language. tures (i.e. Noun:AccusativeCase).
In addition, statistical criteria are often preferred and
it is difficult to qualify such RPKs from the semantic As it has been difficult to set up the pivot using
point of view and in a language independent man- a multilingual embedding (joining multiple spaces,
ner. In this paper, we will detail the experiments we one per language) as well as to avoid pairwise align-
conducted to provide a new definition of the RPKs ment based on combinatorial criteria, the pivot has
based on the structured lexical semantic information been started as a natural one using the English edi-
and describe the way so defined RPKs can be used tion of DBNary (Sérasset, 2014). It incrementally
to help the top down ontology building process. evolves to become interlingual. The pivot evolution
relies on sense-based alignments between the lan-
3 Ressources guages of the MLSN and aims at taking into account
the difference of sense ”granularity” in different lan-
In the present section, we will describe the resources guages. For example, as stewin English can be trans-
we used in our experiments. lated as into French as pot-au-feu and ragoût. It
The RezoJDM (Lafourcade, 2007) is a lexical se- reflects the conceptualization discrepancy as ragoût
mantic network (LSN) for French built using crowd- refers to sauté the ingredients and then add water
2 3
https://sourceforge.net/p/texttoonto/wiki/ http://www.jeuxdemots.org
93
whereas pot-au-feu refers to boiling them together in
a larger amount of water than used for the ragoût. In
the MLSN pivot we have the interlingual term cor-
responding to stew (which covers the English term)
with two hyponyms corresponding to 0rench terms.
The alignments are progressively obtained through
external resources or by inference. Thus it can be Figure 1: MLSN: relationship annotation scheme.
considered as a union of word senses lexicalized or
identified in the languages covered by the MLSN.
. Even though we assume the pivot as being inter- graph) labels. At the time of our writing, the MLSN
lingual, it is still close to a natural one. A relation contains 821 781 nodes and 2 231 197 arcs. It covers
r ∈ R is a sextuplet r =< s, t, type, v, ls , lt > 4 languages : English, French, Russian, and Span-
where s and t correspond respectively to the source ish.
and the target term of the relation. The relation MIAM (Desprès, 2016)5 is a modular termino-
type is a typical relation type. It may model differ- ontology for the digital cooking. It provides knowl-
ent features such as taxonomic and part-whole re- edge necessary for the elaboration of general nutri-
lations (r isa, r hypo, r has part, r matter, r holo), tional suggestions. The knowledge model of this
possible predicate-argument relations (typical object ontology gathers expert knowledge on food, food
r object, location r location, instrument r instr of an transformation, cooking actions, relevant dishes that
action), ”modifier” relations (typical characteristic reflect french culinary tradition, recipes necessary
r carac, typical manner r manner) and more4 . The to cook such dishes. MIAM contains about 7
relationship valuation v corresponds to the charac- 000 nodes and 30 000 semantic (non subsump-
teristics of the relation which are its weight, con- tion/ontological is-a) relations.
fidence score, and annotation. The relation weight
may be negative in order to model noise and keep the 4 Method: Immersion - Projection
information about erroneous relations easy to access
programmatically so they could not affect the infer- 4.1 Summing up the Method
ence processes. The confidence score is a score at- Our method is built upon the idea of projecting a
tributed to a particular data origin (external resource, model (the MIAM model) onto a multilingual or
inference process). In practice, this feature is an ar- monolingual LSN (respectively MLSN and Rezo-
ray as different origins may provide the same rela- JDM) in order to extract an intermediary resource
tion. The confidence information is provided as an that can be used by ontology or domain experts in
argument to the function that maps from some ex- the scope of information retrieval or validation of the
ternal knowledge resource to the MLSN. In case of automatically suggested pieces of knowledge.
relation calculated by an inference process, it cor-
responds to the precision evaluated on a sample of Our method differs from others by the definition of
candidate relations returned by this process. To an- the RPK and by the use of a non ontological seman-
notate a relation we add a complementary informa- tically structured resource for ontology building. We
tion that allows qualifying this relation. The figure 1 define RPK as follows : ”a relevant piece of knowl-
details and exemplifies the annotation scheme. edge is either a term or a relation or a semantic
structure which is known as qualified and qualify-
The labels ls and lt correspond to the language (sub- ing”. Qualified refers to the possibility to describe
4
We also introduced more specific relation types such as r entailment, 5 http://www-limics.smbh.univ-paris13.fr/
r cause, r telic role, r incompatible, r before, r after etc. ontoMIAM/
94
the RPK in a discrete way (i.e. by enumerating the output is the action of inferring terms and relations
typed relations). If the RPK is a term, it needs to in the target LSN (MLSN, RezoJDM).
have a high in-degree (which reveals its conceptual
In their general form, the mapping rules state: ”If
role as it is used to define other terms of the net-
x and y are respectively domain and range of an
work). If the RPK is a relation, it needs to be con-
Object Property p of the ontology to be immersed
textualized (through the annotation mechanism rep-
and y is a subclass of C, then x has a relation R
resented on the figure 1 or through the constraints
with y and y has a relation is-a with C in the re-
put on source and/or target terms of the relation). If
ceiving (target) LSN”. Such rules have been de-
the presumed RPM is a graph structure (path, sub-
fined for the multilingual experiment for two rea-
graph), it needs to possess a certain number of oc-
sons. First, for each of the 93 MIAM proper-
currences in the network. Qualifying refers to the
ties, we determined relevant MLSN semantic re-
possibility to use the candidate RPK for endogenous
lation types (or set of types). Thus, the Ob-
inference process. If the RPK is a term, it needs
ject Property aPourProduitInitial (hasIni-
to have hypernyms, hyponyms, synonyms among its
tialProduct) corresponds to the substance and part-
neighbours. It has to be aligned with other terms
whole meronymy (MLSN relations typed r has part
pertaining to the other languages of the LSN (if such
and r matter). Second, we mapped the ontology
LSN is multiligual). If the RPK is a relation, it must
labels to the MLSN terms by coincidence (3 930
not be unique (other real or potential6 relations must
terms; i.e. poulet basquaise formally denotes a
exist in the network). If the candidate RPK is a struc-
MIAM concept, a lexical item with the same la-
ture, its terms and relations must be qualifying.
bel already exists in the MLSN) or by composition
Here we detail the experiments that have been con- (4 135 terms, i.e.: unité mesure capacité doesn’t
ducted on the basis of the MLSN in order to pro- correspond to any existing MLSN term because it
pose ”pseudo-class” and ”pseudo-property” candi- doesn’t correspond to any commonly used colloca-
date RPKs to enhance the MIAM ontology and those tion in French; this label is split and integrated into
concerning the enrichment of an ontology draft us- MLSN with the semantic relations that link its parts.
ing the monolingual LSN, RezoJDM (Lafourcade,
As part of the monolingual experiment, 115 descrip-
2007). These experiments rely on lexical knowl-
tors have been automatically expressed in French
edge. Therefore, the resulting RPKs have no pre-
on the basis of their Uniform Resource Identifier
tension to the ontological validity. The decision per-
(URI) strings. All the terms except one were al-
tains to the human expert.
ready present in the RezoJDM network. We ex-
ploited the relations typed r carac (typical charac-
4.2 Immersion
teristic) for this experiment. This relation has been
The projection of a given ontology model onto a annotated using the URIs of the ontology properties
LSN starts by the immersion of such model. The aPourDescripteurBruit (hasSoundDescrip-
immersion mechanism uses a set of manually de- tor is immersed as follows: croûte r carac :: bruit
fined mapping rules. It is possible to generate them croustillant (”croûte has typical characteristic linked
automatically for the ontological resources that ex- to the noise croustillant”7 ). The premises of the
ploit standard vocabularies (such as RDFS, SKOS mapping rules rely on the contextualization of the
and other machine readable formats). The input of LSN relations. Such contextualization is possible
the immersion algorithm is the reference ontology when using sets of hypernyms and neighborhood se-
(MIAM) and the set of mapping rules whereas its mantic relations of the source and target terms of a
6 7
Relations that can be calculated using inference. crust has typical characteristic linked to the sound crusty
95
relation. Meta-information attached to the LSN re- poultry”). This analysis allowed selecting a subset
lations (annotations, weight) may also be used. For of relation types to consider during the experiment.
instance: pétrir r object pâte ∧ pétrir r isa tech- The RPK(ci) inference includes two steps: valida-
nique de base ∧ pâte r isa préparation8 ). tion of the hierarchical chain (figure 2) and RPK(ci)
As part of the immersion process the ontology labels candidate suggestion.
become LSN terms that can be polysemous.
4.3 Projection
4.3.1 Inference in the LSN Context
In the MLSN context we set up several algorithms Figure 2: Hierarchy chain validation. Ti are the terms of the
to discover relevant pieces of knowledge (RPK) of hierarchy chain. We check by triangulation their semantic
the types ”class/individual” (ci) and ”ontology prop- relatedness and use a subset of relations types (such relations
erty”(op). To discover the RPK(ci) we compare are noted R) for that.
the neighborhood terms inside an hierarchical chain
which goes up to a high level MIAM concept im- We calculated and validated hierarchical chains cor-
mersed into the LSN. For the RPK(op) we look for responding to 1 322 top MIAM concepts pertaining
(real or possible) MLSN relations similar to the im- to the Aliment module. First we obtained 132 213
mersed MIAM properties. The inference scheme we chains. After filtering them by weight, the set has
use is the abduction scheme. When we have two been reduced to 53 749 chains (40% of the initial
similar terms (such as cohyponyms) the relations de- set). Still, a certain number of redundancies may ex-
tained by one of them can be proposed for the other. ist inside this statistically pre-filtered set since a long
For a term T, the abduction implies selecting a set of chain may include several shorter ones. The logical
similar terms (according to some criteria) in order to filtering by triangulation left us with a set containing
propose the relations detained by those similar terms 9 600 hierarchical chains (18% of the number of sta-
to T . tistically filtered chains and 7% of the initial number
of candidates).
4.3.2 Discovering the RPK(ci)
Hierarchical chain examples after filtering:
In MIAM, the general axioms concern the disjunc- (1) ”baguette complète→pain complet →
tion between the MIAM classes which is the basis of pain→ingrédient de recette de cuisine→aliment”
the ontology consistence. To translate them in terms (2) ”angélique→confiserie→bonbon”.
of MLSN, we considered the labels of the classes RPK(ci) examples:
listed in the axioms in order to identify the crite- truffe chocolat subClassOf chocolat
ria that could have determined the disjunction. We pomme à cidre subClassOf pomme
manually analyzed a subset of the MIAM axioms sucre de pomme subClassOf confiserie
and came up with the following criteria: affiliation The analysis and validation of the hierarchical
(r has part i.e. a specific label (organic)), trans- chains corresponds to the important memory load.
formation (r carac i.e. boiled mixture, cubed veg- The complexity of the algorithm depends on the im-
etable), composition (r matter i.e. produit à base de portance of the MIAM concept being processed as
poisson ”fish based product”), category based dis- well as on the length of the hierarchical chains that
tinctions (r hypo i.e. volaille type dinde ”turkey type are considered. The use of a subset of the seman-
8
knead r object dough ∧ knead r isa basic technique ∧ dough r isa tic relations types available in the MLSN reduces
mixture the number of combinations to process. Thus, given
96
that the highest in-degree typed r isa in the French The MIAM ontology we try to enrich counts 21 565
sub-graph of the MLSN is 5 264 (for the term al- instances of Object Properties. Once they are im-
iment) and that the maximum length l = 9, the mersed into the MLSN, one could consider that we
complexity in the worst case would be O(dlisa or have the same number of inference rule instances
O(5 2649 ) = 3, 103436942 × 1033 . that can be used for the cross-lingual RPK(op) dis-
covery. A naive approach would be setting up a
The table 1 introduces the results obtained for the
cross-lingual inference mechanism. However, such
discovery of RPK(ci) related to the top level concept
approach would be error-prone due to the poten-
Aliment in the French sub-graph of the MLSN. The
tial alignment and polysemy issues. In addition,
#candidates #valid %valid #new %new as MIAM has been built according to the top-down
11 520 11 289 98% 4 741 42% methodology by a community of domain experts, it
contains a variable number of instances per property.
Table 1: The RPK(ci) discovery. The naive approach would reiterate this imbalance.
To refine the RPK(op) discovery, we experimented
automatic evaluation of the proposed RPK(ci) would a rule based approach. First, the validity of the rule
mostly rely on similarity measures. However, the for the source language is calculated. Second, struc-
projection step implicitly relies on relatedness and tures similar to those specified by the rule are being
similarity between the LSN terms. Thus, in our fu- discovered in the MLSN (in other languages).
ture work, the RPK(ci) evaluation will need human
The rule has the following form:
expert decisions.
property=aPourEtatPhysique
(property name)
4.3.3 The RPK(op) Discovery src=?s (domain, set of terms)
The RPK(op) discovery seems to be particularly reltype=r_carac (relation type)
useful in the context of multilingual ontology tgt=?o(range, set of terms)
building or localization of an existing ontology. source_isa=aliment(src hypernyms)
In our experience, each module of the MIAM target_isa=etat physique
ontology has its own hierarchy of properties. While (tgt hypernyms)
immersing them into the MLSN, these properties annotation=int:physical state
have been expressed in terms of semantic relations (meta-information)
contextualized using annotations. The choice of src_feat=OUT/r_pos/int:Noun
the MLSN semantic relation type made for these (in and out relations that characterize terms in the
properties allows us to distinguish the following source set)
cases for the MIAM Object Properties (OPs): tgt_feat=OUT/r_pos/int:Adj (in and out
composition based (aPourProduitInitial, relations that characterize terms in the target set)
”hasInitialProduct”); related to processes ( If a given rule allowed detecting enough structures
aPourMethodeDeConservation, ”has- in the source language (at least, 2 structures), it is
ConservationMethod”); temporal and spatial considered as a valid one and can generate the qual-
relation based (aPourMoisPrimeur, ”hasEar- ifying object. Thanks to this object, candidate struc-
lyMonth”); characteristic based (aPourEtat, tures are detected in the other language sub-graphs
”hasState”); OPs with a specific sub-graph9 of the MLSN. The mechanism of RPK(op) discov-
(aPourAlimentAmi, ”hasFriendlyFoodItem”). ery reveals the following elements that allowed dis-
covering new pieces of knowledge : possibly anno-
9
A subset of MLSN terms connected through semantic relations. tated semantic relation (case of the properties such
97
as aPresenceLactose, aPresenceGluten); 4.3.4 Towards the Automatic Suggestion of the
specific pattern (defined by rules); complex struc- RPKs(op)
ture for properties related to processes. The results
To extend the RPK discovery experiment, we tried
obtained using possibly annotated relations(in par-
to automatically suggest pseudo ontology proper-
ticular, Data Properties)) are presented in the table
ties to be submitted to the domain and ontology hu-
2. the potential improvement is estimated as an in-
man experts. We considered the ontology Senso-
crease compared to the initial number of property
MIAM10 for this experiment. This ontology is a
instances (impr.%).
MIAM module but we considered it as a ”draft” on-
tology as the sensory aspect modeling is a flourish-
ing research and development area and the Senso-
#DP #MIAM #RPK(op) filt. +%
aTeneurLipide 0 4 741 3 271 - MIAM could be improved. We used the monolin-
aPresenceLactose 2 593 530 408 +16% gual LSN (RezoJDM). The SensoMIAM contains
aPresenceGluten 289 820 762 +263% sensory descriptors such as DescripteurTact
(”TactileDescriptor”) = {astringent, filandreux,
Table 2: The RPK(op) discovery on the basis of simple se- ..., nerveux}; DescripteurSubstance (”Sub-
mantic relations. stanceDescriptor”)= {aéré, dense, . . . , épais}
To calculate the RPK descriptors RPK(desc), we
explored the semantics of the source terms of the
Example of the output of the rule-based algorithm: relations typed r carac. If the set of outgoing
ru:jarkoje aPourProduitDiscriminant podlivka relations of such terms connects them to a food
(”stew hasDiscriminatingProduct sauce”) and item and if they have a set of typical character-
en:stew aPourProduitInitial en:vegetable (produce) istics shared with other terms with an hypernym
(”stew hasInitialProduct vegetable”). Our rule- ≈ ”food”, the target term of their outgoing rela-
based RPK(op) experiment (given the actual state of tions typed r carac that is not present in the Senso-
the MLSN) yielded the results listed in the table 3. MIAM can be suggested as a potential RPK(desc).
The relation typed r carac is annotated. The pro-
cess is represented on the figure 3. The experi-
Fully automated structure-based evaluation such as ence allowed to suggest the RPK(desc) such as:
described in (Fernández et al., 2009) may be cho- DescripteurArome={sucré-salé, miellé, . . . ,
sen to compare to other resources available on the vinaigré} or DescripteurTact = {écailleux,
Web such as (Dooley et al., 2018). To address the spongieux,. . . , floconneux}.
ontology accuracy, completeness, conciseness, ef-
ficiency, consistency, and other features (Raad and We automatically suggested and semi-automatically
Cruz, 2015), a combination of methods is needed. In validated 342 RPKs(desc). We explored the pos-
particular, gold standard ontology, specific tasks and sibility of suggesting relevant RPKs to human ex-
corpora may be used for evaluation. A task-based perts. We defined 3 pseudo-properties for test-
evaluation such as semantic analysis (Bebeshina- ing: aPourComposantFlaveur (”hasFlavourCompo-
Clairet, 2019), dietary conflict detection from dish nent”), aPourComposantToucher (”hasTouchCom-
titles (Clairet, 2017) have been used for the MLSN. ponent”), and aPourComposantAspect (”hasAspect-
To evaluate the output of the immersion- projection Component”). To populate them, we explored the
method, we need to organize our triples into a fully RezoJDM relations typed r has part and r matter
structured ontology. This will be one of the priorities and considered the characteristics that can be shared
of our future work. 10
www-limics.smbh.univ-paris13.fr/sensoMIAM/
98
m prop #in en fr es ru #out %aug
A aPourProdInit. 2031 292 1208 203 2 245 3 039 +149%
A aPourEtatPhys. 543 30 29 10 53 85 +16%
A aPourForme 39 77 78 5 37 132 +338%
A aPourLabel 114 15 11 3 1 29 26%
A aPourMethodC. 115 94 101 13 156 309 +269%
A aPourMois 116 117 221 23 28 116 +288%
A aPourRegion 289 98 71 2 57 216 +75%
A aPourProdCon. 98 256 302 143 103 570 +582%
A aPourProdInitialA. 41 94 147 12 567 259 +633%
P aPourTypeDeCuis. 23 155 124 80 285 686 +2 986%
P aPourDomCul. 82 112 92 120 1313 1276 +1 557%
P aPourDecoupe 82 82 78 56 77 272 +332%
S aPourSaveur 752 51 78 47 98 232 +31%
S aPourDescripteurBr. 119 67 80 10 6 159 +134%
S aPourCouleur 233 192 451 59 423 911 +391%
S aPourAspectSurf. 176 40 35 12 52 101 +58%
S aPourSensationT. 54 84 77 21 12 155 +287%
- Total 5388 2384 3960 937 4953 9531 +177%
Table 3: Rule-based approach. m -name of module (Aliment (A), Preparation (P), Sensory (S)), prop - property, #in -
MIAM triples, en, fr, es, ru - MLSN sub-graphs. #out - overall number of suggested RPK(op) after filtering, %aug -
potential improvement.
99
References Mathieu Lafourcade. 2011. Lexique et analyse
sémantique de textes - structures, acquisitions, calculs,
Mamoun Abu Helou, Mustafa Jarrar, Matteo Palmonari, et jeux de mots. (Lexicon and semantic analysis of texts
and Ch Fellbaum. 2014. Towards building lexical ontol- - structures, acquisition, computation and games with
ogy via cross-language matching. pages 346–354. words). https://tel.archives-ouvertes.fr/tel-00649851.
Keith Allan. 2001. Natural Language Semantics. Black-
well. Ora Lassila and Deborah McGuinness. 2001. The role of
frame-based representation on the semantic web. Techni-
Nadia Bebeshina-Clairet. 2019. Construction d’une
cal report, Knowledge Systems Laboratory Report KSL-
ressource termino-ontologique multilingue pour les do-
01-02, Stanford University, Stanford (USA).
maines de la cuisine et de la nutrition. Theses, Université
Paris 13. Jacek Marciniak. 2013. Building wordnet based ontolo-
Christian Biemann. 2005. Ontology learning from text: gies with expert knowledge. In LTC.
A survey of methods. LDV Forum 20:75–93. Thibault Mondary. 2011. Construction d’ontologies
Nadia Clairet. 2017. Dish classification using knowledge à partir de textes. L’apport de l’analyse de con-
based dietary conflict detection. In RANLP 2017. cepts formels.. Theses, Université Paris-Nord - Paris
Sylvie Desprès. 2016. Construction d’une ontolo- XIII. Equipe RCLN. https://tel.archives-ouvertes.fr/tel-
gie modulaire. application au domaine de la cuisine 00596825.
numérique. Revue d’Intelligence Artificielle 30(5):509– Joe Raad and Christophe Cruz. 2015. A Survey on On-
532. https://doi.org/10.3166/ria.30.509-532. tology Evaluation Methods. In Proceedings of the Inter-
Zhendong Dong, Qiang Dong, and Changling Hao. national Conference on Knowledge Engineering and On-
2010. Hownet and its computation of mean- tology Development, part of the 7th International Joint
ing. In Proceedings of the 23rd International Conference on Knowledge Discovery, Knowledge Engi-
Conference on Computational Linguistics: Demon- neering and Knowledge Management . Lisbonne, Portu-
strations. Association for Computational Linguistics, gal. https://doi.org/10.5220/0005591001790186.
Stroudsburg, PA, USA, COLING ’10, pages 53–56. Lionel Ramadier. 2016. Indexation and learning of terms
http://dl.acm.org/citation.cfm?id=1944284.1944298. and relations from reports of radiology. Theses, Uni-
Damion Dooley, Emma Griffiths, Gurinder S. Gosal, versité de Montpellier. https://hal-lirmm.ccsd.cnrs.fr/tel-
Pier Luigi Buttigieg, Robert Hoehndorf, Matthew 01479769.
C. Lange, Lynn M. Schriml, Fiona S. L. Brinkman, and Christophe Roche. 2007. Le terme et le concept : fonde-
William W. L. Hsiao. 2018. Foodon: a harmonized ments d’une ontoterminologie. In TOTh 2007 : Ter-
food ontology to increase global food traceability, qual- minologie et Ontologie : Théories et Applications. An-
ity control and data integration. npj Science of Food 2. necy, France, pages 1–22. 22 pages. https://hal.archives-
https://doi.org/10.1038/s41538-018-0032-6. ouvertes.fr/hal-00202645.
Miriam Fernández, Chwhynny Overbeeke, Marta Sabou,
Gilles Sérasset. 2014. DBnary: Wiktionary as a Lemon-
and Enrico Motta. 2009. What makes a good ontology? a
Based Multilingual Lexical Resource in RDF. Semantic
case-study in fine-grained knowledge reuse. In Asunción
Web – Interoperability, Usability, Applicability pages –.
Gómez-Pérez, Yong Yu, and Ying Ding, editors, The Se-
To appear. https://hal.archives-ouvertes.fr/hal-00953638.
mantic Web. Springer Berlin Heidelberg, Berlin, Heidel-
berg, pages 61–75. Robyn Speer and Catherine Havasi. 2012. Represent-
E. Gaillard, J. Lieber, and E. Nauer. 2015. Improv- ing general relational knowledge in ConceptNet 5. In
ing ingredient substitution using formal concept analysis Proceedings of the Eighth International Conference on
and adaptation of ingredient quantities with mixed linear Language Resources and Evaluation (LREC-2012). Eu-
optimization. In Computer Cooking Contest Workshop. ropean Languages Resources Association (ELRA), Istan-
Frankfort, Germany. https://hal.inria.fr/hal-01240383. bul, Turkey, pages 3679–3686.
J.U. Kietz, A. Maedche, and R. Volz. 2000. A method for Sylvie Szulman. 2012. Logiciel Terminae - Ver-
semi-automatic ontology acquisition from a corporate in- sion 2012. TERMINAE est une plateforme d’aide
tranet. EKAW-2000 Workshop Ontologies and Text, Juan- à la construction de ressources termino-ontologiques
Les-Pins, France, October 2000 . à partir de ressources textuelles. https://hal.archives-
Mathieu Lafourcade. 2007. Making people play for Lex- ouvertes.fr/hal-00719453.
ical Acquisition with the JeuxDeMots prototype. In Andon Tchechmedjiev. 2016. Semantic Interoperability
SNLP’07: 7th International Symposium on Natural Lan- of Multilingual Lexical Resources in Lexical Linked Data.
guage Processing. Pattaya, Chonburi, Thailand, page 7. Theses, Université Grenoble Alpes. https://tel.archives-
https://hal-lirmm.ccsd.cnrs.fr/lirmm-00200883. ouvertes.fr/tel-01681358.
100
Davide Turcato, Fred Popowich, Janine Toole, Dan Fass,
Devlan Nicholson, and Gordon Tisher. 2000. Adapting
a synonym database to specific domains. In ACL-2000
Workshop on Recent Advances in Natural Language Pro-
cessing and Information Retrieval. Association for Com-
putational Linguistics, Hong Kong, China, pages 1–11.
https://doi.org/10.3115/1117755.1117757.
Piek Vossen. 2012. Ontologies. The Ox-
ford Handbook of Computational Linguistics
https://doi.org/10.1093/oxfordhb/9780199276349.013.0025.
Wilson Wong. 2009. Learning lightweight ontologies
from text across different domains using the web as back-
ground knowledge. Ph.D. thesis.
Manel Zarrouk. 2015. Endogeneous Consolidation of
Lexical Semantic Networks. Theses, Université de Mont-
pellier. https://hal-lirmm.ccsd.cnrs.fr/tel-01300285.
101
Naive Regularizers for Low-Resource Neural Machine Translation
Meriem Beloucif1 , Ana Valeria Gonzalez2 , Marcel Bollmann2 , and Anders Søgaard2
1
Language Technology Group, Universität Hamburg, Hamburg, Germany
2
Dpt. of Computer Science, University of Copenhagen, Copenhagen, Denmark
102
Proceedings of Recent Advances in Natural Language Processing, pages 102–111,
Varna, Bulgaria, Sep 2–4 2019.
Two of the regularizers are derived directly from such as the European Parliament data1 , to regu-
the input, without relying on any additional lin- larize the learning helps improve the translation
guistic resources. This makes them adequate for quality (Miceli Barone et al., 2017). This ap-
low-resource settings, where the availability of lin- proach produces good results, but it is not appli-
guistic resources can generally not assumed. Our cable in low-resource settings because it requires
third regularizer (frequency) only assumes access large amounts of data in the language of interest.
to unlabeled data. (b) We show that regulariz- To the best of our knowledge, our work is the first
ing a standard NMT architecture using naive regu- to introduce naive, linguistically motivated regu-
larization methods consistently improves machine larization methods such as sentence length, punc-
translation quality across multiple low-resource tuation and word frequency.
languages, also compared to using more standard
methods such as dropout. We also show that com- 3 Model Description
bining these regularizers leads to further improve- 3.1 Baseline
ments. (c) Finally, we present examples and analy-
In order to show the impact that our regulariz-
sis showing how the more linguistically motivated
ers have on the translation quality, we use an
regularizers we propose, help low-resource ma-
off-the-shelf NMT system described by Luong
chine translation.
et al. (2017) as our baseline. The model con-
sists of two multi-layer recurrent neural networks
2 Related Work
(RNNs), one that encodes the source language and
End-to-end neural machine translation is based on one that decodes onto the target language. For
encoder–decoder architectures (Sutskever et al., the encoder cell, we use a single Long Short-Term
2014; Cho et al., 2014; Luong et al., 2015a, 2017), Memory (LSTM) layer (Hochreiter and Schmid-
in which a source sentence x = (x1 , x2 , ..., xn ) huber, 1997) and output the hidden state, which
is encoded into a vector (or a weighted average then gets passed to the decoder cell.
over a sequence of vectors) z = (z1 , z2 , ..., zn ). We train our models to minimize the cross-
The hidden state representing z is then fed to the entropy loss and back-propagate the loss to up-
transducer (also called decoder) which generates date the parameters of our model. We update net-
translations, noted as y = (y1 , y2 , ..., ym ). work weights using Adam optimization (Kingma
Neural machine translation has achieved state- and Ba, 2014), which calculates the exponential
of-the-art performance for various language pairs moving average of the gradient and squared gradi-
(Luong et al., 2015a; Sennrich et al., 2015; Lu- ent, and combines the advantages of AdaGrad and
ong and Manning, 2016; Neubig, 2015; Vaswani RMSProp. For the purpose of comparison, we set
et al., 2017), especially when trained on large vol- the dropout to 0.2, similar to Luong et al. (2015b).
umes of parallel data, i.e., millions of parallel sen-
3.2 Regularized NMT
tences (also called bi-sentences), humanly trans-
lated or validated. Such amounts of training data, To apply our new regularizers, we add each reg-
however, are difficult to obtain for low-resource ularizer to the loss function during the training of
languages such as Slovene or Vietnamese, and in the NMT model (Luong et al., 2015a; Luong and
their absence, neural machine translation is known Manning, 2016; Luong et al., 2017). Since we
to come with diminishing returns, suffering from aim to minimize the cross-entropy loss, this means
overfitting (Koehn and Knowles, 2017). that we favor training instances which have a low
In order to avoid overfitting, NMT models are penalty from the regularizers (e.g., a small length
often trained with L1 or L2 regularization, as well difference). Importantly, we do not use dropout
as other forms of regularization such as momen- in this scenario, as we want to contrast our naive,
tum training or dropout (Srivastava et al., 2014; but linguistically motivated signals with a tradi-
Wang et al., 2015; Miceli Barone et al., 2017). tional, but not linguistically motivated regulariza-
However, these regularization methods are very tion method, i.e., dropout.
general and do not carry any language specific in- Furthermore, we do not explore alternative
formation. ways for adding regularizers to the loss func-
On the other hand, it has been shown that trans- tion here (other alternatives could be to have a
fer learning approaches using out of domain data, 1
http://www.statmt.org/europarl/
103
weighted penalty which is then tuned to find the methodologies such as punctuation-based align-
best penalty and added to the loss function for test- ment (Chuang et al., 2004). Our second regular-
ing). The main purpose of this work is to study izer is based on this simple idea, as it penalizes
the effect of naive linguistically motivated regu- training instances where the quantities of punctu-
larizers and show that they can improve translation ation marks differ between input and MT output
quality; we leave it to future work to find the op- sentences. Example (2) is taken from the training
timal configuration of regularizers that maximizes set of the French–English translation task:
the overall translation quality.
(2) I N Pas parce qu’ils sont moins bons, pas
4 Naive Regularizers parce qu’ils sont moins travailleurs.
R EF And it’s not because they’re less smart,
4.1 Length-Based Regularizer and it’s not because they’re less diligent.
NMT models have shown to suffer “the curse of O UT And . . . . . . . . . . . . .
sentence length”, and it has been hypothesized that
We note that the punctuation in the French input
this is due to a lack of representation at the de-
sentence matches the punctuation of the desired
coder level (Cho et al., 2014; Pouget-Abadie et al.,
English reference. However, during an early train-
2014). Our proposed sentence-length-based regu-
ing step, the NMT model translates the input to a
larizer penalizes relative differences between the
sequence containing six times the number of punc-
input and the MT output lengths during the train-
tuation marks in the input sentence, which is ob-
ing of the NMT model:
viously incorrect. Our punctuation regularizer fur-
ther penalizes examples like this one.
reglength = |l0 − l1 | (1)
4.3 Frequency-Based Regularizer
Here, l0 and l1 represent the input sentence and the
MT output sentence lengths, respectively, as mea- Our last regularizer is based on the distribution of
sured by the number of words (not to be confused word frequencies between the source and the tar-
with L1 and L2 regularization methods). get sentences. Generally speaking, if the source
Note that this regularizer is different from the sentence contains an uncommon word, we assume
word penalty feature in phrase-based machine that its translation in the target language is also
translation (Zens and Ney, 2004), which only pe- uncommon. The intuition behind this regularizer
nalizes the target sentence length. The relative dif- is that if the source sentence contains one uncom-
ference between the input and the MT output sen- mon word and three common words, then its ac-
tence lengths is also used as a feature in Marie and curate translation should contain similar word fre-
Fujita (2018). quencies. The example below is extracted from
the English–French translation task:
4.2 Punctuation-Based Regularizer
(3) I N But now there is a bold new solution to
The punctuation-based regularizer penalizes train- get us out of this mess.
ing instances whenever the amount of punctua- R EF Mais il exist une solution audacieuse
tion marks in the input sentence differs from the pour nous en sortir.
amount in the MT output sentence. It is computed O UT Mais maintenant il y a une solution pour
as follows: nous en sortir.
regpunct = |p0 − p1 | (2) The English sentence contains the frequent word
there and the less frequent word bold. The French
Here, p0 and p1 is the total number of punctuation output sentence is acceptable, but it is not accu-
marks in the input and the MT output sentence, rate since the English word bold (audacieuse in
respectively. the reference translation) was omitted in the out-
Unfortunately, the only available methods to put. During training, the frequency regularizer pe-
generate more efficient NMT models have in- nalizes such cases that have a big divergence be-
cluded data intensive methods such as sentence tween the word frequencies in the input and output
alignment (Bahdanau et al., 2014). Some very sentences.
early research done in alignment used simple The purpose of our frequency-based regularizer
104
Languages #Words Languages Sentence Pairs
Czech 1.7M Train Development Test
English 85.57M Czech 122,382 480 1,327
French 55.72M French 232,825 890 1,210
German 35.47M German 206,112 888 1,080
Russian 2.5M Russian 178,165 887 1,701
Slovene 1.45M Slovene 17,125 1,144 1,411
Vietnamese 3.5M Vietnamese 133,317 1,553 1,268
Table 1: The size of the Wikipedia dumps Table 2: The size of the training data in sentence
(#words) used to calculate word frequencies for pairs. To test our proposed models, we experi-
each language. ment by translating to/from English for every non-
English language.
is to calculate how different the MT output sen-
tence is from the source input in terms of vocab-
O UT C’était une femme forte portant une
ulary distribution. For instance, the frequency of
fourrure autour du cou
using the word chauve-souris in French is almost
similar to the frequency of using its English trans- Example (4) shows an input sentence and its MT
lation bat in English. The same could be applied output, for which we would compute the fre-
for the more frequent words such as et in French quency vectors as follows:
and its English translation and.
We start by computing the frequency vectors −
v→
in = hf (‘it’), f (‘was’), . . . ,f (‘neck’)i
−
v→ −→ − →
in and vout , containing the frequency for every vout = hf (‘c’était’), f (‘une’), . . . ,f (‘cou’)i
word wi in the input and MT output sentence, re-
spectively:
5 Experiments
→
−
v = hf (w1 ), . . . , f (wn )i (3)
5.1 Data
To calculate the word frequencies f (w) for each The purpose of our experiments is to show that sig-
language, we use the Wikipedia database2 as an nals such as sentence length, punctuation or word
external resource. Table 1 contains the size of frequency help improve the translation quality of
the datasets (in number of words) used to estimate a standard neural machine translation architecture.
these. We note that there is considerably more data To that effect, we experiment with 12 translation
for English and French than for e.g. Vietnamese tasks, translating from English to six low-resource
(cf. Table. 1); we discuss the effect that this might languages, and vice versa.
have on the results in Sec. 6. The six languages represent the following lan-
We interpret the resulting frequency vectors → −
v guage families: Slavic, Romance, Germanic, and
as distributions, for which we now calculate the Austro-Asian. We further vary the size of the
Kullback-Leibler (KL) divergence to obtain our training data to test how our regularization meth-
regularization term: ods affect the quality of the MT output in differ-
ent setups. Table 2 contains the size of the train-
regfreq = DKL (−
v→ −→
in , vout ) (4) ing, development and test set for every language
Essentially, this regularizer penalizes transla- pair. Note that the training sets vary considerably
tions if their word frequency distributions diverge in size, from 17k sentence pairs for Slovene to al-
too strongly from those of the source sentence. most 233k for French.
The data is from the International Workshop
(4) I N It was a big lady who wore a fur around on Spoken Language Translation (IWSLT), ex-
her neck cept for Russian, Slovene and Vietnamese which
R EF C’était une dame forte qui portait une are from IWSLT 2015, the data for the remain-
fourrure autour du cou ing translation tasks is from IWSLT 2017 (Cettolo
2
https://en.wikipedia.org/wiki/Wikipedia:Database et al., 2012).
105
Preprocessing The purpose of our experiments System BLEU
is to learn how to efficiently translate low-resource
Luong et al. (2015) 23.30
languages. For that purpose, we do not use any
Luong et al. (2017) (dropout=0.2) 25.10
advanced preprocessing for any of our translation
Baseline (dropout=0.2) 26.43
tasks except tokenization where we use the script
+ Length 26.77
from the Moses toolkit (Koehn et al., 2007). We
+ Punct 26.71
also set the maximum sentence length to 70 tokens
+ Frequency 26.12
and the vocabulary size to 50k.
+ Combined 27.13
5.2 Training Details
Table 3: Baseline vs. our proposed models on
We use the attention-based model described in Lu- the English–Vietnamese translation task, using the
ong et al. (2015b). Our model is composed of two same dataset as Luong et al. (2015b). The re-
LSTM layers each of which has 512-dimensional sults in bold represent statistically significant re-
units and embeddings; we also use a mini-batch sults compared to the baseline according to Mul-
size of 128. Adding an attention mechanism in tEval (Clark et al., 2011).
neural machine translation helps to encode rele-
vant parts of the source sentence when learning the
model. We propose to add additional regularizers Translation Task #Steps
on top of the attention-based model at each trans- Lang→English
lation step.
We have noticed that the convergence highly de- Czech 12K
pends on the language pairs involved. While our French 20K
baseline model is identical to the NMT model de- German 20K
scribed by Luong et al. (2015b), we deviate from Russian 22K
their training procedure by continuing the training Slovene 10K
until convergence, which for us took 15 epochs in- Vietnamese 15K
stead of the 12 epochs described by the authors. English→Lang
The convergence in our case is measured by the
models having no improvements on the develop- Czech 12K
ment set over five epochs. French 22K
German 20K
Table 3 shows that our baseline is +1.5 BLEU
Russian 18K
points better than the scores reported by Luong
Slovene 11K
et al. (2015b). On top of that, our length-based and
Vietnamese 15K
punctuation-based models produce a statistically
significant improvement over the baseline (+0.5
Table 4: Number of steps it took until the models
BLEU points).
stopped improving for all the translation tasks.
We train all our models automatically until con-
vergence. In Table 4, we report the number of
epochs it took to converge by translation task gram based metrics BLEU (Papineni et al., 2002)
when translating to/from English. We note that and METEOR (Banerjee and Lavie, 2005), as well
except for Czech and Slovene, which converged as the error-rate based metric TER (Snover et al.,
the quickest, most of the translation tasks took be- 2006). The evaluation metric BLEU (Papineni
tween 15k and 20k steps to converge. et al., 2002) is based on n-gram matching between
the input and the output, whereas the error-rate
6 Evaluation based metric TER (Snover et al., 2006) measures
In order to show that the naive regularizers which how many edits are needed so that the machine
we propose in this paper significantly boost the translation resembles the man-made reference.
translation quality, we test the machine transla-
6.1 Results
tion output using the toolkit MultEval defined in
Clark et al. (2011). In this paper, we report the Table 5 shows the results for all language pairs and
results using three commonly used metrics: the n- all metrics. We observe an improvement over the
106
System Languages
Czech French German Russian Slovene Vietnamese
Baseline 14.01 32.13 22.07 12.87 5.60 26.43
Length 14.65 32.32 21.64 12.81 4.98 26.77
EN→Lang
Punct 14.98 32.79 22.89 13.06 5.64 26.71
Frequency 14.75 33.47 22.14 13.50 1.95 26.12
Baseline 21.32 31.51 24.41 15.39 8.85 24.94
Length 21.83 31.09 24.56 15.29 9.05 25.87
Lang→EN
Punct 21.96 32.43 25.17 16.36 9.63 25.32
Frequency 21.88 32.26 24.87 15.90 9.18 24.35
(a) BLEU
Baseline 17.62 51.11 40.47 16.12 26.52 11.46
Length 18.41 51.10 39.93 16.80 27.03 12.01
EN→Lang
Punct 18.43 51.67 41.18 16.77 27.00 12.30
Frequency 18.16 52.10 40.57 16.79 26.95 12.29
Baseline 24.66 31.77 27.23 20.63 16.28 28.11
Length 25.07 31.55 27.11 20.65 15.95 28.71
Lang→EN
Punct 25.10 32.31 27.75 21.45 17.05 28.48
Frequency 25.27 32.16 27.43 20.80 16.85 27.86
(b) METEOR
Baseline 62.64 49.21 57.17 70.17 77.20 54.29
Length 62.18 48.96 57.90 70.85 79.51 53.93
EN→Lang
Punct 61.69 48.57 57.24 70.04 77.02 54.03
Frequency 62.46 48.87 57.63 69.40 87.20 54.99
Baseline 57.06 46.42 53.31 63.62 72.46 53.66
Length 55.68 46.44 53.29 63.31 72.54 52.74
Lang→EN
Punct 56.29 45.37 52.31 62.24 72.11 53.51
Frequency 57.32 45.55 52.75 62.10 75.73 54.72
(c) TER
Table 5: Contrasting our three proposed models to the baseline (NMT; Luong et al., 2017) across 12
translation tasks. We evaluate all the models using BLEU, METEOR and TER. The bold values rep-
resent the models that show statistically significant improvements over the baseline (p < 0.001; Clark
et al., 2011). Note that for BLEU and METEOR, higher is better, while for TER, lower is better. All
regularization schemes almost consistently lead to improvements, with the punctuation-based regularizer
achieving the highest gains.
baseline across almost all language pairs for all frequencies (cf. Table 1; English has around 80M
models and across all metrics. We obtain statis- words and French has around 50M words, whereas
tically significant results for almost all translation all other languages have much less data).
tasks for at least one regularization method.
The most challenging translation tasks are
More specifically, the punctuation regularizer Slovene–English and English–Slovene, especially
outperforms all the other models on all transla- in terms of error rate. The results show that
tion tasks except for French–English and English– with 17k sentence pairs as a training set, it be-
French. For the latter, we observe that the word comes more challenging to efficiently learn any-
frequency regularizer is better than the other sys- thing. The results we obtained are between 2
tems. This could be explained by the fact that and 5 BLEU points when translating from En-
the English vocabulary has many words borrowed glish. The Slovene output contained many non-
from French, which makes the word frequency translated words. Specifically, this task greatly
regularizer a better signal than punctuation or sen- suffers when using the word frequency regularizer,
tence length for this specific task. It also could with an error increase of about 10 TER points from
be due to the fact that both English and French English to Slovene. We do not observe such losses
have the largest vocabulary for training the word for the Czech–English and English–Czech transla-
107
tion tasks, even though the vocabulary size for es- R EF That’s 90 percent of our time sur-
timating the word frequencies is lower for Czech. rounded by architecture .
We hypothesize that this is due to the Czech train- BASE <unk> percent of our time via archi-
ing set being seven times larger than the Slovene tecture .
one. We hypothesize that this is due to the fact F REQ <unk> percent of our time sur-
that for Slovene we only have 17K sentence pairs rounded by architecture .
for the training step; whereas for Czech, we have (6) I N Débloquer ce potentiel est dans
122K sentence pairs, which helped control the l’intérêt de chacun d’entre nous .
model compared to Slovene. R EF Unlocking this potential is in the inter-
One case where the punctuation regularizer suc- est of every single one of us .
ceeds consistently is on the English–German and BASE <unk> that potential is in all of us .
German–English translation tasks, with an error F REQ <unk> that potential is in the interest
reduction of about 1 TER point. This reflects the of all of us .
similarity in punctuation between these languages.
Although we also observe improvements using the More precisely, entourés in French is almost as
other regularization methods, e.g. the length-based frequent as surrounded in English, which is a word
method, these are not statistically significant here that our model with frequency-based regulariza-
as calculated by MultEval (Clark et al., 2011). tion translates correctly, while the baseline does
Table 3 shows the BLEU scores of seven differ- not. Additionally, in Example (6), our model has
ent systems including the one where we combine a better fluency and adequacy than the baseline
our three regularizers on the English–Vietnamese since it not only correctly translates l’intérêt to in-
translation task. The combined regularizer does terest, but also correctly produces of all instead of
not only produce a statistically significant im- in all, as in the baseline output.
provement of almost 1-BLEU point over the at-
tention based baseline, but it also outperforms all 7.2 Punctuation-Based Regularizer
the other regularizers achieving a BLEU score of
The punctuation-based regularization performs
27.23.
best in the German–English and English–German
7 Translation Examples translation tasks. This regularizer penalizes cases
where the difference in the number of punctuation
The punctuation regularizer outperforms the base- between the source and the target sentences is par-
line in most cases, and all of our regularization ticularly large. As seen in Example (7), simply
methods show statistically significant improve- introducing this bias into a translation model leads
ments in at least one language. Below we present to an output which more closely matches the punc-
examples, extracted from the test data, of how tuation of the source and target sentences.
each of the regularization methods affects the out-
put in comparison to the baseline model. The pur- (7) I N Und die Antwort , glaube ich , ist ja .
pose of the examples is to show how each objec- [ " F = T ∇ Sτ " ] . Was Sie ger-
tive function in the learning component affects the ade sehen , ist wahrscheinlich die beste
performance component. Entsprechung zu E = mc2 für Intelli-
genz , die ich gesehen habe .
7.1 Frequency-Based Regularizer R EF And the answer , I believe , is yes .
The frequency-based regularization method penal- [ " F = T ∇ Sτ " ] What you’re see-
izes cases where the distribution of the target vo- ing is probably the closest equivalent to
cabulary greatly differs from the source vocabu- an E = mc2 for intelligence that I’ve
lary. We have noted a significant improvement seen .
for this specific regularizer when translating from BASE And the answer , I think , is yes .
French to English and vice-versa. Examples (5) P UNC And the answer , I think , is yes . [ " R
and (6) show how this regularizer is improving the = T T <unk> " ] What you’re looking
translation output. at is probably the best <unk> <unk>
<unk> of intelligence that I’ve seen .
(5) I N 90 % de notre temps entourés par
l’architecture . The baseline MT output completely fails to cap-
108
ture anything from the input except for the first Example (10) shows another case of how the out-
part up to “. . . is yes.” Our punctuation-based put of the frequency-based regularization system
model, however, manages to capture most parts of actually shows overall improvements in an ex-
the sentence. tremely low-resource language. The output of our
system is semantically closer to the reference than
7.3 Length-Based Regularizer the baseline output, up to the word educate. In ad-
Finally, the length-based regularization method dition, the system preserves a similar length as the
leads to noticeable improvements in the Czech– source sentence.
English and English–Czech translation tasks. Ex-
(11) I N Mi smo tu na vrhu .
ample (8) shows that introducing an input sen-
R EF We are here on top .
tence length bias led to an MT output that is much
BASE What we are at the top .
closer to the reference than the baseline. The input
F REQ We are here at the top .
sentence consists of 12 tokens (including punctu-
ation), the baseline output consists of 10 tokens, Finally, Example (11) shows a low-resource case
while our length based regularization model pre- where our system manages to make subtle changes
serves the length of 12 tokens. in order to reach the correct translation, whereas
the baseline system does not.
(8) I N V roce 2009 jsem ztratila někoho ,
koho jsem velmi milovala . 8 Conclusion
R EF In 2009 , I lost someone I loved very
much . We have shown that using naive regularization
BASE In 2009 , I lost somebody who I loved . methods based on sentence length, punctuation,
L EN In 2009 , I lost somebody who I loved and word frequency consistently improves the
very much . translation quality in twelve low-resource transla-
tion tasks. The improvement is consistent across
7.4 General Improvements multiple language pairs and is not dependent on
The Slovene dataset is our smallest with about the language family. We have reported and dis-
17k sentence pairs for training. Despite the low cussed examples demonstrating why and how each
amount of resources available in Slovene, we regularizer is improving the translation quality.
found that introducing very naive linguistic biases Our proposed approach shows that even naive,
into our machine translation models actually leads but linguistically motivated, regularizers help
to subtle differences that result in an output closer improve the translation quality when training
to the reference, not only lexically, but also seman- NMT models. We believe this shows the useful-
tically. In Example (9), we compare the output ness of using task-related regularizers for improv-
of the frequency based system against the baseline ing neural models, and opens the door for future
for the Slovene to English translation: work to exploit these regularization methods in an
even more efficient manner by experimenting with
(9) I N In kaj potem ? different ways of combining the regularizers with
R EF And so , what after that ? the loss function.
BASE And then then ?
F REQ And then , what ?
References
In this particular case, the frequency based regu- Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua
larization model takes care of the translation of the Bengio. 2014. Neural machine translation by
word what, and although the word so is not trans- jointly learning to align and translate. CoRR
lated, the overall meaning of the source is pre- abs/1409.0473. http://arxiv.org/abs/1409.0473.
served. Satanjeev Banerjee and Alon Lavie. 2005. Me-
teor: An automatic metric for mt evaluation
(10) I N Imeti moraš otroke , da preživiš . with improved correlation with human judgments.
R EF You need to have children to survive . In Proceedings of the ACL Workshop on Intrin-
sic and Extrinsic Evaluation Measures for Ma-
BASE Well you have the kids that you chine Translation and/or Summarization. Associa-
need to educate . tion for Computational Linguistics, pages 65–72.
F REQ You have to have kids to educate . https://www.aclweb.org/anthology/W05-0909.
109
Mauro Cettolo, Christian Girardi, and Marcello Fed- Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris
erico. 2012. Wit3 : Web inventory of transcribed and Callison-Burch, Marcello Federico, Nicola Bertoldi,
translated talks. In Proceedings of the 16th Con- Brooke Cowan, Wade Shen, Christine Moran,
ference of the European Association for Machine Richard Zens, Chris Dyer, Ondřej Bojar, Alexan-
Translation (EAMT). Trento, Italy, pages 261–268. dra Constantin, and Evan Herbst. 2007. Moses:
Open source toolkit for statistical machine trans-
Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bah- lation. In Proceedings of the 45th Annual Meet-
danau, and Yoshua Bengio. 2014. On the prop- ing of the Association for Computational Linguistics
erties of neural machine translation: Encoder– Companion Volume Proceedings of the Demo and
decoder approaches. In Proceedings of SSST- Poster Sessions. Association for Computational Lin-
8, Eighth Workshop on Syntax, Semantics and guistics, Prague, Czech Republic, pages 177–180.
Structure in Statistical Translation. Association https://www.aclweb.org/anthology/P07-2045.
for Computational Linguistics, pages 103–111.
https://www.aclweb.org/anthology/W14-4012. Philipp Koehn and Rebecca Knowles. 2017.
Six challenges for neural machine transla-
Thomas C Chuang, Jian-Cheng Wu, Tracy Lin, Wen- tion. In Proceedings of the First Workshop
Chie Shei, and Jason S Chang. 2004. Bilingual on Neural Machine Translation. Association
sentence alignment based on punctuation statistics for Computational Linguistics, pages 28–39.
and lexicon. In International Conference on Natu- https://www.aclweb.org/anthology/W17-3204.
ral Language Processing. Springer, pages 224–232.
Minh-Thang Luong, Eugene Brevdo, and Rui Zhao.
Jonathan H. Clark, Chris Dyer, Alon Lavie, and 2017. Neural machine translation (seq2seq) tutorial.
Noah A. Smith. 2011. Better hypothesis testing for https://github.com/tensorflow/nmt.
statistical machine translation: Controlling for opti-
mizer instability. In Proceedings of the 49th Annual Minh-Thang Luong and Christopher D. Manning.
Meeting of the Association for Computational Lin- 2016. Achieving open vocabulary neural ma-
guistics: Human Language Technologies. Associa- chine translation with hybrid word-character mod-
tion for Computational Linguistics, pages 176–181. els. In Proceedings of the 54th Annual Meet-
https://www.aclweb.org/anthology/P11-2031. ing of the Association for Computational Lin-
guistics (Volume 1: Long Papers). Association
Daxiang Dong, Hua Wu, Wei He, Dianhai Yu, and for Computational Linguistics, pages 1054–1063.
Haifeng Wang. 2015. Multi-task learning for https://doi.org/10.18653/v1/P16-1100.
multiple language translation. In Proceedings of
the 53rd Annual Meeting of the Association for Minh-Thang Luong, Hieu Pham, and Christopher D.
Computational Linguistics and the 7th Interna- Manning. 2015a. Effective approaches to attention-
tional Joint Conference on Natural Language Pro- based neural machine translation. In Proceed-
cessing (Volume 1: Long Papers). Association ings of the 2015 Conference on Empirical Meth-
for Computational Linguistics, pages 1723–1732. ods in Natural Language Processing. Association
https://doi.org/10.3115/v1/P15-1166. for Computational Linguistics, pages 1412–1421.
https://doi.org/10.18653/v1/D15-1166.
Orhan Firat, KyungHyun Cho, and Yoshua Ben-
gio. 2016. Multi-way, multilingual neu- Minh-Thang Luong, Hieu Pham, and Christo-
ral machine translation with a shared at- pher D. Manning. 2015b. Effective approaches to
tention mechanism. CoRR abs/1601.01073. attention-based neural machine translation. CoRR
http://arxiv.org/abs/1601.01073. abs/1508.04025. http://arxiv.org/abs/1508.04025.
Jonas Gehring, Michael Auli, David Grangier, De- Benjamin Marie and Atsushi Fujita. 2018. A smorgas-
nis Yarats, and Yann N. Dauphin. 2017. Con- bord of features to combine phrase-based and neural
volutional sequence to sequence learning. CoRR machine translation. In AMTA.
abs/1705.03122. http://arxiv.org/abs/1705.03122.
Antonio Valerio Miceli Barone, Barry Haddow, Ul-
Sepp Hochreiter and Jürgen Schmidhuber. 1997. rich Germann, and Rico Sennrich. 2017. Regu-
Long short-term memory. Neural computation larization techniques for fine-tuning in neural ma-
9(8):1735–1780. chine translation. In Proceedings of the 2017 Con-
ference on Empirical Methods in Natural Language
Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Processing. Association for Computational Lin-
Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Tho- guistics, Copenhagen, Denmark, pages 1490–1495.
rat, Fernanda B. Viégas, Martin Wattenberg, Greg https://www.aclweb.org/anthology/D17-1156.
Corrado, Macduff Hughes, and Jeffrey Dean. 2016.
Google’s multilingual neural machine translation Graham Neubig. 2015. lamtram: A toolkit for lan-
system: Enabling zero-shot translation. CoRR guage and translation modeling using neural net-
abs/1611.04558. http://arxiv.org/abs/1611.04558. works. https://github.com/neubig/lamtram.
Diederik P. Kingma and Jimmy Ba. 2014. Adam: Kishore Papineni, Salim Roukos, Todd Ward, and
A method for stochastic optimization. CoRR Wei-Jing Zhu. 2002. B LEU : a method for au-
abs/1412.6980. http://arxiv.org/abs/1412.6980. tomatic evaluation of machine translation. In
110
Proceedings of the 40th Annual Meeting of
the Association for Computational Linguistics.
https://www.aclweb.org/anthology/P02-1040.
Jean Pouget-Abadie, Dzmitry Bahdanau, Bart van Mer-
riënboer, Kyunghyun Cho, and Yoshua Bengio.
2014. Overcoming the curse of sentence length for
neural machine translation using automatic segmen-
tation. Syntax, Semantics and Structure in Statistical
Translation page 78.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2015. Neural machine translation of rare words
with subword units. CoRR abs/1508.07909.
http://arxiv.org/abs/1508.07909.
Matthew Snover, Bonnie Dorr, Richard Schwartz, Lin-
nea Micciulla, and John Makhoul. 2006. A study of
translation edit rate with targeted human annotation.
In Proceedings of the 7th Conference of the Associ-
ation for Machine Translation in the Americas. The
Association for Machine Translation in the Ameri-
cas, pages 223–231. http://mt-archive.info/AMTA-
2006-Snover.pdf.
111
Exploring Graph-Algebraic CCG Combinators
for Syntactic-Semantic AMR Parsing
Sebastian Beschke
University of Hamburg, Germany
beschke@informatik.uni-hamburg.de
112
Proceedings of Recent Advances in Natural Language Processing, pages 112–121,
Varna, Bulgaria, Sep 2–4 2019.
Champagne and dessert followed
constituent or to the left. This directionality is ex-
N conj N S [dcl ]\NP
and follow-01
pressed by forward and backward slashes. E.g.,
op1 op2 ARG1 given the atomic syntactic categories S for sen-
champagne h0i h1i h0i
dessert tences and NP for noun phrases, the complex cat-
lex lex
NP NP
conj
egory (S \NP )/NP represents a transitive verb,
and accepting first an NP to the right and then another
op1 op2
h0i dessert
NP to the left to produce a sentence.
NP \NP Semantic categories contain building blocks for
NP
<
sentential meaning representations. Traditionally,
op1
and λ-calculus is used to represent the compositional-
op2
champagne dessert ity of semantic categories, while the object lan-
< guage that is being composed is a logical repres-
S [dcl ]
ARG1 entation of sentence meaning. This paper deviates
follow-01 and
op1 op2 from that tradition by using a graph representation
champagne dessert for semantic categories which is defined Section
2.3.
Figure 1: An example graph-algebraic CCG deriv-
ation. 2.2 Abstract Meaning Representation
The Abstract Meaning Representation (AMR;
parser, which makes it compute-intensive and dif- Banarescu et al., 2013) is a meaning representa-
ficult to manage. We address this issue by per- tion language that underlies much recent work in
forming lexicon induction in a separate step. semantic parsing. In AMR, meaning is annotated
Besides AMR parsing, CCG has also been used on the sentence level in the form of a labeled, dir-
for joint syntactic-semantic parsing in other con- ected graph. While the nodes of the graph rep-
texts (Krishnamurthy and Mitchell, 2014; Lewis resent instances of concepts, edges represent roles
et al., 2015). that these entities play with respect to each other.
2 Background 2.2.1 Evaluation of AMRs
This paper uses Combinatory Categorial Gram- AMR parsers are usually evaluated with respect to
mar (CCG) to derive Abstract Meaning Repres- the Smatch metric (Cai and Knight, 2013), which
entations (AMR) using a graph-algebraic modific- measures precision and recall of semantic triples
ation of CCG’s syntax-semantics interface. These in an AMR graphs with respect to a gold standard
concepts are briefly introduced in this below. graph. The computation of Smatch relies on find-
ing an optimal alignment between the two graphs,
2.1 CCG for Semantic Parsing which is usually approximated.
Combinatory Categorial Grammar (CCG) de-
scribes syntax and semantics as part of the same 2.3 Graph-Algebraic CCG
derivation process (Steedman, 2000). CCG deriv- Graph algebras are an established means to model
ations are trees where every node is annotated with the derivation of AMRs (Koller, 2015). A modi-
both a syntactic and a semantic category. The cat- fication of CCG that applies graph-algebraic oper-
egories at the leaves of the tree are drawn from a ators to semantic categories has first been presen-
lexicon, while categories at the inner nodes result ted by Beschke and Menzel (2018). They define a
from the application of combinatory rules to the set of semantic operators that apply to s-graphs,
child nodes’ categories. The syntax-semantics in- which contain specially marked source nodes,
terface in CCG is transparent, meaning that the which are consecutively indexed starting from 0.
same rule is always applied to syntactic and se- They also define three semantic operators:
mantic categories.
In CCG, categories are understood as n-ary • Application, which 1) merges the root of
functions. Syntactic categories essentially express an argument graph with the highest-indexed
the type of the associated semantic category by source node of the function graph and 2)
specifying the types of constituents that can be merges all source nodes that have the same
accepted as arguments, either to the right of the index.
113
Combinator Left Operand Right Operand Result
> X/Y : X h0i Y : X ⇒ X: X X
< Y : X X\Y : X h0i ⇒ X: X X
B> X/Y : X h0i Y /Z : X h0i ⇒ X/Z : X X h0i
B× < Y /Z : X h0i X\Y : X h0i ⇒ X/Z : X X h0i
B >
2
X/Y : X h0i (Y /Z1 )/Z2 : X h0i ⇒ (X/Z1 )/Z2 : X X h0i
h1i h1i
conj conj : conj h0i X: X h0i ⇒ X\X : conj h1i
h1i X h0i
rp X: X .: ⇒ X: X
lp .: X: X ⇒ X: X
Table 1: The set of binary combinators used in our system. Circles and diamonds correspond to arbitrary
AMR subgraphs. X and Y represent arbitrary syntactic categories. The conj node represents any concept
corresponding to a conjunction, such as and or contrast. Edge labels are omitted.
The combinators Forward Application (>), Backward Application (<), Forward Composition (B>),
Backward Crossed Composition (B× <), and Forward Generalised Composition (B2 >) all use the Ap-
plication semantic operator. The Conjunction (conj) combinator uses the Conjunction semantic operator,
and Left and Right Punctuation (lp, rp) use Identity.
• Conjunction, which 1) merges the root of an • A set of alignments linking tokens in the sen-
argument graph with the 1-indexed source tence to nodes in the meaning representation,
node of the conjunction graph, and 2) re- obtained from automatic alignment tools
names the 0-indexed source node of the con-
junction graph so that it becomes the highest- A set of lexical items explaining the sentence
indexed source node in the combined graph can then be obtained by walking down the syn-
(thus becoming accessible for application). tactic parse tree, starting at the root with the full
sentential meaning representation. At each binary
• Identity, in which the function graph is derivation step, the meaning representation is par-
empty and the argument graph is returned un- titioned into two subgraphs by unmerging nodes
changed. as appropriate. Each split is done in such a way
that it can be reversed using a graph algebraic
An overview of the rules as well as how they are combinator and the token-to-node alignments are
applied in the context of CCG derivations is given honored.
in Table 1. For any token, this procedure may generate sev-
An example derivation is given in Figure 2. eral or no lexical entries. If the alignments do not
uniquely specify how the meaning representation
3 Lexicon Induction should be divided in a splitting step, all alternat-
For parsing with graph algebraic CCG, a lexicon ives are explored. Also, splitting may abort at an
must first be obtained. We achieve this using the inner node of the derivation if there is no combin-
recursive splitting algorithm by Beschke and Men- ator that satisfies the alignment constraint.
zel (2018), which uses the following information This work adds two steps to the lexicon induc-
to induce lexical items from an AMR-annotated tion process: the delexicalisation of lexical items,
sentence: followed by filtering for the most probable deriva-
tion for each sentence according to EM estimates.
• The sentence’s AMR
3.1 Delexicalisation
• A syntactic CCG parse obtained from a syn- We achieve generalisation over content words by
tax parser delexicalising lexical entries. We follow the ap-
114
proach from Kwiatkowski et al. (2011) which di- Algorithm 1 Variation of the inside-outside al-
vides lexical entries into templates and lexemes. A gorithm to estimate parameters over CCG deriv-
template is a graph wherein up to one node has ations. See Section 3.2 for function definitions.
been replaced by a lex marker. A lexeme x—y is a Input: Data set S; scoring function S COREi
pair of a word x and a node label y. For examples Output: Distributions PTi+1 and PLi+1
of templates and lexemes, see Table 2. 1: countT [j] ← 0 for 0 ≤ j < |T |
The idea of the delexicalisation algorithm is that 2: countL [j] ← 0 for 0 ≤ j < |L|
a node in the graph which corresponds to the lex- 3: for s ∈ S do
ical meaning of the lexical entry is replaced by a 4: chart ← S PLIT P(s)
marker, converting it into a template. Since it is 5: likelihood ← e∈chart[0,|s|−1] INi+1 (e)
not known in advance which node carries the lex- 6: for e ∈ chart do i (e) OUT i+1 (e)
ical meaning, we replace every node in turn and 7: c ← S CORElikelihood
add all resulting templates to the lexicon. Every 8: for (t, l) ∈ D ELEX(e) do
replaced node label is associated with the token 9: countT [t] ← countT [t] + c
currently under consideration and stored as a lex- 10: countL [l] ← countL [l] + c
eme. 11: end for
Not all lexical entries contain a node with lex- 12: end for
ical meaning, e.g. in the case of function words. 13: end for
i+1 countT [t]
Therefore, the original meaning representation is 14: PT (t) = P
0 countT [t0 ]| for t ∈ T
t ∈T
also added to the lexicon as a template along with i+1
15: PL (l) = P countL [l] 0 for l ∈ L
an empty lexeme. l0 ∈L countL (l )|
115
section and output the predicted token-wise su-
X pertag distributions (clipping at 99% cumulative
S COREi (e) = PTi (t)PLi (l) probability). To obtain supertag predictions on the
(t,l)∈D ELEX(MR(e)) train section, we employ 5-way jackknifing: the
data is split into five parts and predictions for each
A given meaning representation MR(e) can be
part are obtained by training on the remaining four
created by either instantiating a lexical entry with
parts.
probability S COREi (e), or by deriving it using
During training, the occurrence of the correct
any of its pairs of children (l, r) with probability
label within the top-10 predictions for every token
IN i (l) IN i (r). All of these are alternative choices;
is monitored and training aborted when this meas-
therefore, the probabilities are summed to make
ure stops improving (early stopping).
up the inside probability. The outside probability
is composed of the entry’s parents’ outside prob- 4.1.1 Limited Supervision
abilities and the entry’s neighbours’ inside prob- The grammar induction process as described in
abilities. Section 3 attempts to find lexical items for every
Algorithm 3.2 describes how an updated set of individual token, but may stop early if no combin-
parameters is estimated using these calculations. atory rule fitting the alignment constraint is avail-
able. In this case, no supervision for training the
4 Parsing tagger is available at the token level. We over-
Our parser uses a CKY-style chart parsing al- come this issue by labelling the respective tokens
gorithm to parse sentences to AMR. For each as UNK (the same label used for rare templates oc-
token, template-lexeme pairs are drawn from the curring only once) and masking UNK tokens in the
lexicon. Recursively, derivation nodes are created loss function.
according to CCG/AMR rules. All candidate de- This allows the tagger to fill in the gaps with
rivation nodes are evaluated with respect to a lin- reasonable templates that are in the lexicon. How-
ear model. A beam search limitation is applied, ever, it also means that not every sentence from the
meaning that only the top n candidates from each train set can be perfectly parsed any more, because
chart cell are kept. it is possible that its meaning representation can-
The flip side of using a delexicalised lexicon is not be constructed using the induced token-level
that every template can now be applied to every lexical entries.
token. To limit the number of leaves that have
to be considered, we employ a supertagger which
5 Training
predicts the most suitable template for each token. To drive the parser, we train a linear model over
We then limit our search to the most probable tem- graph algebraic CCG derivations. Since we do not
plates as predicted by the supertagger. observe derivations in the data, this is an instance
of latent variable learning and a supervision signal
4.1 Supertagging
must be generated. We take a dual approach by
For supertagging, we use a single-layer BiLSTM. combining two weak supervision signals:
For inputs, the raw tokens and syntactic CCG cat-
egories predicted by a CCG supertagger are used. 1. An oracle is used to heuristically generate
The model is then trained to predict the template silver-standard derivations, which can then
instantiated by each token. be used for training.
The following preprocessing steps are applied: 2. The derivations found by the parser are eval-
uated and used for cost-sensitive parameter
• Tokens are embedded using the third layer
updates.
produced by ELMo (Peters et al., 2018).
5.1 Model
• CCG supertags, as well as templates, that oc-
cur in less than two sentences are replaced by We train a linear model using a structured per-
UNK. ceptron algorithm (Collins, 2002) with Adadelta
updates (Zeiler, 2012). We use features over paths
To predict supertags on the dev and test sec- in the graph as well as the identities of invoked
tions, we train the supertagger on the entire train templates, lexemes, and combinators.
116
Template Lexeme Combined
N : lex weapons—weapon N : weapon
N P/N : h0i the—∅ N P/N : h0i
mod mod
N/N : h0i lex nuclear—nucleus N/N : h0i nucleus
ARG0 ARG0
(S\N P )/N P : lex h0i said—say-01 (S\N P )/N P : say-01 h0i
ARG ARG
1 h1i 1 h1i
N P : country Iran—“Iran” N P : country
name name
name name
op1 op1
lex “Iran”
Table 2: Selected templates and lexemes from the induced lexicon. The templates are among the 20 most
highly scored according to EM parameters; the lexemes among the top 50.
Figure 2: A derivation for a subsequence of PROXY NYT ENG 20020406 0118.25, as produced by
our parser.
117
5.2 Oracle Parsing derivation. Cost-sensitive updates let the parser
In latent variable learning in structured prediction search for complete derivations and enforce a
settings, the challenge is to obtain an unobserved margin between the best derivations in the beam
derivation for a known gold-standard result. In and all the others. We follow Singh-Miller and
this case, gold-standard sentential AMRs are an- Collins (2007) by implementing a cost-sensitive
notated in the AMR corpus, but they are not re- perceptron algorithm which weights hypotheses
lated to the sentence by a grammatical derivation. according to their Smatch f1 score.
A common approach to this challenge is forced 5.4 Combined Update Strategy
decoding (Artzi et al., 2015): the parser is used
to construct derivations which lead to the correct While early updates are efficient, our oracle is im-
result by pruning all hypotheses from the search perfect. To allow the parser to improve over oracle
space which deviate from the gold-standard AMR. parses, we use a cost-sensitive update whenever a
E.g., all AMRs that contain elements not present parse has been found whose Smatch f1 score sur-
in the gold-standard could be pruned. passes that of the oracle parse.
However, as noted in Section 4.1, not every
6 Experimental Setup
gold-standard AMR can be reconstructed perfectly
using the induced lexicon due to the incomplete- We evaluate our parser1 on the proxy section of
ness of the splitting algorithm, which defies find- the AMR 1.0 corpus (LDC2014T12; Knight et al.,
ing correct parses using forced decoding. 2014). This section consists of newswire texts.
Instead, we train the parser using an oracle Sentences are tokenised and lemmatised us-
driven by a heuristic scoring function which scores ing Stanford NLP (Manning et al., 2014). We
the correctness and completeness of an intermedi- use EasyCCG to obtain CCG parses and super-
ate hypothesised AMR. We parse the sentence us- tags (Lewis and Steedman, 2014). Token-to-
ing CKY with beam search, ranking intermediate AMR alignments are obtained by combining out-
results according to the harmonic mean of the fol- puts generated by the JAMR aligner (Flanigan
lowing values: et al., 2014) and the ISI aligner (Pourdamghani
et al., 2014), as described by Beschke and Men-
• Triple precision: the proportion of node- zel (2018).
edge-node triples in the intermediate result First, we induce a CCG/AMR lexicon from
that also occur in the gold standard meaning the entire proxy-training section, delexicalise the
representation. entries, and filter for the best derivations using
• Alignment recall: the proportion of node la- EM, as described in Section 3. We perform 100 it-
bels that are linked by an alignment edge to erations of EM. Sentences longer than 100 tokens
one of the intermediate result’s tokens that are filtered out. The resulting lexicon contains
also occur in the intermediate result. 15630 templates and 10504 lexemes.
Next, we extract template tag sequences and
This scoring function is designed to rank res- train our suppertagger on them. First, tags for the
ults in proportion to their deviation from the gold training data are predicted using 5-way jackknif-
standard, achieving a soft form of pruning. ing. Then, a model is trained on the entire training
Having obtained a set of derivations using or- section and used to predict tags for the dev and test
acle parsing, we finally re-rank these derivations sections of the corpus. Since only templates are
by their Smatch f1 scores and use the best deriva- predicted that occur in at least two training sen-
tion to perform a parameter update using an early tences, a set of 2453 templates is used for predic-
update strategy (Collins and Roark, 2004). tion. The top-10 recall of the annotated supertags
is 96.4% on a randomly chosen held-out portion
5.3 Cost-Sensitive Update of the training set.
Another approach to training with weak super- Finally, the induced lexicon as well as the pre-
vision for structured prediction are cost-sensitive dicted tag sequences are used to parse the proxy-
updates. While the gold-standard to update to- test section of the AMR corpus. We use a beam
wards is unknown, an evaluation metric is avail- 1
For information on reproducing the experiments, see
able for the AMR that results from a specific https://gitlab.com/nats/gramr-ranlp19/.
118
System P R F 7.1 Discussion
This paper 0.688 0.423 0.524 The parser output in Figure 5 shows some of
Artzi et al. (2015) 0.668 0.657 0.663 the most common errors produced by our parser.
Misra and Artzi (2016) 0.681 0.642 0.661 Firstly, the sequence International Science and
Liu et al. (2018) - - 0.731 Technology Center is not recognised as a contigu-
ous named entity. Additionally, Technology Cen-
Table 3: Smatch results on the proxy-test section ter is misrecognised as a country. Both of these
of LDC2014T12. Liu et al. (2018) did not report issues can be classified as supertagging errors, as
precision and recall in their paper. P stands for they result from the templates chosen from the lex-
precision, R for recall, F for f1 score. icon. In this specific case, the supertagger’s beha-
viour could likely be improved by adding named
entity features to its input. In general, the super-
size of 15 during parsing and 20 for finding oracle tagging task is challenging, especially in the case
derivations (see Section 5.2). Parses whose root of function words, which tend to be highly poly-
categories do not match any of the top-10 deriva- semous.
tions produced by EasyCCG are dropped from the Additionally, the scopes of and and of are inver-
parser output2 . ted. This can be interpreted as a weakness of the
The smatch tool3 is used to calculate Smatch parsing model, which misjudges the probability of
precision, recall, and f1 scores for the parser out- the respective scope assignments. Although one
put. would hope for a semantic parser to improve pre-
cisely upon these semantically informed syntactic
7 Results decisions, this behaviour is perhaps to be expected
given that we train a sparse linear model with a re-
We compare our system to two previous CCG- latively small amount of training data. Replacing
based AMR parsers (Artzi et al., 2015; Misra and the linear classifier with a neural model that com-
Artzi, 2016), as well as the current state of the art putes embeddings of graph meanings, such as the
in AMR parsing on this data set (Liu et al., 2018). architecture proposed by (Misra and Artzi, 2016),
The results are shown in Table 7. The system in- could improve the parser’s judgment.
troduced in this paper achieves comparable preci-
sion to the other CCG-based systems, but lower 8 Conclusion
recall.
We have introduced a pipeline for training a CCG
This gap is largely, but not completely, ex-
parser which jointly models syntax and semantics.
plained by sentences that were not parsed at all:
A central element of our architecture are efforts to
when unparsed sentences are excluded from the
reduce the lexicon size. With 2453 delexicalised
evaluation, our system achieves a precision of
templates, our parser uses a relatively small lex-
0.701 and a recall of 0.6154 . Oracle parsing
icon despite the templates being induced automat-
achieves a Smatch precision of 0.886 and an f1
ically. We employ a semantic construction mech-
score of 0.706.
anism that is less powerful with λ-calculus, but
The evaluation set contains 823 sentences in
still achieve competitive precision.
total, of which 170 were not parsed, resulting
Future directions in this line of work could in-
in a coverage of 79%. Of these sentences, 68
clude applications that make use of the system’s
were skipped because they were longer than 40
transparency, such as the interactive training of
tokens. The remaining 102 are unparsed because
parsers without gold-standard annotations, or the
the parser failed to find a complete parse.
application of external constraints such as contex-
2
This restriction was included because the parser tended tual knowledge to the parser.
to favour interpretations of sentences as NP instead of S.
3
https://github.com/snowblink14/smatch, Acknowledgments
revision ad7e655
4
The precision improves when unparsed sentences are The author would like to thank Wolfgang Menzel
excluded because the smatch tool does not permit empty for discussions and advice, as well as the anonym-
AMRs to be specified. Unparsed sentences are therefore rep-
resented by single-node placeholder AMRs, which are penal- ous reviewers.
ised in terms of precision.
119
References Discriminative Graph-Based Parser for the Ab-
stract Meaning Representation. In Proceed-
Yoav Artzi, Kenton Lee, and Luke Zettlemoyer. ings of the 52nd Annual Meeting of the Asso-
2015. Broad-coverage CCG Semantic Parsing ciation for Computational Linguistics (Volume 1:
with AMR. In Proceedings of the 2015 Con- Long Papers). Association for Computational Lin-
ference on Empirical Methods in Natural Lan- guistics, Baltimore, Maryland, pages 1426–1436.
guage Processing. Association for Computational http://www.aclweb.org/anthology/P14-1134.
Linguistics, Lisbon, Portugal, pages 1699–1710.
http://aclweb.org/anthology/D15-1198. Jonas Groschwitz, Matthias Lindemann, Meaghan
J. K. Baker. 1979. Trainable grammars for Fowlie, Mark Johnson, and Alexander Koller. 2018.
speech recognition. The Journal of the Acous- AMR dependency parsing with a typed semantic al-
tical Society of America 65(S1):S132–S132. ht- gebra. In Proceedings of the 56th Annual Meet-
tps://doi.org/10.1121/1.2017061. ing of the Association for Computational Linguist-
ics (Volume 1: Long Papers). Association for Com-
Laura Banarescu, Claire Bonial, Shu Cai, Madalina putational Linguistics, Melbourne, Australia, pages
Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin 1831–1841. http://aclweb.org/anthology/P18-1170.
Knight, Philipp Koehn, Martha Palmer, and Nathan
Schneider. 2013. Abstract Meaning Represent- Kevin Knight, Laura Baranescu, Claire Bonial,
ation for Sembanking. In Proceedings of the Madalina Georgescu, Kira Griffitt, Ulf Herm-
7th Linguistic Annotation Workshop and Interoper- jakob, Daniel Marcu, Martha Palmer, and Nathan
ability with Discourse. Association for Computa- Schneider. 2014. Abstract Meaning Representa-
tional Linguistics, Sofia, Bulgaria, pages 178–186. tion (AMR) Annotation Release 1.0 LDC2014T12.
http://www.aclweb.org/anthology/W13-2322. Linguistic Data Consortium, Philadelphia. ht-
tps://catalog.ldc.upenn.edu/LDC2014T12.
Sebastian Beschke and Wolfgang Menzel. 2018. Graph
algebraic combinatory categorial grammar. In Pro- Alexander Koller. 2015. Semantic construction
ceedings of the Seventh Joint Conference on Lex- with graph grammars. In Proceedings of
ical and Computational Semantics. Association for the 11th International Conference on Compu-
Computational Linguistics, New Orleans, Louisi- tational Semantics. Association for Computa-
ana, pages 54–64. https://doi.org/10.18653/v1/S18- tional Linguistics, London, UK, pages 228–238.
2006. http://www.aclweb.org/anthology/W15-0127.
Austin Blodgett and Nathan Schneider. 2019. An Jayant Krishnamurthy and Tom M. Mitchell. 2014.
improved approach for semantic graph compos- Joint Syntactic and Semantic Parsing with Com-
ition with CCG. In Proceedings of the 13th binatory Categorial Grammar. In Proceedings
International Conference on Computational Se- of the 52nd Annual Meeting of the Associ-
mantics - Long Papers. Association for Computa- ation for Computational Linguistics (Volume 1:
tional Linguistics, Gothenburg, Sweden, pages 55– Long Papers). Association for Computational Lin-
70. https://www.aclweb.org/anthology/W19-0405. guistics, Baltimore, Maryland, pages 1188–1198.
https://doi.org/10.3115/v1/P14-1112.
Shu Cai and Kevin Knight. 2013. Smatch: an
Evaluation Metric for Semantic Feature Struc- Tom Kwiatkowski, Luke Zettlemoyer, Sharon
tures. In Proceedings of the 51st Annual Meet- Goldwater, and Mark Steedman. 2011. Lex-
ing of the Association for Computational Linguistics ical generalization in CCG grammar induction
(Volume 2: Short Papers). Association for Compu- for semantic parsing. In Proceedings of the
tational Linguistics, Sofia, Bulgaria, pages 748–752. Conference on Empirical Methods in Nat-
http://www.aclweb.org/anthology/P13-2131. ural Language Processing. pages 1512–1523.
Michael Collins. 2002. Discriminative Training Meth- http://dl.acm.org/citation.cfm?id=2145593.
ods for Hidden Markov Models: Theory and Ex-
periments with Perceptron Algorithms. In Proceed- Mike Lewis, Luheng He, and Luke Zettlemoyer.
ings of the 2002 Conference on Empirical Meth- 2015. Joint A* CCG Parsing and Semantic Role
ods in Natural Language Processing. Association Labelling. In Proceedings of the 2015 Con-
for Computational Linguistics, pages 1–8. ht- ference on Empirical Methods in Natural Lan-
tps://doi.org/10.3115/1118693.1118694. guage Processing. Association for Computational
Linguistics, Lisbon, Portugal, pages 1444–1454.
Michael Collins and Brian Roark. 2004. Incre- http://aclweb.org/anthology/D15-1169.
mental Parsing with the Perceptron Algorithm. In
Proceedings of the 42nd Meeting of the Associ- Mike Lewis and Mark Steedman. 2014. A* CCG
ation for Computational Linguistics (ACL’04), Main Parsing with a Supertag-factored Model. In
Volume. Barcelona, Spain, pages 111–118. ht- Proceedings of the 2014 Conference on Em-
tps://doi.org/10.3115/1218955.1218970. pirical Methods in Natural Language Pro-
cessing (EMNLP). Association for Computa-
Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, tional Linguistics, Doha, Qatar, pages 990–1000.
Chris Dyer, and Noah A. Smith. 2014. A http://www.aclweb.org/anthology/D14-1107.
120
Yijia Liu, Wanxiang Che, Bo Zheng, Bing Qin, and as Sequence-to-Graph Transduction. ht-
Ting Liu. 2018. An AMR Aligner Tuned by tps://arxiv.org/abs/1905.08704.
Transition-based Parser. In Proceedings of the 2018
Conference on Empirical Methods in Natural Lan-
guage Processing. Association for Computational
Linguistics, Brussels, Belgium, pages 2422–2430.
https://www.aclweb.org/anthology/D18-1264.
121
Quasi Bidirectional Encoder Representations from Transformers for
Word Sense Disambiguation
122
Proceedings of Recent Advances in Natural Language Processing, pages 122–131,
Varna, Bulgaria, Sep 2–4 2019.
Consider the sentence (1) above. It features attrac- framing of WSD as a tagging task (Raganato et al.,
tors, i.e. words or phrases pushing the sense inter- 2017a; Vial et al., 2018).
pretation in one direction or the other, with the left
context providing a strong cue for the food sense Transfer Learning WSD Systems One of the
and the right for the devotion sense. In this pa- best performing WSD systems (Yuan et al., 2016)
per, we propose a modification of the usual CLM employs a semi-supervised neural architecture,
architecture for transfer learning that enables us whereby a unidirectional LSTM was trained to
to train a high-performance WSD system. In this predict a masked token on huge amounts of un-
context, we make the following contributions: labeled data (over 100B tokens). The trained
LSTM was used to produce contextualized embed-
• we introduce the BiTransformer, a novel dings for tagged tokens in SemCor; then kNN or
Transformer-based (Vaswani et al., 2017) co- a more sophisticated label propagation algorithm
attentive layer allowing deeper bidirectional- was used to predict a sense. The size of the train-
ity; ing data makes replication difficult – a reimple-
mentation attempt with a smaller corpus led to
• we introduce QBERT (Quasi Bidirectional worse results (Le et al., 2018). A similar approach
Encoder Representations from Transform- using ELMo contextualized embeddings has been
ers), a novel Transformer-based architecture presented by Peters et al. (2018), but the results
for CLM making use of the BiTransformer; were underwhelming. Another attempt at using
transfer learning in WSD has been carried out by
• we train a WSD model using QBERT contex-
Melacci et al. (2018). The authors enhanced IMS
tualized embeddings, outperforming on the
with context2vec (Melamud et al., 2016), obtain-
standard evaluation datasets both the pre-
ing performance roughly on a par with Yuan et al.
viously established state of the art (by a
(2016).
large margin) and a comparable model using
ELMo; Contextualized Embeddings Most of the ap-
proaches to contextualized embeddings involve
• we use QBERT to beat ELMo on the re-
CLM pretraining of directional (either attentive or
cently established Word-in-Context (WiC)
recurrent) networks. Very successful CLM-based
task (Pilehvar and Camacho-Collados, 2019).
models include ELMo, in which two separate di-
2 Related Work rectional LSTMs are fed the output of a shared
character-based Convolutional Neural Network
Despite the limited availability of training data, the (CNN) encoder (Peters et al., 2018), and Ope-
WSD systems offering the best performances are nAI GPT, using Transformers instead of LSTMs
supervised ones. Many of the approaches are still and a BPE vocabulary (Sennrich et al., 2016) with
end-to-end, i.e. they only make use of the infor- regular embeddings instead of the CNN encoder
mation learned during the WSD training. (Radford et al., 2018). Another popular approach,
Flair, features character-level LSTMs, outputting
End-to-end WSD Systems In WSD traditional
hidden states at word boundaries (Akbik et al.,
machine learning techniques are still very compet-
2018). As CLM architectures are normally uni-
itive because they are not as data-hungry as neural
directional, one alternative in order to guarantee a
networks. The very popular It Makes Sense (IMS)
joint encoding of the context is the Masked Lan-
system (Zhong and Ng, 2010), based on Support
guage Modeling (MLM) of BERT (Devlin et al.,
Vector Machines and hand-crafted features, per-
2019), which, however, requires a variety of tricks
forms very well when word embeddings are used
at training time.
as additional features (Iacobacci et al., 2016); the
classifier by Papandrea et al. (2017) also gets com- 3 The QBERT Architecture
petitive results. The system of Weissenborn et al.
(2015) attains very high performances, but only Similarly to other LM-based approaches to con-
disambiguates nouns. More recently, neural mod- textualized embeddings (Peters et al., 2018; Rad-
els have been developed (Kagebäck and Salomon- ford et al., 2018; Howard and Ruder, 2018; Devlin
sson, 2016; Uslu et al., 2018; Luo et al., 2018). et al., 2019), the architecture we hereby propose
Some of the most successful offer an intuitive has two main components, which we will refer to
123
QBERT Encoder
Past Transformer Stack
BiTransformer
futuremask. futuremask. Position Shift
Positional e1 e1 p
1
1
p
1
1
embeddings P≫ F≪
feedforward
feedforward
pastmask.
PAD
1 1 n
p p P f
e2 e2 K 2 2 K 2
…
Q Q
W
e3 e3 V p
1
3
p
1
3
V f
1
n
p
n
1
f
3
n
K CLM
1 1
PAD n
Q
n n
Prediction
e4 e4 p p f p f V
4 4 2 2 4
Head
p
n
n p
n
PAD
feedforward
attnMMH 1 f
attnMMH 3 3
n
p f
n
C
attnMMH
pastmask.
2
pastmask. 4
e1
n
p
PAD
e1 e1 f
1
1
f
1
1
3
f
2
n
PAD
WSD
e2
feedforward
feedforward
n
p
4 n n
futuremask.
1 1 f p
Prediction
e2 e2 f f
K 2 2 K 3 1 K
e3
…
Q Q Q
e3 e3 V f
1
f
1
V F f
4
n
p
n
2 V
Head
3 3
e4
e4 e4 f
4
1
f
4
1
PAD p
n
3
Adaptive F≪ P≫
Future Transformer Stack
as Encoder and task-specific Prediction Head. In second (F ). To combine the shifted sequences
Figure 1 we show a high-level view of our sys- we use a novel Transformer layer variant taking
tem. Raw tokens are fed to the encoder, which em- them both as input, the BiTransformer, featuring a
beds them into context-independent fixed-length co-attentive mechanism in which P attends over
vector representations (the word embeddings, W F and F attends over P . The Encoder is
in Figure 1), then uses them to produce context- trained on CLM using an Adaptive Softmax layer
dependent hidden representations (the contextual- (Grave et al., 2017) as Prediction Head.
ized embeddings, C), where the context is some
subset of the sequence itself. The Prediction Head 3.2 Transformer Variants in QBERT
exploits the vectors produced by the Encoder to In the QBERT Encoder we employ three dis-
perform a task. tinct variants of the plain Transformer: the future-
masked Transformer, the past-masked Trans-
3.1 Encoder
former and the BiTransformer. To introduce them
As will become clear in what follows, the En- we first need to elaborate further into the inner
coder of the QBERT architecture is able to com- workings of the layer. A vanilla Transformer layer
pute the hidden representation of a word wt in a (Vaswani et al., 2017) can be defined as a multi-
sequence w1:n as a function of the weights and head self attention submodule followed by a time-
of the whole sequence except wt itself, i.e. of wise 2-layer feedforward network, with additional
w1:t−1 and wt+1:n . To embed tokens, the En- residual connection (He et al., 2016) and layer nor-
coder uses the Adaptive Input layer (Baevski and malization stabilizing training (Ba et al., 2016).
Auli, 2018). Sinusoidal positional embeddings
are added to the output and passed to two sep- Core (Self) Attention The intuition behind the
arate stacks of masked Transformers, computing attention mechanism is very simple (Bahdanau
two directional encodings of the sequence, with et al., 2015; Luong et al., 2015). We have a se-
one (P ) having past and present (w1:t ) informa- quence of vectors (the nq queries Q of dimen-
tion encoded in the present-token hidden vector sion dq ) and we want to compute relevance scores
and the other (F ) having present and future (wt:n ) against some other sequence of vectors (the nk
information instead. Since in the CLM training keys K of dimension dk ) specific to each couple
information about the present token must be hid- of vectors q and k. The nq × nk relevance score
den from the output layer, we shift and pad the matrix is then used to compute nq weighted means
sequences in order to have only the past tokens of another sequence of vectors (the nk values V
encoded in the output of the first stack (P ) and of dimension dv ). So, if we pack Q, K, V into
only future tokens encoded in the output of the matrices, the mechanism can be distilled into the
124
formula: from P and the first from F and add padding to
QK T the opposite sides. This effectively shifts the hid-
attn(Q, K, V ) = softmax( √ )V (1)
dk den representation by one place to the left and
√ by one place to the right. We refer to the result-
where dk is a normalization factor meant to
ing sequences as, respectively, P and F . As
prevent the dot products from getting too large.
a result, the ith position of P encodes informa-
In the case of the self attention mechanism Q,
tion from tokens w1:i−1 while F encodes wi+1:n .
K, V stand for the same matrix. In the Trans-
Formally:
former (Vaswani et al., 2017), multi-head attention
P =P AD ⊕ P1:n−1
is used, in which n attentions (the heads) are com- (3)
puted in parallel, concatenated and then combined F =F2:n ⊕ P AD
through dot product with a dq n × do matrix W o .
where ⊕ denotes concatenation along the timestep
For each attention head hi , there are three weight
dimension. The padding vector P AD is learned
matrices WiQ , WiK , WiV , multiplying respectively
during training. The process is visualized in Fig-
Q, K and V . Formally, we define multi-head at-
ure 1, where tokens are aligned according to their
tention (attnM H ) as:
shifted positions.
attnMH (Q, K, V ) =
3.2.2 BiTransformer
n
M (QWiQ )(KWiK )T To combine P and F we employ the BiTrans-
[softmax( √ )(V WiQ )]W o
i=0
dk former, a novel Transformer layer variant that uses
(2) a masked coattentive multihead attention mecha-
nism over two sequences. Masking allows P
where ⊕ denotes concatenation along the second to attend over F while keeping present-token
dimension. knowledge hidden from the network, and vice
Transformer Masking In the past and future versa. This allows deeper bidirectionality in that
stacks, as well as in the BiTransfomer layer, we the resulting output is not a naive combination of
employ a masking mechanism on attention to en- two separate directional representations but rather
force directionality, i.e. to force the relevance the result of a whole-sequence attention, albeit
scores computed between Q and K to be 0 for computed in a two-step process, where the first
tokens following or preceding the current one, as step can be arbitrarily deep (the masked Trans-
needed. Masking can be implemented by per- former stacks) and the second is always shallow
forming an elementwise sum between QK T and (the BiTransformer). Unfortunately, BiTransform-
a nq × nk masking matrix M , whose values are ers cannot be stacked as each timestep in the out-
set to −∞ if Ki and Vj are to be excluded from put of the layer encodes information about every
the attention computation, else 0. In our architec- token in the sequence but the one it has to pre-
ture we employ two different masking matrices: dict in CLM, so any further use of attention would
a future masking-matrix Mf which is set to −∞ make pretraining impossible.
when i < j and a past-masking matrix Mp set to The BiTransformer requires modifications to
−∞ when i > j; note that Mf = MpT . Mul- the first part of the vanilla Transformer intramod-
tihead and masked attention can be combined by ule architecture. First, both input sequences
simply using the same masking matrix in each at- are layer normalized separately. We compute
tention head. We use Mf in the past Transformer a masked multihead attention using the future-
stack and Mp in the future stack, producing, re- masked sequence P as Q, the past-masked se-
spectively, P and F . To encourage the network quence F as K and V , using the past-masking
to encode comparable representations we tie the matrix Mp . To give an insight into what happens,
weights between layers at the same depth on the the position i of the n queries, encoding informa-
past and future stack. tion about words 1 to i − 1, is allowed to look
at positions i to n of the keys, encoding words
3.2.1 Timestep Shift wi+1:n , wi+2:n and so on. Then we compute the
Present-token information is still encoded in both reverse, using F , P and future-masking ma-
P and F . To remove it, we use a simple shift- trix Mf . This process results in two sequences to
ing approach where we detach the nth timestep which input residuals are added, and then added
125
together via a simple elementwise sum. The rest size of 80000 tokens. As an optimizer we employ
of the layer works just like a regular Transformer regular Nesterov-accelerated SGD, with a learn-
layer. We formally describe this coattentive mech- ing rate that first increases linearly from 10−5 to 1
anism as follows: during a warmup phase lasting 2000 updates, and
0 then varies from a maximum of 1 to a minimum
P =LayerNorm(P )
of 10−5 according to a Cyclical Learning Rate
0
F =LayerNorm(F ) (Smith, 2017) policy with cosine scheme, with a
0 0 0
O =attnM M H (P , F , F , Mp )+ period of 2000 updates. With each cycle, the pe-
0
attnM M H (F 0
, P 0
, P , Mf ) + P + F riod is multiplied by 1.5 while both the maximum
(4) and minimum values are halved. We train until
convergence.
O goes through the 2-layer feedforward to produce We implement the system and training logic in
contextualized embeddings, which are used as in- pytorch with the help of the fairseq library.
put for the task-specific Prediction Heads. We de-
scribe them in the relevant paragraphs of Sections 4.2 Comparison Systems
4.1 and 5. In our experiments we compare QBERT with three
different contextualized embeddings systems:
4 Experimental Setup
1. Off-the-shelf pretrained ELMo (Peters et al.,
In what follows we first describe the Encoder ar-
2018). We employ a model featuring 4096-
chitecture hyperparameters and CLM pretraining
sized bidirectional LSTMs and 512-sized
details (Section 4.1). In Section 4.2 we describe
contextualized embeddings1 , pretrained on
the contextualized embeddings systems we use as
the concatenation of a Wikipedia dump and a
comparison in the WSD and Word-in-Context ex-
few English monolingual news corpora2 , for
periments. Finally, we report the setup and results
a total of 5.5B tokens.
of the experiments in Section 5.
126
5 Evaluation tasks Training and Test Data For each comparison
system we train two WSD classifiers, one using
As the first and main experiment we train and eval- only SemCor as training corpus and the other us-
uate a WSD Transformer classifier (Section 5.1.1) ing the concatenation of SemCor and the corpus
using QBERT and comparison contextualized em- of WordNet’s Tagged Glosses5 (WTG). WTG in-
beddings. To corroborate the results, as further ex- cludes 117659 manually disambiguated WordNet
periments we evaluate the performance of the con- synset glosses, with 496776 annotated tokens. We
textualized embeddings on the Word-in-Context test the performance of the models on the En-
task (Pilehvar and Camacho-Collados, 2019) (Sec- glish all-words evaluation datasets from the Sen-
tion 5.2). sEval and SemEval WSD evaluation campaigns,
namely Senseval-2 (Edmonds and Cotton, 2001),
5.1 Word Sense Disambiguation Senseval-3 (Snyder and Palmer, 2004), SemEval-
07 (Pradhan et al., 2007), SemEval-13 (Navigli
5.1.1 Setup
et al., 2013), SemEval-15 (Moro and Navigli,
To perform our WSD experiment we train a sim- 2015) and their concatenation (ALL). We use
ple Transformer-based classifier, which we eval- SemEval-2015 as our development set, to select
uate on all-words WSD benchmark datasets. We the best epoch of the run. We use the version of
use F1 on the test set as a measure of performance. SemCor and the evaluation datasets included in the
WSD framework6 of Raganato et al. (2017b).
Architecture Our Transformer classifier takes
as input the w-weighted mean between the word 5.1.2 Results
embeddings produced by the Encoder (the Adap- We show in Table 1 and Table 2 the results of
tive Input layer in the case of QBERT, the the evaluation on all-words WSD of the Predic-
character-level CNN in the case of ELMo) and tion Head trained on top of QBERT and the com-
the contextualized embeddings. We freeze the En- parison systems. Our best model beats all the
coder and only train w and the weights of the previously established results on all evaluation
Transformer classifier. As Flair has no word em- datasets. While the performance of the systems
beddings, we concatenate the outputs of the for- using SBERT and ELMo are also very competi-
ward and backward models with GloVe embed- tive, in many cases exceeding the state of the art,
dings (Pennington et al., 2014), and substitute the QBERT consistently outperforms them, achieving
weighted mean with a dense layer projecting the one of the largest performance gains in years.
concatenated matrix to the Transformer hidden
dimension. The classifier produces a probabil- SemCor If we restrict the comparison to mod-
ity distribution over an output vocabulary which els trained on SemCor (Table 1), QBERT beats
includes all the possible synsets plus a special the state of the art with a margin of 0.7 points on
<untagged> symbol for words with no associ- Semeval-07 and 1.5 points on SemEval-13. On
ated tag. During training only, we treat monose- our development set, SemEval-15, we get a score
mous words as tagged. At test time, we predict the 2 points over the state of the art. On Semeval-2
synset with the highest probability among those and Senseval-3 our F1 score is in the same ballpark
associated with the lemma of the target word. Im- as, respectively, Yuan et al. (2016) and Uslu et al.
portantly, we do not employ any Most Frequent (2018). QBERT also performs well measured
Sense backoff strategy. against our comparison systems. SBERT achieves
lower performance across the board, but attains
Hyperparameters All models are trained with overall competitive results on all the datasets.
Adam for a maximum of 60 epochs. We use a sim- ELMo performs on a par with SBERT on the con-
ilar learning rate scheduling scheme as in the CLM catenation of all datasets, but gets better results
training, first linearly increasing the value from than QBERT on the development set. Flair, per-
10−5 to 10−3 , then using a cosine CLR scheduler haps as a result of its purely character-based na-
with period 200, maximum learning rate 10−3 and ture, is severely outperformed on most datasets.
minimum learning rate 10−4 ; with each cycle the 5
http://wordnetcode.princeton.edu/
maximum and minimum are halved, while the pe- glosstag.shtml
riod is doubled. 6
http://lcl.uniroma1.it/wsdeval/
127
Systems Dev. set S2 S3 S07 S13 S15 ALL
IMS (Melacci et al., 2018) – 0.702 0.688 0.622 0.653 0.693 0.681
IMSWE (Melacci et al., 2018) – 0.722 0.699 0.629 0.662 0.719 0.696
IMSC2V+P R (Melacci et al., 2018) – 0.738 0.719 0.633 0.682 0.728 0.713
supWSDEmb (Papandrea et al., 2017) – 0.727 0.706 0.631 0.668 0.718 –
BiLSTMatt+lex (Raganato et al., 2017a) S07 0.720 0.694 0.637 0.664 0.724 0.699
GASext (concat) (Luo et al., 2018) S07 0.722 0.705 – 0.672 0.726 0.706
BiLSTM (Vial et al., 2018) WTG 0.735† 0.709† 0.625† 0.676† 0.716† 0.705†
BiLSTM+VR (ensemble) (Vial et al., 2018) WTG 0.731 0.706 0.613 0.712 0.716 0.718
LSTM+LP (Yuan et al., 2016) – 0.738 0.718 0.635 0.695 0.726 –
fastSense (Uslu et al., 2018) S2 0.735 0.735 0.624 0.662 0.732 –
SotA (single model) – 0.735 0.735 0.637 0.695 0.728 0.713
SotA (ensemble) – 0.735 0.735 0.637 0.712 0.728 0.718
ELMo + WSD Pred. Head S15 0.719 0.718 0.607 0.703 0.762 0.714
Flair + WSD Pred. Head S15 0.702 0.702 0.615 0.694 0.732 0.699
SBERT + WSD Pred. Head S15 0.731 0.719 0.640 0.694 0.741 0.715
QBERT + WSD Pred. Head? S15 0.734 0.732 0.644 0.710 0.743 0.724
Table 1: Results of the evaluation on the English datasets of models trained on SemCor. We include as
competitors supervised systems capable of performing all-words WSD on the whole WordNet inventory.
We report in the ‘Dev set.’ column the development corpus used (if any). The † symbol indicates that
the result is an average of 20 training runs. Bold means that the result is the highest one among non
ensemble models. We use ? to mark significant improvement against best single model performance
on ALL according to a z-test (p < 0.05). We report in the four row blocks 1) competitor SVM-based
systems; 2) competitor neural networks; 3) state of the art as the maximum value in the previous rows;
4) QBERT and our comparison systems.
Table 2: Results of the evaluation on the English datasets of models trained on the concatenation of
SemCor and WTG. We use the same notation as in Table 1, employing ? to mark significance against
the single model state of the art. Models from Vial et al. (2018), marked by SMP, use a random sample
of sentences from SemCor and WTG as development. In the row blocks we report 1) competitor neural
networks; 2) the state of the art as the maximum value in the previous rows and in Table 1; 3) QBERT
and our comparison systems.
SemCor and WTG When we report in the com- ALL. With respect to our own comparison sys-
parison systems trained on the concatenation of tems, QBERT performs better than ELMo, Flair
SemCor and WTG (Table 2), QBERT beats the and SBERT in this setting as well. ELMo gets very
state of the art more consistently and by a larger competitive results compared to the previous state
margin. We reach 1.3 points above the previ- of the art, which it beats on many datasets. Com-
ous state of the art on Senseval-2, 0.4 points on pared to QBERT, however, it gets worse results
Senseval-3, 2.4 on Semeval-07, 3.8 on Semeval- on almost every dataset, with the single excep-
13 and 4.6 on Semeval-15 (which is however our tion of SemEval-13. Flair underperforms also in
development set). On the concatenation of all this setting. SBERT achieves good performances,
datasets, our margin is of 3 points. Even if we con- but still consistently lower than QBERT, except
sider the ensemble of 20 models trained on Sem- for SemEval-07, which is however a small dataset
Cor and WTG by Vial et al. (2018), we get bet- whose F1 scores show high variance across differ-
ter results on every dataset with the exception of ent training runs.
SemEval-07, with a difference of 2.2 points on
128
5.2 Word-in-Context System Acc. µ Acc. σ
Elmo (ours) 59.97 1.41
Flair 60.23 0.91
The Word-in-Context task (WiC) was recently SBERT 60.03 1.13
established by Pilehvar and Camacho-Collados QBERT 60.74 1.22
(2019). Like WSD, WiC requires identification
Table 3: Results of our ELMo, SBERT and
of a contextually appropriate meaning, but it is
QBERT models on the WiC dataset evaluation
framed as a simpler binary classification task:
dataset. We report the mean and standard devia-
given two contextual occurrences of the same
tion of the accuracy for 5 runs.
lemma, predict whether the pair shares the same
sense. The dataset includes 8320 context pairs, di-
vided between training and development (the test In this setting ELMo performs on a par with
set has not yet been released). By including the SBERT and Flair, while QBERT achieves the best
same target word in each element of the pair, the result. Note that the quasi-deeply bidirectional en-
dataset is constructed in such a way that context- coding that QBERT can exploit through the Bi-
insensitive word embeddings would not perform Transformer might see its effectiveness reduced in
better than the random baseline. Thus, the dataset this setting since many pairs feature limited con-
is an ideal evaluation set for assessing the quality text, even as short as 2 or 3 words. Still, the re-
of the semantic information encoded in contextu- sults of the WiC task corroborate those of the all-
alized embeddings. words WSD, providing evidence that joint encod-
ing is crucial to better performance in word-level
semantics.
5.2.1 Setup
6 Conclusion
Among the baselines offered by Pilehvar and
Camacho-Collados (2019), one uses ELMo con- In this paper we showed that the use of contextu-
textualized embeddings as input to a simple two- alized embeddings enables a WSD system to beat
layer feed-forward classifier. We replicate the the previous state of the art. Moreover, we demon-
same setting, but using the concatenation of strated that the use of the BiTransformer coat-
QBERT Encoder word and contextualized embed- tentive mechanism in the QBERT contextualized
dings as input instead. Also, as the authors have embeddings model itself results in even stronger
not yet released the gold keys for the development performance. As a result, we attain one of the
set and evaluation can only be performed by up- largest gains in WSD performance in years, with
loading a prediction file to the Codalab competi- a margin of 3 points over the best reported sin-
tion page7 , we take 10
1
of the training instances as gle model on the concatenation of all datasets, and
our development set, and use the provided devel- of 2.2 points over the best ensemble model in the
opment set for testing. We train the system for a literature. We leave for future work the assess-
max of 40 epochs, submitting the epoch with best ment of whether the gains brought about by the
accuracy on the development split. WiC’s scorer use of the BiTransformer in QBERT carry over
reports the accuracy calculated on the predictions. to other tasks, helping to bridge the gap between
We implement the same system employing ELMo, CLM-based and fully bidirectional MLM-based
Flair (using the concatenation of GloVe and con- contextualized embeddings. We release the code
textualized embeddings) and SBERT as well. Per- to train the QBERT Encoder and the WSD clas-
formance is measured by mean accuracy over 5 sifier, along with pretrained models at https:
runs. //github.com/mbevila/qbert.
Acknowledgments
5.2.2 Results
The authors gratefully acknowledge
In Table 3 we show the results of the evaluation on the support of the ERC Consolida-
the WiC development set. tor Grant MOUSSE No. 726487 un-
der the European Union’s Horizon
7
https://competitions.codalab.org/
2020 research and innovation pro-
competitions/20010 gramme.
129
References Jeremy Howard and Sebastian Ruder. 2018. Uni-
versal Language Model Fine-tuning for Text
Alan Akbik, Duncan Blythe, and Roland Vollgraf. Classification. In Proc. of ACL. pages 328–339.
2018. Contextual String Embeddings for Se- https://www.aclweb.org/anthology/papers/P/P18/P18-
quence Labeling. In Proc. of COLING. pages 1031/.
1638–1649. https://aclanthology.info/papers/C18-
1139/c18-1139. Ignacio Iacobacci, Mohammad Taher Pilehvar,
and Roberto Navigli. 2016. Embeddings for
Lei Jimmy Ba, Ryan Kiros, and Geoffrey E. Hinton. Word Sense Disambiguation: An Evalua-
2016. Layer normalization. CoRR abs/1607.06450. tion Study. In Proc. of ACL. pages 897–907.
http://arxiv.org/abs/1607.06450. http://aclweb.org/anthology/P/P16/P16-1085.pdf.
Alexei Baevski and Michael Auli. 2018. Adap- Mikael Kagebäck and Hans Salomonsson. 2016.
tive input representations for neural lan- Word Sense Disambiguation using a Bidirectional
guage modeling. CoRR abs/1809.10853. LSTM. In Proc. of COLING. pages 51–56.
http://arxiv.org/abs/1809.10853. https://aclanthology.info/papers/W16-5307/w16-
5307.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben-
gio. 2015. Neural machine translation by jointly Guillaume Lample and Alexis Conneau. 2019. Cross-
learning to align and translate. In Proc. of ICLR. lingual language model pretraining. CoRR
https://arxiv.org/abs/1409.0473. abs/1901.07291. http://arxiv.org/abs/1901.07291.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, Minh Le, Marten Postma, Jacopo Urbani, and Piek
and Kristina Toutanova. 2019. BERT: Vossen. 2018. A Deep Dive into Word Sense Dis-
Pre-training of Deep Bidirectional Trans- ambiguation with LSTM. In Proc. of COLING.
formers for Language Understanding. In Association for Computational Linguistics, pages
Proc. of NAACL-HLT. pages 4171–4186. 354–365. https://aclanthology.info/papers/C18-
https://aclweb.org/anthology/papers/N/N19/N19- 1030/c18-1030.
1423/.
Fuli Luo, Tianyu Liu, Qiaolin Xia, Baobao Chang, and
Philip Edmonds and Scott Cotton. 2001. SENSEVAL- Zhifang Sui. 2018. Incorporating Glosses into Neu-
2: Overview. In Proc. of SENSEVAL-2. pages 1–5. ral Word Sense Disambiguation. In Proc. of ACL.
https://www.aclweb.org/anthology/papers/S/S01/S01- Association for Computational Linguistics, pages
1001/. 2473–2482. https://aclanthology.info/papers/P18-
1230/p18-1230.
Christiane Fellbaum. 1998. WordNet: An Electronic
Lexical Database. Language, Speech, and Commu- Thang Luong, Hieu Pham, and Christopher D.
nication. MIT. Manning. 2015. Effective Approaches to
Attention-based Neural Machine Transla-
Edouard Grave, Armand Joulin, Moustapha tion. In Proc. of EMNLP. Association for
Cissé, David Grangier, and Hervé Jégou. Computational Linguistics, pages 1412–1421.
2017. Efficient softmax approximation for http://aclweb.org/anthology/D/D15/D15-1166.pdf.
GPUs. In Proc. of ICML. pages 1302–1310.
http://proceedings.mlr.press/v70/grave17a.html. Stefano Melacci, Achille Globo, and Leonardo
Rigutini. 2018. Enhancing Modern Su-
Lushan Han, Abhay L. Kashyap, Tim Finin, pervised Word Sense Disambiguation
James Mayfield, and Jonathan Weese. 2013. Models by Semantic Lexical Resources.
UMBC ebiquity-CORE: Semantic Textual In Proc. of LREC. pages 1012–1017.
Similarity Systems. In Proceedings of the https://www.aclweb.org/anthology/papers/L/L18/L18-
Second Joint Conference on Lexical and 1163/.
Computational Semantics. pages 44–52.
http://aclweb.org/anthology/S/S13/S13-1005.pdf. Oren Melamud, Jacob Goldberger, and Ido Da-
gan. 2016. context2vec: Learning Generic
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Context Embedding with Bidirectional
Sun. 2016. Deep Residual Learning for Image LSTM. In Proc. of COLING. pages 51–61.
Recognition. In Proc. of CVPR. pages 770–778. http://aclweb.org/anthology/K/K16/K16-1006.pdf.
https://doi.org/10.1109/CVPR.2016.90.
Andrea Moro and Roberto Navigli. 2015. SemEval-
Luheng He, Kenton Lee, Omer Levy, and Luke 2015 Task 13: Multilingual All-Words Sense Dis-
Zettlemoyer. 2018. Jointly Predicting Predi- ambiguation and Entity Linking. In Proceedings
cates and Arguments in Neural Semantic Role of the 9th International Workshop on Semantic
Labeling. In Proc. of ACL. pages 364–369. Evaluation, SemEval@NAACL-HLT 2015, Denver,
https://aclanthology.info/papers/P18-2058/p18- Colorado, USA, June 4-5, 2015. pages 288–297.
2058. http://aclweb.org/anthology/S/S15/S15-2049.pdf.
130
Roberto Navigli. 2009. Word sense disambiguation: Alessandro Raganato, Jose Camacho-Collados, and
A survey. ACM Comput. Surv. 41(2):10:1–10:69. Roberto Navigli. 2017b. Word Sense Disambigua-
https://doi.org/10.1145/1459352.1459355. tion: A Unified Evaluation Framework and Empir-
ical Comparison. In Proc. of ACL. pages 99–110.
Roberto Navigli, David Jurgens, and Daniele Vannella. https://www.aclweb.org/anthology/papers/E/E17/E17-
2013. SemEval-2013 Task 12: Multilingual Word 1010/.
Sense Disambiguation. In Proc. of SemEval. pages
222–231. http://aclweb.org/anthology/S/S13/S13- Rico Sennrich, Barry Haddow, and Alexandra Birch.
2040.pdf. 2016. Neural Machine Translation of Rare Words
with Subword Units. In Proc. of ACL. pages 1717–
Simone Papandrea, Alessandro Raganato, and Clau- 1725. http://aclweb.org/anthology/P/P16/P16-
dio Delli Bovi. 2017. SupWSD: A Flex- 1162.pdf.
ible Toolkit for Supervised Word Sense Dis-
ambiguation. In Proc. of EMNLP. pages Leslie N. Smith. 2017. Cyclical Learning Rates for
103–108. https://aclanthology.info/papers/D17- Training Neural Networks. In Proc. of WACV. pages
2018/d17-2018. 464–472. https://doi.org/10.1109/WACV.2017.58.
Tommaso Pasini and Roberto Navigli. 2017. Train- Benjamin Snyder and Martha Palmer. 2004. The En-
O-Matic: Large-Scale Supervised Word Sense Dis- glish all-words task. In Proc. of SENSEVAL-3.
ambiguation in Multiple Languages without Man- pages 41–43. https://aclanthology.info/papers/W04-
ual Training Data. In Proc. of EMNLP. pages 78– 0811/w04-0811.
88. https://aclanthology.info/papers/D17-1008/d17-
Kaveh Taghipour and Hwee Tou Ng. 2015. One Mil-
1008.
lion Sense-Tagged Instances for Word Sense Disam-
Jeffrey Pennington, Richard Socher, and biguation and Induction. In Proc. of CoNLL. pages
Christopher D. Manning. 2014. Glove: 338–344. http://aclweb.org/anthology/K/K15/K15-
Global Vectors for Word Representation. 1037.pdf.
In Proc. of EMNLP. pages 1532–1543. Tolga Uslu, Alexander Mehler, Daniel Baumartz,
http://aclweb.org/anthology/D/D14/D14-1162.pdf. Alexander Henlein, and Wahed Hemati. 2018. Fast-
Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Sense: An Efficient Word Sense Disambiguation
Gardner, Christopher Clark, Kenton Lee, and Luke Classifier. In Proc. of LREC. pages 1042–1046.
Zettlemoyer. 2018. Deep Contextualized Word https://www.aclweb.org/anthology/papers/L/L18/L18-
Representations. In Proc. of NAACL-HLT. pages 1168/.
2227–2237. https://aclanthology.info/papers/N18- Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
1202/n18-1202. Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
Kaiser, and Illia Polosukhin. 2017. Attention is
Mohammad Taher Pilehvar and José Camacho-
All you Need. In Proc. of NIPS. pages 6000–
Collados. 2019. Wic: the word-in-context dataset
6010. https://papers.nips.cc/paper/7181-attention-
for evaluating context-sensitive meaning represen-
is-all-you-need.pdf.
tations. In Proc. of NAACL-HLT. pages 1267–1273.
https://aclweb.org/anthology/papers/N/N19/N19- Loı̈c Vial, Benjamin Lecouteux, and Didier
1128/. Schwab. 2018. Improving the coverage and
the generalization ability of neural word sense
Sameer Pradhan, Edward Loper, Dmitriy Dligach, disambiguation through hypernymy and hy-
and Martha Palmer. 2007. SemEval-2007 Task- ponymy relationships. CoRR abs/1811.00960.
17: English Lexical Sample, SRL and All http://arxiv.org/abs/1811.00960.
Words. In Proc. of SemEval 2007. pages 87–92.
http://aclweb.org/anthology/S07-1016. Dirk Weissenborn, Leonhard Hennig, Feiyu Xu, and
Hans Uszkoreit. 2015. Multi-Objective Optimiza-
Ofir Press and Lior Wolf. 2017. Using the Out- tion for the Joint Disambiguation of Nouns and
put Embedding to Improve Language Mod- Named Entities. In Proc. of ACL. pages 596–605.
els. In Proc. of EACL. pages 157–163. http://aclweb.org/anthology/P/P15/P15-1058.pdf.
https://aclanthology.info/papers/E17-2025/e17-
2025. Dayu Yuan, Julian Richardson, Ryan Doherty, Colin
Evans, and Eric Altendorf. 2016. Semi-supervised
Alec Radford, Karthik Narasimhan, Tim Salimans, and Word Sense Disambiguation with Neural Mod-
Ilya Sutskever. 2018. Improving Language Under- els. In Proc. of COLING. pages 1374–1385.
standing by Generative Pre-Training . http://aclweb.org/anthology/C/C16/C16-1130.pdf.
Alessandro Raganato, Claudio Delli Bovi, and Zhi Zhong and Hwee Tou Ng. 2010. It Makes
Roberto Navigli. 2017a. Neural Sequence Sense: A Wide-Coverage Word Sense Disam-
Learning Models for Word Sense Disambigua- biguation System for Free Text. In Proc.
tion. In Proc. of EMNLP. pages 1156–1167. of ACL, System Demonstrations. pages 78–83.
https://aclanthology.info/papers/D17-1120/d17- http://www.aclweb.org/anthology/P10-4014.
1120.
131
Evaluating the Consistency of Word Embeddings from Small Data
132
Proceedings of Recent Advances in Natural Language Processing, pages 132–141,
Varna, Bulgaria, Sep 2–4 2019.
sistency. We show that in spite of being a simple sic evaluation (Schnabel et al., 2015) are interest-
evaluation, consistency actually depends on vari- ing, but require domain experts, as well as instruc-
ous combinations of factors, including the nature tions that elicit the desired type of semantic rela-
of the data itself, the model used to train the seman- tion (i.e. similarity). Extrinsic evaluation requires a
tic space, and the frequency of the learned terms, downstream task that can be evaluated, but in this
both in the background space and in the in-domain use case we are interested in the information en-
data of interest. This leads us to conclude that the coded by the DS model itself. QVEC (Tsvetkov
evaluation of in-domain word embeddings from et al., 2015) evaluates by aligning dimensions of
small data has to be controlled extremely carefully a semantic space to linguistic features, but we are
in order not to draw incorrect conclusions from ex- interested only in evaluating some vectors rather
perimental results. than an entire space (target term vectors but not
the background space vectors), and this approach
2 Related Work requires language-specific resources.
Nooralahzadeh et al. (2018) evaluate domain-
Learning embeddings for rare words is a very chal- specific embeddings by building a query inven-
lenging process (Luong et al., 2013). Word2Vec tory for their domain from a glossary containing
(W2V, Mikolov et al., 2013a)’s skipgram model synonym, antonym and alternative form informa-
can learn embeddings from tiny data after modifica- tion. Unfortunately, such structured glossaries are
tion, as shown by Herbelot and Baroni (2017) when generally not available for specific philosophers.
it consists of just a single highly informative defini- Hellrich and Hahn (2016) test their models for re-
tional sentence. However, philosophical data is typ- liability in a study investigating historical English
ically small data rather than tiny data. While tiny and German texts, another relatively low-resource
data consists of a single definitional sentence, our domain. Their reliability metric involves training
small data consists of multiple context sentences three identically parametrized models, and compar-
per term that are not necessarily definitional. Her- ing the nearest neighbors of each word in each
belot and Baroni’s (2017) Nonce2Vec (N2V) has model using a modified Jaccard coefficient. This
not been tested on this type of data. W2V has been metric does not require any language-specific data,
tested on smaller datasets, but was found to be sub- but it mainly serves as a test of the impact of the
optimal (Asr et al., 2016) and surpassed by SVD sources of randomness in Word2Vec, and not as
models on a 1 million word dataset (Sahlgren and a measure of the systematic semantic differences
Lenci, 2016). across various data sources.
Different DS evaluations test different aspects
of the learned embeddings (i.e. Wang et al., 2019). 3 Consistency Metric
Most existing methods are however not easily ap-
plicable to our task. The typical evaluation of com- We propose consistency as a useful metric to eval-
paring embedding similarities to a gold standard uate word embeddings in the absence of domain-
of word similarity scores, such as the SimLex-999 specific evaluation datasets. We consider a model
dataset (Hill et al., 2015) cannot be applied, be- to be consistent if its output does not vary when its
cause we are interested in the representation of input should not trigger variation (i.e. because it is
specific terms: even if these terms are present in sampled from the same text). Thus, a model can
the evaluation set, their meaning in the philosophi- only be as consistent as the input data it is trained
cal domain is likely to differ. Manually creating a on and it requires the experimenter to compute data
domain-specific resource requires labor-intensive consistency in addition to vector space consistency.
effort by domain experts, which makes it imprac- To evaluate data consistency, we create vectors
tical to port standard datasets to a specific type of for target terms in a domain corpus under two con-
corpora. This holds also for other evaluation meth- ditions: a) random sampling; b) equal split. The
ods such as analogy scores (Mikolov et al., 2013b), ‘equal split’ condition simply corresponds to split-
as well as coherence (Schnabel et al., 2015), which ting the data in the middle, thus obtaining two
is based on the idea that pairs of similar words subcorpora of equal size and in diachronic order.
should be close in semantic space. Given a pre-trained background space kept frozen
Methods where human raters directly respond across experiments, the vector representation of a
to output of the model, such as comparative intrin- target is generated by simple vector addition over
133
its context words. Therefore, the obtained vector 4 Task Description
directly represents the context the target term oc-
curs in, and consequently, similar representations Our overall aim is to obtain consistent embeddings
(in terms of cosine similarity) mean that the target of terms central to the works of Willard Van Or-
term is used in a similar way in different parts of man Quine (1908-2000), an influential 20th cen-
a book/corpus, and is thus consistently learnable. tury philosopher and logician. As the meaning of
Crucially, though, this measure may interact with terms may differ between authors and even be-
data size. Kabbach et al. (2019) recently noted a tween books by the same author, we need to learn
sum effect in the additive model, where summed such embeddings from small data, bounded by the
vectors are close to each other. It may be the case occurrences of the term in one particular book.
that additive model vectors summed over more con-
text data contain more information and may have Quine datasets We build novel datasets based
higher similarity between each other, resulting in on a corpus of 228 philosophical articles, books
higher consistency scores. We test this in Section 6. and bundles written by Quine, with a focus om
When randomly sampling, we limit the number of two of Quine’s books: A System of Logistic (Quine,
sentences per sample to control for this. 1934) and Word & Object (Quine, 1960). These
Quine texts are part of a larger corpus of philo-
To evaluate space consistency, we create iden- sophical texts, which is still being compiled, that
tically parametrized models as in Hellrich and are central to the history of scientific ideas (Betti
Hahn’s (2016) reliability metric, but over different and van den Berg, 2016). We focus on these partic-
parts of the data, with the data being split in the ular works from the corpus because testing consis-
middle, as just described. We consider two ways tency is best done on homogeneous data, and our
of comparing two vectors a~1 and a~2 : by similar- philosophy domain experts informed us that Quine
ity, where a higher cosine similarity indicates more was a remarkably stable philosopher in his outlook
consistency, or by nearest neighbor rank, where a (Betti and Oortwijn, p.c.).
higher rank of a~1 among the nearest neighbors of The first book is a formula-heavy logic book, de-
a~2 indicates more consistency. Every vector in the viating strongly from ordinary language. Such a
background space, as well as a~2 , is ranked by co- technical book is particularly likely to be internally
sine similarity to a~1 to compute this rank value. consistent. It contains 80,279 tokens after tokeniza-
tion and manual replacement of formulas with spe-
Although it is more complex than having a sin-
cial tokens. The second book is more textual and
gle metric, we must consider both rank and similar-
consists of standard philosophical argumentation.
ity simultaneously: rank is a more relative metric
Our domain experts consider it conceptually con-
and helps to ground the similarity value in the local
sistent. It contains 133,240 tokens after tokeniza-
context of the target term. A vector with 0.8 simi-
tion. The full corpus of the 228 Quine articles con-
larity but lower rank is a worse result than a vector
tains 1.7 million tokens and is pre-processed with
with 0.8 similarity and a high rank, as the low rank
NLTK (Bird et al., 2009) for sentence splitting and
means that the vectors are in a dense part of the se-
tokenization. A less common preprocessing step
mantic space and a very high similarity is required
we took was to remove one-character tokens from
to consistently identify which of the neighbouring
the texts. These works contain many one-letter vari-
vectors refers to the same concept. Conversely, a
able names, logical symbols and other formal lan-
low-similarity, high-rank vector can be a cause for
guage that a model might otherwise use to position
scepticism, as it may have been placed far out from
vectors of Quine terminology in particular areas
the rest of the semantic space.
of the semantic space, as these tokens are highly
We take consistency to be a desirable property of infrequent in the general domain.
word embeddings at the level of a certain domain. To obtain terms that would be relevant to model,
Of course, consistency only measures one specific we automatically extract terms from the two books’
desirable property of embeddings and should thus indexes, as the most important terminology is
not be interpreted as a general quality or accuracy likely to be listed there. We include multi-word
score. But even taken on its own, we will show that terms, and divide the terms into 70%/30% subsets
it exhibits complex behavior with respect to data, for training and testing, resulting in a 157 / 67 split
background vectors and term frequency. for Logistic and a 184 / 79 split for Word & Ob-
134
ject. The target terms thus differ per book, as each Section 3. To obtain an embedding for a new tar-
book lists different terms in its index. Instead of get term, we use an additive model over its con-
this automatic approach to obtaining target terms, text words, using as background space ordinary lan-
an expert-created resource could provide a better guage representations. For the in-domain context,
set of target terms, if available. If neither this nor a we use a window size of 15, with the window being
terms glossary or index of terms is available, key- restricted to the sentence. The background space
word extraction methods could be used as an al- is based on a Wikipedia snapshot of 1.6B words
ternative way of obtaining terms for evaluation. In trained with Word2Vec’s Gensim implementation
cases where the model will not be used for any with default parameters, and containing 259,376
analysis of domain-specific content downstream, it word vectors in 100 dimensions. For each target
may be sufficient to randomly select words from term, context words undergo subsampling, which
the text as target terms. randomly drops higher-frequency words.3 The vec-
Next, we derive datasets from this corpus using tors of the remaining context words are summed
our two conditions for data consistency: random to create a vector for the target term. This additive
sampling and equal split. In random sampling, for model was used by Lazaridou et al. (2017) for their
each target term that meets a frequency cutoff of 10, textual data, and was shown by Herbelot and Ba-
we randomly select five non-overlapping samples roni (2017) to work reasonably well on tiny data.
of up to 10 random sentences that contain the target We calculate the vectors separately per sample (or
term, divided evenly across the samples if the term book half), yielding comparable term vectors.
occurs in fewer than 50 sentences. This gives us the Next, we turn to space consistency. We use our
datasets Quine-WordObject-rnd (with Word & Ob- consistency metric to evaluate two models that are
ject core terms as target terms), Quine-Logistic-rnd suited to learning embeddings from small data:
(with System of Logistic core terms) for our two Nonce2Vec (Herbelot and Baroni, 2017) and an
books of interest, and Quine-all-rnd sampled from SVD-reduced count-based model over concatena-
the full Quine corpus, where we also use the Word tions of our datasets with general-domain data.
& Object core terms as target terms.1 In the equal The first model, Nonce2Vec, modifies W2V’s
split condition, we divide a book into two halves at ‘skip-gram’ model (Mikolov et al., 2013a) in a
a chapter boundary, and extract all sentences con- way that is inspired by fast mapping (Carey and
taining index terms that meet a frequency cuf-off Bartlett, 1978) in humans. Human learners can
of 2 in each half, resulting in the datasets Quine- acquire new words from just a single token and
WordObject and Quine-Logistic. With random sam- this process of fast mapping appears to build on
pling, we intend to capture the basic consistency concepts that are already known (Coutanche and
of the model. With equal split, we aim to capture Thompson-Schill, 2014). Nonce2Vec models this
consistency across potential meaning development through incremental learning, an initial high learn-
throughout the book.2 ing rate, greedy processing, and parameter decay.
To simulate the existence of background knowl-
Wikipedia dataset For cross-domain compari-
edge, Nonce2Vec maps its novel word vectors into
son, we apply our method to a 140M word pre-
a previously learned semantic space, based on the
processed Wikipedia snapshot using the same ran-
aforementioned Wikipedia snapshot and the same
dom sampling process. As target terms, we used
subsampling procedure. Target term vectors are ini-
300 randomly sampled one-word Wikipedia page
tialized to their sum vector from the additive model.
titles, following Herbelot and Baroni (2017).
For each sentence, the model is trained on the target
term, only updating the weights for that term and
5 Method
freezing all other network parameters. The learning
Before evaluating whether we have space consis- rate and context window size decay in proportion
tency, we must establish to what extent we have to the number of times the target term has been
data consistency, following our argumentation in seen, and the subsampling rate increases per sen-
1
tence.
Word & Object touches upon much of Quine’s work, so
its terminology can be considered representative. Secondly, we try a count-based approach, creat-
2
While our datasets are derived from copyrighted works
3
and cannot be shared, we provide replication instructions, Some promising alternative subsampling methods for tiny
term lists and code here: https://bloemj.github.io/quine2vec/ data were recently discussed by Kabbach et al. (2019).
135
Dataset cos-sim rank n cos-sim
Quine-WordObject 0.938 1 1 0.794
Quine-Logistic 0.907 22.4 2 0.837
Quine-WordObject-rnd 0.919 1 3 0.905
Quine-Logistic-rnd 0.935 1 4 0.923
Quine-all-rnd 0.953 1 8 0.956
Wiki-rnd 0.927 1.001 all 0.987
Table 1: Consistency metrics on different data sets Table 2: Data consistency for the term analytical
using the additive model. hypotheses in Word & Object when varying the
number of sentences per sample n.
ing vectors over the general-domain and in-domain
data at the same time. In this procedure, we con- sistency score than random sampling from that
catenate a particular Quine dataset with a 140M book (Quine-Logistic-rnd), possibly because the
word Wikipedia corpus sample, in which the Quine meaning of a term may evolve from the first half
terms are marked with special tokens in order to the second half of a book. This suggests some
to be trained separately from the same term in utility of the data consistency measure in quanti-
the Wikipedia data. We create embeddings from fying meaning development throughout a text, as
this corpus, applying PPMI weighting and singu- long as other factors are controlled for. We also see
lar value decomposition (SVD) to reduce the mod- that the Wikipedia domain data (Wiki-rnd) is less
els to 100 dimensions, to match the dimensionality consistent than the Quine domain data (Quine-all-
of our other models and because factorized count rnd), which is to be expected as it contains more
models have been shown to work well on smaller diverse text.
datasets (Sahlgren and Lenci, 2016). We use the These results seem to indicate that the addi-
Hyperwords implementation of Levy et al. (2015), tive model provides consistent embeddings. This
with a window size of 5, and other hyperparameters means that it must be possible to learn consistent
set to the default values. embeddings from these datasets, at least up to the
In both the above approaches, we can then com- consistency values reported here, as the additive
pute vector space consistency between different model directly represents the contexts that predic-
vectors learned for the same term over different tive models use for training.
splits of the data. As already mentioned, however, the factor of
data size may interfere with consistency. We do
6 Consistency of Data
observe in Table 1 that the consistency of data
We start by applying the additive model to quantify sampled across the full Quine corpus is higher. Al-
data consistency on the different datasets described though we limited our samples to 10 sentences per
in Section 4. We compute average similarities and term, not every core Quine term is used frequently
nearest neighbor ranks over the vectors of all target enough to have 5 samples with the maximum size
terms in a dataset. For the randomly sampled data of 10 sentences. Specifically, in the full Quine
sets, we have five vectors per term, one from each dataset, 68.6% of terms reach the maximum size,
sample, and compute the metrics over all unique while in the Word & Object dataset, only 32.1%
combinations of 2 vectors. For the equal split set- of terms reach it. In the Wiki set, this is 90.9%,
ting, we compare the term vectors summed over showing that its lower consistency is not explained
each half of the book. by limited data. To fully control for data size, we
The additive model produces highly consistent would need to use artificial data: if we control for
embeddings on the training data: for most terms, the number of sentences, the number of words and
the vectors summed over each book half are each the number of words subsampled still affect data
other’s nearest neighbors in the background space. size. As we are mainly interested in the quality of
This trend is also observed for the test sets pre- models on our own philosophical corpus, we leave
sented in Table 1, where we observe high consis- this for future work.
tency for the embeddings from both books. Instead, we test the effect of data size by sum-
Using the book halves of System of Logistic ming two vectors for the same term over varying
(Quine-Logistic) gives us a slightly lower data con- numbers of sentences, and computing the consis-
136
tency between them. Table 2 shows a clear effect Dataset cos-sim rank
of data size: vectors summed over more sentences Quine-WordObject 0.686 1.21
have higher data consistency. This shows that data Quine-Logistic 0.748 1.48
consistency should ideally be computed within the Quine-WordObject-rnd 0.695 2.11
constraints of a particular data size, because vec- Quine-Logistic-rnd 0.743 1
tors summed over more context are more informa- Quine-all-rnd 0.717 1.59
tive and thus more consistent. Wiki-rnd 0.589 507.8
137
Group of terms similarity Dataset cos-sim rank
System of Logistic 0.323 Quine-WordObject-rnd 0.352 22,947
Word & Object 0.366 Quine-Logistic-rnd 0.353 24,513
Q-High-freq W-Low-freq 0.735 Quine-all-rnd 0.382 17,262
Q-Low-freq W-Low-freq 0.417 Wiki-rnd 0.475 2,902
Q-Low-freq W-High-freq 0.109
Q-High-freq W-High-freq 0.078 Table 6: Average similarities between learned
in-domain term vectors and pretrained general-
Table 5: Average similarities between Quine vec- domain background vector on different data sets
tors and Wiki vectors in our SVD model. Q = for the Nonce2Vec-based models.
Quine, W = Wiki.
ble to learn good representations from this data by
Wikipedia general-domain spaces and the Quine using background knowledge from the Wikipedia
terminology. To quantify the difference between domain, but the models we tested did not reach the
domains, we selected all sentences from the Quine level of consistency of the additive model.
corpus containing 35 target terms and concatenated For better or for worse, our models do move
them with our 140M word Wikipedia sample, as away from what is in the background space. In
in the previous experiment. These terms were se- our Nonce2Vec experiment on the Quine-all-rnd
lected to be either high-frequent or low-frequent in dataset, we also measured the average cosine simi-
the Quine domain and either high-frequent or low- larity and nearest neighbour rank of the pretrained
frequent in the general domain. Again, the Quine Word2Vec term vector from the background space,
terms were marked in order to be trained sepa- compared to the vectors we learned for that same
rately from the same term in the Wikipedia domain, term from the in-domain data. These numbers,
and we created a SVD model. In this SVD model, shown in Table 6, reveal that the model does not
we computed the cosine similarities between each stay close to the pre-trained background vectors
Quine term and its Wikipedia counterpart, and take in order to achieve high consistency, which could
this to be a measure of domain difference. be a risk if consistency was used as a learning
Table 5 shows a clear effect of term frequency. signal in combination with an invariant initializa-
We grouped all terms according to two factors: tion. Furthermore, the vectors learned from the
their frequency in the Quine book they were se- Wiki data are closer to the pre-trained vectors than
lected for (low, relative frequency4 < 0.0005 or those learned from the Quine data. This is expected
high, relative frequency > 0.001) and their fre- of a good model, as there is no domain gap to
quency in the Wikipedia domain (low, RF < bridge when training with Wikipedia context sen-
0.000025 or high, RF > 0.00005).5 We observe tences into a Wikipedia background space. This
that infrequent terms with a dominant philosophi- also means that the vector representations for terms
cal sense such as stimulus have more similar vec- as used by Quine become more distinct after train-
tors in both domains despite their sparsity in both ing, as our philosophy domain experts would ex-
corpora. Generally, terms that are highly frequent pect of a good meaning representation of these in-
in the Quine-domain but have low frequency in domain terms.
the Wikipedia domain are more similar between
We must again note that consistency is not the
the two domains (Q-High-freq W-Low-freq). To a
only desirable property of word embeddings. Un-
lesser extent, this is also true for terms that are low-
fortunately, other properties are more difficult to
frequent in both domains.
evaluate on low-resource data. Without a domain-
This result indicates that bridging the domain specific evaluation set, we can only explore issues
gap should be easier with these philosophical core with quality by examining nearest neighbors of
terms than with frequent Wikipedia terms. The fact vectors that our metric marks as perfectly consis-
that our models are less consistent on Wikipedia tent. We observe both in our results, illustrated
data also indicates that the generality of this do- by cherry-picked examples from the Nonce2Vec
main is more relevant than any specific differences model on the Quine-WordObject dataset. Table 7
with the Quine domain. It must therefore be possi- shows that the nearest neighbours for both book
4F
where F is the term frequency and C is the corpus size. half vectors for the term talking (Word & Object)
C
5
Different thresholds are necessary for the larger corpus. look bad. The vectors’ nearest neighbours are some
138
Term a~1 NNs a~2 NNs ical terminology, consistency helps us assess the
1 wrongfulness axiomatically quality of embeddings for philosophical terms,
talking 2 axiomatically epiphenomenon which may differ in meaning across a book or an
3 particularized impredicative author’s work, and for which no gold standard eval-
1 logophoric logophoric uation sets are available. These embeddings can
2 deverbal resumptive then be used to aid in the examination of large
verbs
3 adpositions countability volumes of philosophical text (Ginammi et al., in
4 uninflected adverbials press). Beyond our use case, the consistency met-
ric is quite broadly applicable — a relevant back-
Table 7: Qualitative examination of some nearest ground semantic space is necessary, but this can be
neighbours of target term vectors computed over constructed from out-of-domain data.
book halves 1 and 2 of Word & Object.
Like any metric, the consistency metric does not
apparently unrelated words yet they are closest to answer all of our questions about the quality of our
each other (similarity 0.751). We thus have high embeddings. Although the additive model is more
consistency, but not a good semantic representa- consistent than the others, both its dependence on
tion. The word verb is an example that does work: data size and the not-always-great qualitative re-
all nearest neighbours from the background space sults show that exploring other models is worth-
are linguistic terms. The two verbs vectors are also while for small data. Further research is required
closest to each other (similarity 0.625). to determine whether the representations produced
by the additive model are useful for downstream
8 Conclusion tasks. Using the knowledge of domain experts
in a structured evaluation task would be a good,
Our results show that it is possible to learn consis-
though resource-intensive, next step. Our metric
tent embeddings from small data in the context of
helps quantify the reliability of a model before in-
a low-resource domain, as such data provides con-
vesting more resources into evaluation.
sistent contexts to learn from. Applying an addi-
tive model that sums general-domain vectors from Our observation that the consistency metric de-
a pre-trained background space resulted in similar pends on a variety of other factors shows that con-
vectors for the same terms across different contexts sistency is a non-trivial aspect of the evaluation
from the same domain. The Nonce2Vec model also of distributional semantic models that should not
results in consistent embeddings that are closer to be overlooked. In future work, we will apply the
vectors of the same term trained on different con- consistency metric to evaluate other models, and
text sentences than to vectors of other terms. The datasets from other domains.
summed vectors from the additive model applied
to our philosophical small data are highly discrim-
inative, distinguishing the target terms from back- Acknowledgments
ground terms almost perfectly.
Our results show the benefits of using consis- We are grateful to Yvette Oortwijn and Arianna
tency as an intrinsic evaluation metric for dis- Betti for their input as Quine domain experts. We
tributional semantic models, particularly for low- also thank them as well as Lisa Dondorp, Thijs Os-
resource situations in which no gold standard sim- senkoppele and Maud van Lier for their work on
ilarity scores are available. While the metric may the Quine in Context corpus. We thank Pia Som-
appear simple, it proved useful both for evaluating merauer for help with the SVD setup, and the UvA
the homogeneity of a dataset and for evaluating e-Ideas group for their fruitful discussion of a draft
the stability of vector spaces generated by a given of this paper. We also thank the anonymous review-
model. Consistency turns out to depend on vari- ers for their time and valuable comments. This re-
ous combinations of factors, including the nature search was supported by VICI grant e-Ideas (277-
of the data itself, the model used to train the seman- 20-007) awarded to Arianna Betti and VENI grant
tic space, and the frequency of the learned terms, Reading between the lines 275-89-029 awarded to
both in the background space and in the in-domain Antske Fokkens, both financed by the Dutch Re-
data of interest. search Council (NWO).
For the specific purpose of modeling philosoph-
139
References Johannes Hellrich and Udo Hahn. 2016. Bad company
- neighborhoods in neural embedding spaces con-
Fatemeh Torabi Asr, Jon Willits, and Michael Jones. sidered harmful. In Proceedings of COLING 2016,
2016. Comparing predictive and co-occurrence the 26th International Conference on Computational
based models of lexical semantics trained on child- Linguistics: Technical Papers, pages 2785–2796.
directed speech. In CogSci.
Aurélie Herbelot and Marco Baroni. 2017. High-risk
Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and learning: acquiring new word vectors from tiny data.
Christian Jauvin. 2003. A neural probabilistic lan- In Proceedings of the 2017 Conference on Empiri-
guage model. Journal of machine learning research, cal Methods in Natural Language Processing, pages
3(Feb):1137–1155. 304–309.
Arianna Betti and Hein van den Berg. 2016. Towards Felix Hill, Roi Reichart, and Anna Korhonen. 2015.
a Computational History of Ideas. In Proceedings Simlex-999: Evaluating semantic models with (gen-
of the Third Conference on Digital Humanities in uine) similarity estimation. Computational Linguis-
Luxembourg with a Special Focus on Reading His- tics, 41(4):665–695.
torical Sources in the Digital Age. CEUR Workshop
Proceedings, CEUR-WS.org, volume 1681, Aachen. Alexandre Kabbach, Kristina Gulordava, and Aurélie
Herbelot. 2019. Towards incremental learning of
Arianna Betti, Hein van den Berg, Yvette Oortwijn, word embeddings using context informativeness. In
and Caspar Treijtel. 2019. History of Philosophy in Proceedings of the 57th Conference of the Asso-
Ones and Zeros. In Mark Curtis and Eugen Fischer, ciation for Computational Linguistics: Student Re-
editors, Methodological Advances in Experimental search Workshop, pages 162–168, Florence, Italy.
Philosophy, Advances in Experimental Philosophy, Association for Computational Linguistics.
pages 295–332. Bloomsbury, London.
Angeliki Lazaridou, Marco Marelli, and Marco Baroni.
Steven Bird, Ewan Klein, and Edward Loper. 2009. 2017. Multimodal word meaning induction from
Natural language processing with Python: analyzing minimal exposure to natural text. Cognitive science,
text with the natural language toolkit. ” O’Reilly 41:677–705.
Media, Inc.”.
Omer Levy, Yoav Goldberg, and Ido Dagan. 2015. Im-
Elia Bruni, Nam-Khanh Tran, and Marco Baroni. 2014. proving distributional similarity with lessons learned
Multimodal distributional semantics. Journal of Ar- from word embeddings. Transactions of the Associ-
tificial Intelligence Research (JAIR), 49(1-47). ation for Computational Linguistics, 3:211–225.
Susan Carey and Elsa Bartlett. 1978. Acquiring a sin- Thang Luong, Richard Socher, and Christopher Man-
gle new word. Papers and Reports on Child Lan- ning. 2013. Better word representations with re-
guage Development, 15:17–29. cursive neural networks for morphology. In Pro-
ceedings of the Seventeenth Conference on Computa-
Stephen Clark. 2015. Vector space models of lexical tional Natural Language Learning, pages 104–113.
meaning. The Handbook of Contemporary semantic
theory, pages 493–522. Matthew McGrath and Devin Frank. 2018. Proposi-
tions. In Edward N. Zalta, editor, The Stanford En-
Marc N Coutanche and Sharon L Thompson-Schill. cyclopedia of Philosophy, spring 2018 edition. Meta-
2014. Fast mapping rapidly integrates information physics Research Lab, Stanford University.
into existing memory networks. Journal of Experi-
mental Psychology: General, 143(6):2296. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey
Dean. 2013a. Efficient estimation of word represen-
Katrin Erk. 2012. Vector space models of word mean- tations in vector space. In Proceedings of the ICLR
ing and phrase meaning: A survey. Language and Workshop.
Linguistics Compass, 6(10):635–653.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor-
Annapaola Ginammi, Jelke Bloem, Rob Koopman, rado, and Jeff Dean. 2013b. Distributed representa-
Shenghui Wang, and Arianna Betti. in press. tions of words and phrases and their compositional-
Bolzano, Kant, and the Traditional Theory of Con- ity. In Advances in neural information processing
cepts: A Computational Investigation. In The Dy- systems, pages 3111–3119.
namics of Science: Computational Frontiers in His-
tory and Philosophy of Science. Pittsburgh Univer- Farhad Nooralahzadeh, Lilja Øvrelid, and Jan Tore
sity Press. Lønning. 2018. Evaluation of domain-specific word
embeddings using knowledge resources. In Proceed-
Anna Gladkova and Aleksandr Drozd. 2016. Intrinsic ings of the Eleventh International Conference on
evaluations of word embeddings: What can we do Language Resources and Evaluation (LREC-2018).
better? In Proceedings of the 1st Workshop on Eval-
uating Vector-Space Representations for NLP, pages Willard Van Orman Quine. 1934. A system of logistic.
36–42. Harvard University Press.
140
Willard Van Orman Quine. 1960. Word & Object. MIT
Press.
Magnus Sahlgren and Alessandro Lenci. 2016. The ef-
fects of data size and frequency range on distribu-
tional semantic models. In Proceedings of the 2016
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 975–980.
Tobias Schnabel, Igor Labutov, David Mimno, and
Thorsten Joachims. 2015. Evaluation methods for
unsupervised word embeddings. In Proceedings of
the 2015 Conference on Empirical Methods in Natu-
ral Language Processing, pages 298–307.
Yulia Tsvetkov, Manaal Faruqui, Wang Ling, Guil-
laume Lample, and Chris Dyer. 2015. Evaluation of
word vector representations by subspace alignment.
In Proceedings of the 2015 Conference on Empiri-
cal Methods in Natural Language Processing, pages
2049–2054.
Peter D Turney and Patrick Pantel. 2010. From fre-
quency to meaning: Vector space models of se-
mantics. Journal of artificial intelligence research,
37:141–188.
Bin Wang, Angela Wang, Fenxiao Chen, Yuncheng
Wang, and C.-C. Jay Kuo. 2019. Evaluating word
embedding models: methods and experimental re-
sults. APSIPA Transactions on Signal and Informa-
tion Processing, 8:e19.
141
Cross-Domain Training for Goal-Oriented Conversational Agents
142
Proceedings of Recent Advances in Natural Language Processing, pages 142–150,
Varna, Bulgaria, Sep 2–4 2019.
tomer support, providing certain information or In order to train a model for a more complex do-
general help with reservations etc., is partly due to main such as customer support, the improvement
the difficulty of obtaining large annotated datasets. has to be validated on multiple datasets from dif-
With each new domain and each new conversa- ferent domains. Also, because the cost of anno-
tion flow introduced by a new task, newly anno- tating data for such a domain is very high, trans-
tated data need to be fed into the system in order fer learning methods should be studied for possi-
to assimilate them and later provide the best an- ble improvements that increase the automation of
swer for different inputs. The efforts for selecting, domain-specific conversations with few data.
categorising and annotating the data are substan- The rest of the paper is organized as follows.
tial, no matter the previous experience or domain- Section 2 presents the related work for Goal Ori-
knowledge. This paper analyses the possibility to ented Chatbots. The model used in our exper-
alleviate the data annotation endeavor through the iments is detailed in Section 3 and the datasets
inter-domain transfer learning technique (Ilievski in Section 4. The results of our experiments are
et al., 2018). Alongside with unsupervised pre- described in detail and interpreted in Section 5.
training and others, transfer learning has proven Finally, future improvements and conclusions are
a significant contribution to deliver better results, presented in Section 6.
but it has only been tested with datasets from a
small number of domains. We experiment with 2 Related Work
larger datasets, wider scenarios and we offer a
richer understanding of the method, premises and We have already classified the existing solutions
results for overcoming the lack of data. used for building machine learning chatbots in two
In this paper, we provide a thorough study categories, based on the learning method: super-
on the impact of transfer learning in goal- vised learning and reinforcement learning. In this
oriented chatbots, starting from the work pre- section, we are providing a more in depth analysis
sented by (Ilievski et al., 2018). They proposed of these two alternatives.
the possibility to reuse the knowledge gained from Serban et al. (2016) propose a solution for
a source domain to boost the training and testing non-goal-driven systems, which uses an encoder-
performance of a machine learning chatbot model decoder model and word embeddings to generate
on a different target domain, as described in more the response of the agent starting from the utter-
detail in the following sections. They identify two ance of the user as input. The architecture is com-
cases: posed from two RNNs: one for the utterance level,
which treats the dialogue as a sequence of utter-
• domain overlap - the source and target do- ances, and one at the word level, which processes
mains are different, but share a fraction of ac- an utterance as a sequence of words. This model
tions, and is trained on movie scripts and the dialogues in-
clude the speech acts. A detail worth mention-
• domain extension - the source domain is ex-
ing here is that the pre-training is performed on a
tended by the target domain.
large related, but non-dialogue, corpus. The con-
In both cases, there are common actions that jus- sequence is that the model accomplishes slightly
tify the transfer learning between domains instead better results compared with an initialization with
of independently training models for each of them. fixed word embeddings.
This approach has two effects: (1) the success rate Another supervised learning model for chatbots
of the model obtained with transfer learning is sig- is presented by (Wen et al., 2017). The architec-
nificantly higher than that of the model trained ture is significantly more complex, and is divided
without any prior knowledge, and (2) transfer in several modules. The utterances received from
learning can be an alternative or complementary the user are converted into two representations:
to warm starting, which also requires labeled data. a probability distribution over the slot-value pairs
The results presented by the aforementioned au- called the belief state, and an intent representation
thors represent a significant improvement for GO generated by an intent network. A database oper-
Chatbots, but they are obtained with only three ator selects the most probable values in the belief
domains, with relatively small datasets: Movie state and makes a query to the database. A policy
Booking, Restaurant Booking, and Tourist Info. network takes as input the intent representation,
143
database result, and belief state and returns a rep- essary to have the same state distribution in both
resentation of the next system action. Finally, a domains, therefore the bots trained on the source
generation network uses the action representation domain must be aware of the actions in the target
to generate a template sequence, which is filled domain. They obtain an improvement of 65% in
with actual values from the database. This system success rate in the case of domain extension and
is very similar to the one used in the current pa- 20% for domain overlap. This represents a note-
per, but the former is trained in a supervised fash- worthy result and one of the reasons we chose to
ion and, therefore, it is possible to fail at finding study this mechanism in more detail. Another rea-
a good policy due to the shortcomings in dialogue son is the faster learning resulted from the combi-
exploration. Firstly, the policy is learned by a net- nation of transfer learning with warm start.
work instead of using RL (our case) and, secondly, In a more recent paper, Wolf et al. (2019)
the − greedy policy used in our experiments en- present the improvement brought by transfer
sures the exploration of unknown states, instead learning in generative tasks such as open-domain
of relying entirely on seen training data and rigid dialog generation. A Transformer model (Vaswani
choices. et al., 2017) is pre-trained on a large unlabeled
A possible solution for the disadvantage of us- dataset, followed by a fine-tuning step in which
ing supervised training in the model presented two loss functions are optimized: (1) a next-
above is proposed by Su et al. (2016). The ar- utterance classification loss, and (2) a language
chitecture is similar, but there is a difference in modeling loss. As a result, the model outperforms
the policy network training: it receives the current the existing systems by a significant margin ob-
state and predicts the next system action in a su- taining 51% absolute improvement in perplexity
pervised fashion in the first phase, followed by a on the validation dataset.
reinforcement learning phase. The purpose of the Given that many works successfully engage un-
second phase is to improve the generalization ca- supervised learning in various manners, there still
pacity of the policy by a better exploration of the remains the question: how does unsupervised pre-
action space using reinforcement learning. training work? An answer is formulated by (Erhan
A step forward in solving complex tasks is done et al., 2010) and a possible explanation is that the
by Peng et al. (2017). They introduce a hier- pre-training guides the learning towards basins of
archical deep reinforcement learning architecture attraction of minima that support better general-
for solving composite tasks for travel planning, ization from the training dataset. Therefore, it acts
that are a collection of subtasks such as: book like a regularizer for the supervised fine-tunning
air ticket, reserve hotel room, buy train ticket, phase, when the parameters are restricted to a rel-
etc. This type of tasks are a challenge for RL atively small space. This assumption is reinforced
approaches because of the reward sparsity, the by the results that show an effectiveness’ upgrade
slot constraints between different subtasks and of pre-training as the number of units per layer in-
the agent’s tendency to switch between different creases, a better generalization performance, but
subtasks frequently when conversing with users, worse training errors, and worse performance than
which leads to poor user experience. The dialogue random initialization for small networks, all char-
manager consists of (1) a top-level dialogue pol- acteristics of regularization. They also show a
icy that selects subtasks, (2) a low-level dialogue growth in the probability of finding a local min-
policy that selects the actions in a given subtask, ima by increasing the depth of a network with ran-
and (3) a global state tracker that supervises the dom initialization, compared to an unsupervised
cross-subtask constraints. pre-training.
Ilievski et al. (2018) use the transfer learning The most important advantage of pre-training is
mechanism for chatbots employing neural models the possibility of using unlabeled data, which is
to reduce the amount of training data and speed up really helpful given the high costs of data annota-
the learning process for new domains. This can tion. Therefore, the effect of pre-training with very
be accomplished with the transfer of the param- large datasets observed in the experiments is the
eters learned in a source domain to a target do- most surprising result of the paper (Erhan et al.,
main, which has some common actions with the 2010): the early examples determine the basin of
former. In order to apply the transfer, it is nec- attraction for the remainder of the training and the
144
supervised fine-tuning cannot escape from it. The user has an internal state su composed of a goal G
hypothesis is that those examples induce changes and an agenda A. The goal consists of constraints
in the magnitude of the weights, which decreases C and requests R. At each step t, the user simulator
the number of regions accessible to the stochas- generates the user action au,t based on the current
tic gradient descent procedure. This is why, in a state su,t and the last agent action aa,t and updates
0
large-scale setting, the influence of unsupervised the current state su,t .
pre-training is still present, in contrast to the clas- Natural Language Generation (NLG) is the
sical regularizers, when the effect disappears with module that generates natural language text for the
more data. user dialogue actions. For better results, a hy-
Nevertheless, fine-tuning large pre-trained brid approach is used, including a model-based
models is parameter inefficient, because each NLG and a template-based NLG. The model-
task requires an entirely new model. A compact based NLG is an LSTM decoder, which takes a
and extensible model is needed in order to solve dialogue action as input and generates a sentence
this shortcoming. For this purpose, (Houlsby with slot placeholders. If the sentence can be
et al., 2019) introduce adapter-based tuning, found in the predefined templates, the template-
which achieves a mean GLUE score of 80.0 based NLG is applied for filling the slots, other-
on several text classification tasks, compared wise, the utterance generated by the model-based
to 80.4 achieved by full fine-tuning, using 1.3 NLG is used.
task-specific parameters in total, compared to 9. Natural Language Understanding (NLU) is
This method also facilitates continual learning the opposite to the NLG module: it takes as in-
(training on a sequence of tasks) and multi-task put an utterance and determines the user’s in-
learning (training on simultaneous tasks). tent and the set of slots associated with it (e.g.
{movie name: ”deadpool”, date: ”saturday”,
3 Model number of people: ”5”}), in order to form a se-
The system used in this paper is a semantic frames mantic frame. It is implemented with an LSTM
system (Li et al., 2017). It represents the dialogue and its objective is to maximize the conditional
as a set of slot-value pairs and at each step t, given probability of the slots and the intent, given the
the user utterance ut , the agent takes an action at , utterance.
which can be either the final result or a request for The Dialogue Management (DM) includes
a value of an empty slot. The architecture consists two submodules: Dialogue State Tracker and
of two parts: a User Simulator module and a Di- Policy Learning module. The goal of the Dia-
alogue Manager module. logue State Tracker is to build a representation of
The purpose of the User Simulator is to inter- the current state for policy learning, using the se-
act with the Dialogue Manager in order to train a mantic frame received from the NLU component.
policy for an agent. First, a user goal is chosen It keeps the history of the user utterances, system
randomly from the goals’ pool and is unknown for actions and the query results from the Knowledge
the agent, but it tries to help the user to accom- Base.
plish it during the dialogue. The goal consists of Policy learning module generates the next ac-
two types of slots: tion of the system at according to the policy π =
P (a|s), given the current state st , in order to ac-
• inform slots - represent the constraints im-
complish the user goal in the smallest number of
posed by the user, hence their values are
steps. The state st includes the latest user action,
known (e.g. {movie name: ”deadpool”, city:
the latest agent action, turn information, history
”Madison Heights”, date: ”saturday”, num-
dialogue turns and the available database results.
ber of people: ”5”}).
A DQN (Mnih et al., 2015) is used to approximate
• request slots - represent the values that the the state-action function Q(s, a|Θ) and contains
agent should provide, hence they enclose the experience replay mechanism.
unknown values to the user (e.g. {price,
start time, critic rating}). 4 Dataset
Then, the user utterance ut is generated follow- In order to study the impact of transfer learn-
ing the Agenda-Based model (Li et al., 2016): the ing in multiple domains, we choose MultiWOZ
145
(a) Hospital with pre-training on Attraction domain.
2.0 (Budzianowski et al., 2018), a large-scale the task respecting the constraints imposed by the
multi-domain corpus of natural human-human user. In the second case, the user sends a list of
conversations, collected through the Wizard-of- request slots besides the list of inform slots, and
Oz framework (Kelley, 1984). It contains about the agent should accomplish the task and answer
10000 samples from seven domains, with an av- to the user’s questions.
erage of turns per dialogue between 8.9 and 15,
depending on the domain. From this dataset, we Source Target Source Target Common
Domain Domain Slots Slots Slots
select the following five domains: hotel, attrac- 1 movie restaurant 6 9 3
tion, train, taxi, hospital, and also keep movie 2 restaurant tourist 9 9 6
and tourist domains used by Ilievski et al. (2018). 3 hotel attraction 13 9 5
4 train taxi 9 6 4
These domains are grouped in source-target pairs
5 movie hotel 17 13 5
according to their common slots, resulting five 6 tourist hotel 9 13 6
new opportunities for transfer learning. The total 7 attraction hospital 9 4 3
number of slots for source and target domains, re-
spectively, as well as the amount of common slots Table 1: Number of slots per domain
is presented in Table1. We call extra source/target
slots the difference between the total domain slots This is a noteworthy detail, because it can influ-
and the common ones. ence the success rate through the experience ac-
cumulated in the warm start phase. In this phase,
The goals can be divided into two categories de- a fixed-size buffer is filled with experiences from
pending on whether they contain request slots or positive-outcome conversations. Thus, the learn-
not. In the first case, the user sends a list of inform ing process gains a boost when having to self-
slots to the agent and the agent should accomplish calibrate based on an experience which will lead
146
(a) Attraction with pre-training on Hotel domain.
to goal-achievement. We have noticed that the small compared with the number of slots from the
best results are obtained with warm start on no- hotel domain, and the agent trained on the source
request goals, following that the agent will man- domain is not able to learn. Consequently, we in-
age to achieve request goals during training. As creased nmax turns to 40 turns.
we increased the percentage of request goals in
the warm start buffer, the overall success rate 5.1 Different Domains
decreases and the learning curve becomes less The first set of experiments aims to analyze the
smooth. convergence for the agent with and without trans-
The total number of goals per domain is dis- fer learning on new domains. We group the exper-
tributed as follows: iments in three categories, according to the num-
ber of extra slots in the target domain, as follows:
• 3000 training and 400 testing user goals in
1. small number of extra slots (less than 2 slots);
hotel, train and attraction;
2. medium number of extra slots (between 3 and
• 2000 training and 200 testing user goals in 6 slots); and 3. big number of extra slots (greater
taxi datasets; than 6 slots). We are interested in the improvement
transfer learning brings to the success rate and the
• 80 training and 15 testing user goals in hos- convergence pace.
pital dataset. Figure 1 presents the learning curve for the tar-
get domains with a small number of extra slots.
5 Experiments
For hospital domain with pre-training on attrac-
The experiments are executed on overlapping do- tion, both agents converge to a success rate greater
mains, with the setup from (Ilievski et al., 2018) than 95%, but the agent with transfer learning con-
and running each experiment 10 times, with verges in a few epochs (<5), while the scratch
nepochs = 100 epochs. The second set of ex- agent needs 40 epochs to reach similar accu-
periments mimic testing on extension domain by racy values. In the case of taxi domain with
restricting the slots in the source domain to the pre-training on train domain, the improvement of
common ones. The warm start technique with transfer learning is 15% for train dataset and 19%
experience replay buffer is used for both trans- for test dataset. In absolute terms, the success rate
fer learning and scratch agent (the same version, increases from 80% to 91% for train dataset and
but without transfer learning), and the experience from 76% to 91% on test dataset.
buffer is flushed when the agent reaches, for the For the attraction domain with pre-training on
first time, a success rate of 0.3. We also keep hotel, presented in Figure 2, the model obtained
the maximal number of allowed turns per dialogue with transfer learning has a success rate 7% higher
nmax turns = 20 in most experiments, except in than that of the scratch model on train dataset.
the case of attraction domain with pre-training on This denotes an increase from 68% to 73% in ab-
hotel. In this situation, the number of turns is too solute terms. For test dataset, transfer learning im-
147
Source Target Scratch TL Scratch TL
proves the success rate from 68% to 76% or with Slots Slots Score Score Epochs Epochs
12% in relative terms. 1 6 6 0.81 0.88 100 40
2 6 7 0.75 0.85 90 40
The last category registers the lowest overall 3 6 8 0.63 0.79 70 100
4 6 9 0.55 0.72 100 100
success rate for both agents and we consider that 5 6 10 0.53 0.59 100 100
these results stem from the large number of extra 6 6 11 0.09 0.52 100 90
7 6 12 0.01 0.44 90 100
slots in the target domain. The learning curves are 8 6 13 0.0 0.39 10 100
illustrated in Figure 3 and the success rate is sim- 9 9 6 0.77 0.83 100 50
10 9 7 0.71 0.82 90 100
ilar for train and test datasets. Hotel domain with 11 9 8 0.65 0.76 100 100
pre-training on movie has a success rate of 27% 12 9 9 0.56 0.70 100 100
13 9 10 0.54 0.59 100 100
with transfer learning and 8.5% without transfer 14 9 11 0.11 0.52 100 100
15 9 12 0.04 0.42 90 100
learning, with an improvement of 217%. While 16 9 13 0.04 0.36 100 60
the same domain with tourist as source domain
of transfer learning, registers a relative boost of Table 3: Target Slots number influence
737%, from 4.3% to 36%.
5.2 Same Domains, Different Number of As expected, the number of extra slots in tar-
Slots get domain has a bigger influence over the final
results. Therefore, the relative average decrease of
The second set of experiments targets the evolu-
the success rate for the agent with transfer learning
tion of the success rate according to the number
is 9.5% for each extra slot, while the source do-
of extra slots, both in the source and target do-
main contains only common slots. Another three
main. The selected experiment evaluates the hotel
slots added to the source dataset generates an aver-
domain with pre-training on tourist dataset, given
age decrease of 3.6%, relative to the previous test.
they each have large number of slots with few
In comparison with the scratch agent, the improve-
common ones (see Table1). We keep the setup for
ment increases from 79% in the first case, to 274%
the other parameters and only change the number
in the latter.
of extra slots in one domain, while the other re-
mains constant. 6 Conclusions
Source Target Scratch TL Scratch TL
Slots Slots Score Score Epochs Epochs In this paper, we study the factors that influence
1 6 6 0.81 0.88 100 40
2 7 6 0.81 0.86 100 70
the success of transfer learning approach in Re-
3 8 6 0.81 0.84 100 70 inforcement Learning-based Goal-Oriented Chat-
4 9 6 0.77 0.83 100 50
5 6 9 0.55 0.72 100 100
bots and demonstrate the results on five new cases
6 7 9 0.57 0.63 100 100 of overlapping domains. We found that a big num-
7 8 9 0.54 0.70 80 100
8 9 9 0.56 0.70 100 100 ber of different slots between the source domain
and the target domain leads to a smaller success
Table 2: Source Slots number influence rate. Even so, the transfer learning mechanism
brings a betterment of over 79% over the agent
The final success rate on the test dataset for con- trained with no prior knowledge.
stant slots in target domain is summarized in Ta- The outcomes encourage the use of transfer
ble 2. When the target contains only the common learning with warm start on various cases of over-
slots, we observe a decrease of the success rate lapping and extending source and target domains.
with less than 2% with each extra slot added in the However, the optimal selection in terms of hyper-
source domain, for the model trained with trans- parameters of the system, such as the number of
fer learning. However, it is still by 6.6% greater epochs or the number of maximum turns in a con-
than the success rate of the agent trained with any versation, need to be determined for each particu-
other prior knowledge. An interesting fact is that lar scenario. They are, after all, directly influenced
the same test with the common slots plus three by the amount of data and its characteristics: num-
extra slots in target domain has the effect of di- ber of slots, types and distribution of goals, and the
minishing the success rate by an average of 15% degree of overlapping between source and target
compared with the previous situation. At the same slots.
time, the improvement over the scratch agent is Further work involves developing wider exper-
equal to 24%. iment scenarios for hierarchical deep reinforce-
148
(a) Hotel with pre-training on Movie domain.
ment learning system and introducing the transfer learning? Journal of Machine Learning Research
learning approach into this architecture when the 11:625–660.
sub-tasks share slots. Moreover, we can imagine Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski,
other transfer learning setups such as sharing sub- Bruna Morrone, Quentin De Laroussilhe, Andrea
tasks as the learnt common part, instead of slots, Gesmundo, Mona Attariyan, and Sylvain Gelly.
from one composite task to another. All these at- 2019. Parameter-efficient transfer learning for NLP.
In Kamalika Chaudhuri and Ruslan Salakhutdi-
tempts have the objective of gaining more context nov, editors, Proceedings of the 36th International
information and better performance with less an- Conference on Machine Learning. PMLR, Long
notated data, which is onerous to obtain. Beach, California, USA, volume 97 of Proceedings
of Machine Learning Research, pages 2790–2799.
http://proceedings.mlr.press/v97/houlsby19a.html.
149
Gao. 2017. End-to-end task-completion neural di- Thomas Wolf, Victor Sanh, Julien Chaumond, and
alogue systems. In Proceedings of the Eighth In- Clement Delangue. 2019. Transfertransfo: A trans-
ternational Joint Conference on Natural Language fer learning approach for neural network based con-
Processing, IJCNLP 2017, Taipei, Taiwan, Novem- versational agents. CoRR abs/1901.08149.
ber 27 - December 1, 2017 - Volume 1: Long Papers.
pages 733–743.
Xiujun Li, Zachary C. Lipton, Bhuwan Dhingra, Li-
hong Li, Jianfeng Gao, and Yun-Nung Chen. 2016.
A user simulator for task-completion dialogues.
CoRR abs/1612.05688.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver,
Andrei A. Rusu A., Joel Veness, and Marc G. Belle-
mare et al. 2015. Human-level control through deep
reinforcement learning. Nature 518(7540):529–
533.
Baolin Peng, Xiujun Li, Lihong Li, Jianfeng Gao,
Asli Celikyilmaz, Sungjin Lee, and Kam-Fai Wong.
2017. Composite task-completion dialogue policy
learning via hierarchical deep reinforcement learn-
ing. In Proceedings of the 2017 Conference on Em-
pirical Methods in Natural Language Processing,
EMNLP 2017, Copenhagen, Denmark, September
9-11, 2017. pages 2231–2240.
Iulian V. Serban, Alessandro Sordoni, Yoshua Bengio,
Aaron Courville, and Joelle Pineau. 2016. Build-
ing end-to-end dialogue systems using generative hi-
erarchical neural network models. In Proceedings
of the Thirtieth AAAI Conference on Artificial In-
telligence, February 12-17, 2016, Phoenix, Arizona,
USA.. pages 3776–3784.
Pei-Hao Su, Milica Gasic, Nikola Mrksic, Lina Rojas-
Barahona, Stefan Ultes, David Vandyke, Tsung-
Hsien Wen, and Steve Young. 2016. Continu-
ously learning neural dialogue management. CoRR
abs/1606.02689.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural net-
works. In Z. Ghahramani, M. Welling, C. Cortes,
N. D. Lawrence, and K. Q. Weinberger, editors,
Advances in Neural Information Processing Sys-
tems 27, Curran Associates, Inc., pages 3104–
3112. http://papers.nips.cc/paper/5346-sequence-
to-sequence-learning-with-neural-networks.pdf.
Ashish Vaswani, Noam Shazeer, Niki Parmar, and
Jakob Uszkoreit et al. 2017. Attention is all you
need. In Advances in Neural Information Process-
ing Systems 30: Annual Conference on Neural In-
formation Processing Systems 2017, 4-9 December
2017, Long Beach, CA, USA. pages 6000–6010.
Tsung-Hsien Wen, David Vandyke, Nikola Mrksic,
Milica Gasic, Lina M. Rojas-Barahona, Pei-Hao Su,
Stefan Ultes, and Steve Young. 2017. A network-
based end-to-end trainable task-oriented dialogue
system. In Proceedings of the 15th Conference of
the European Chapter of the Association for Com-
putational Linguistics, EACL 2017, Valencia, Spain,
April 3-7, 2017, Volume 1: Long Papers. pages 438–
449.
150
Learning Sentence Embeddings for Coherence Modelling and Beyond
Abstract
We present a novel and effective tech-
nique for performing text coherence tasks
while facilitating deeper insights into the
data. Despite obtaining ever-increasing
task performance, modern deep-learning
approaches to NLP tasks often only pro-
vide users with the final network deci- How coherent is it? Suggest a coherent sentence order
the data. In this work, we show that a coherence. If far, low coherence. appear along the x-axis.
151
Proceedings of Recent Advances in Natural Language Processing, pages 151–160,
Varna, Bulgaria, Sep 2–4 2019.
question of whether we can achieve good perfor- of a sentence in a document given only the raw
mance on NLP tasks while simultaneously provid- text. By training a neural network on this task,
ing users with easily obtainable insights into the it is forced to learn how the location of a sen-
data. This is precisely what the work in this paper tence in a structured text is related to its syntax
aims to do in the context of coherence analysis, and semantics. As a neural model, we use a bi-
by providing a tool with which users can quickly directional recurrent neural network, and train it
and visually gain insight into structural informa- to take sentences and predict a discrete distribu-
tion about a text. To accomplish this, we rely on tion over possible locations in the source text. We
the surprising importance of sentence location in demonstrate the effectiveness of predicted position
many areas of natural language processing. If a distributions as an accurate way to assess docu-
sentence does not appear to belong where it is lo- ment coherence by performing order discrimina-
cated, it decreases the coherence and readability tion and sentence reordering of scientific abstracts.
of the text (Lapata and Barzilay, 2005). If a sen- We also demonstrate a few types of insights that
tence is located at the beginning of a document these embeddings make available to users that the
or news article, it is very likely to be a part of a predicted location of a sentence in a news article
high quality extractive summary (See et al., 2017). can be used to formulate an effective heuristic for
The location of a sentence in a scientific abstract is extractive document summarization – outperform-
also an informative indicator of its rhetorical pur- ing existing heuristic methods.
pose (Teufel et al., 1999). It thus follows that the The primary contributions of this work are thus:
knowledge of where a sentence should be located 1. We propose a novel self-supervised approach
in a text is valuable. to learn sentence embeddings which works
Tasks requiring knowledge of sentence position by learning to map sentences to a distribution
– both relative to neighboring sentences and glob- over positions in a document (Section 2.2).
ally – appear in text coherence modelling, with
2. We describe how these sentence embeddings
two important tasks being order discrimination (is
can be applied to established coherence tasks
a sequence of sentences in the correct order?) and
using simple algorithms amenable to visual
sentence ordering (re-order a set of unordered sen-
approximation (Section 2.3).
tences). Traditional methods in this area make use
of manual feature engineering and established the- 3. We demonstrate that these embeddings are
ory behind coherence (Lapata and Barzilay, 2005; competitive at solving text coherence tasks
Barzilay and Lapata, 2008; Grosz et al., 1995). (Section 3) while quickly providing access to
Modern deep-learning based approaches to these further insights into texts (Section 4).
tasks tend to revolve around taking raw words and
directly predicting local (Li and Hovy, 2014; Chen 2 Predicted Position Distributions
et al., 2016) or global (Cui et al., 2017; Li and Ju- 2.1 Overview
rafsky, 2017) coherence scores or directly output
By training a machine learning model to predict
a coherent sentence ordering (Gong et al., 2016;
the location of a sentence in a body of text (condi-
Logeswaran et al., 2018; Cui et al., 2018). While
tioned upon features not trivially indicative of po-
new deep-learning based approaches in text coher-
sition), we obtain a sentence position model such
ence continue to achieve ever-increasing perfor-
that sentences predicted to be at a particular loca-
mance, their value in real-world applications is un-
tion possess properties typical of sentences found
dermined by the lack of actionable insights made
at that position. For example, if a sentence is pre-
available to users.
dicted to be at the beginning of a news article, it
In this paper, we introduce a self-supervised ap- should resemble an introductory sentence.
proach for learning sentence embeddings which In the remainder of this section we describe our
can be used effectively for text coherence tasks neural sentence position model and then discuss
(Section 3) while also facilitating deeper under- how it can be applied to text coherence tasks.
standing of the data (Section 4). Figure 1 provides
a taste of this, displaying the sentence embeddings 2.2 Neural Position Model
for the abstract of this paper. The self-supervision The purpose of the position model is to produce
task we employ is that of predicting the location sentence embeddings by predicting the position in
152
PPD sequence
lem, we aim to determine what quantile of the doc-
PPD(S1)
PPD(S2) ument a sentence resides in. Notationally, we will
=
PPD(S3) refer to the number of quantiles as Q. We can in-
PPD(S4)
terpret the class probabilities behind a prediction
Sentence position model as a discrete distribution over positions for a sen-
Softmax (sentence PPD) tence, providing us with a predicted position dis-
tribution (PPD). When Q = 2 for example, we are
concat
predicting whether a sentence is in the first or last
LSTM LSTM LSTM half of a document. When Q = 4, we are pre-
dicting which quarter of the document it is in. In
LSTM LSTM LSTM Figure 2 is a visualization of the neural architec-
ture which produces PPDs of Q = 10.
concat
153
this. As a simple example, if we include the true 2.3.1 Sentence Ordering
location of a sentence in a text as a feature when To induce a new ordering on a sequence of sen-
training the position model, then instead of learn- tences, S, we simply sort the sentence by their
ing the connection between sentence meaning and weighted average predicted quantile, Q̂(s ∈ S),
position, the mapping would trivially exploit the defined by:
known sentence position to perfectly predict the
sentence quantile position. This would not allow Q
X
us to observe where the sentence seems it should Q̂(s) = i × P P D(s)i , (1)
be located. i=1
Calculate PPDs coh = τ ((Q̂(s), for s = S1 , ..., S|S| ), (1, ..., |S|)).
(2)
3 Experiments
In this section, we evaluate our PPD-based ap-
proaches on two coherence tasks and demonstrate
Extract sentences with
that only minimal performance is given up by our
highest Q1 probability Calculate weighted
average predicted
approach to providing more insightful sentence
Summary sentence quantiles embeddings.
Sentences 1, 2, and 7
154
Order discrimination Reordering
Model Accident Earthquake Acc τ
Random 50 50 15.6 0
Entiry Grid 90.4 87.2 20.1 0.09
Window network - - 41.7 0.59
LSTM PtrNet 93.7 99.5 50.9 0.67
RNN Decoder - - 48.2 0.67
Varient-LSTM+PtrNet 94.4 99.7 51.6 0.72
ATTOrderNet 96.2 99.8 56.1 0.72
PPDs 94.4 99.3 54.9 0.72
Table 2: Results on the order discrimination and sentence reordering coherence tasks. Our approach
trades only a small decrease in performance for improved utility of the sentence embeddings over other
approaches, achieving close to or the same as the state-of-the-art.
ordering of the sentences and the original or- der discrimination dataset and for three trials for
dering is the most coherent (in the original or- the sentence reordering task. The final hyperpa-
der), for twenty such permutations. Since these rameters chosen are in Table 1.
datasets only contain training and testing parti- Baselines. We compare our results against
tions, we follow (Li and Hovy, 2014) and perform a random baseline, the traditional Entity
10-fold cross-validation for hyperparameter tun- Grid approach from (Barzilay and Lapata,
ing. Performance is measured with the accuracy 2008), Window network (Li and Hovy, 2014),
with which the permuted sentences are identified. LSTM+PtrNet (Gong et al., 2016), RNN Decoder
For example, the Entity Grid baseline in Table 2 and Varient-LSTM+PtrNet from (Logeswaran
gets 90.4% accuracy because given a shuffled re- et al., 2018), and the most recent state-of-the art
port and original report, it correctly classifies them ATTOrderNet (Cui et al., 2018).
90.4% of the time. Results. Results for both coherence tasks are
Sentence ordering setup. For sentence order- collected in Table 2. For the order discrimination
ing, we use past NeurIPS abstracts to compare task, we find that on both datasets, our PPD-based
with previous works. While our validation and test approach only slightly underperforms ATTOrder-
partitions are nearly identical to those from (Lo- Net (Cui et al., 2018), with performance similar to
geswaran et al., 2018), we use a publicly available the LSTM+PtrNet approaches (Gong et al., 2016;
dataset2 which is missing the years 2005, 2006, Logeswaran et al., 2018). On the more difficult
and 2007 from the training set ((Logeswaran et al., sentence reordering task, our approach exhibits
2018) collected data from 2005 - 2013). Abstracts performance closer to the state-of-the-art, achiev-
from 2014 are used for validation, and 2015 is ing the same ranking correlation and only slightly
used for testing. To measure performance, we re- lower positional accuracy. Given that the pub-
port both reordered sentence position accuracy as licly available training set for the reordering task
well as Kendall’s rank correlation coefficient. For is slightly smaller than that used in previous work,
example, the Random baseline correctly predicts it is possible that more data would allow our ap-
the index of sentences 15.6% of the time, but there proach to achieve even better performance. In the
is no correlation between the predicted ordering next section we will discuss the real-world value
and true ordering, so τ = 0. offered by our approach that is largely missing
Training and tuning. Hyperparameter tun- from existing approaches.
ing for both tasks is done with a random search,
choosing the hyperparameter set with the best val- 4 Actionable Insights
idation score averaged across the 10 folds for or-
A primary benefit of applying PPDs to coherence-
2
https://www.kaggle.com/benhamner/ related tasks is the ability to gain deeper insights
nips-papers into the data. In this section, we will demon-
155
Figure 4: The PPDs for a CNN article. (full text available at http://web.
archive.org/web/20150801040019id_/http://www.cnn.com/2015/03/13/us/
tulane-bacteria-exposure/). The dashed line shows the weighted average predicted sentence
positions.
strate the following in particular: (1) how PPDs being able to understand where subsections may
can quickly be used to understand how the coher- occur in a document, a writer can make informed
ence of a text may be improved, (2) how the ex- decisions on where to split a long text into more
istence of multiple coherence subsections may be coherent chunks or paragraphs. Knowing where
identified, and (3) how PPDs can allow users to lo- approximate borders between ideas in a document
cate specific types of information without reading exist may also help readers skim the document to
a single word, a specific case of which is extrac- find desired information more quickly, as further
tive summarization. For demonstrations, we will discussed in the next subsection.
use the news article presented in Figure 4.
4.3 Locating Information and
4.1 Improving Coherence Summarization
For a writer to improve their work, understand- When reading a new article, readers well-versed
ing the incoherence present is important. Observ- in the subject of the article may want to skip high-
ing the PPD sequence for the article in Figure 4 level introductory comments and jump straight to
makes it easy to spot areas of potential incoher- the details. For those unfamiliar with the content
ence: they occur where consecutive PPDs are sig- or triaging many articles, this introductory infor-
nificantly different (from sentences 1 to 2, 6 to 7, mation is important to determine the subject mat-
and 10 to 11). In this case, the writer may deter- ter. Using PPDs, locating these types of infor-
mine that sentence 2 is perhaps not as introduc- mation quickly should be easy for readers, even
tory as it should be. The predicted incoherence when the document has multiple potential subsec-
between sentences 10 and 11 is more interesting, tions. In Figure 4, sentences 1 and 11 likely con-
and as we will see next, the writer may realize that tain introductory information (since the probabil-
this incoherence may be okay to retain. ity of occurring in the first quantiles is highest), the
most conclusory-type information is in sentence
4.2 Identifying Subsections 10, and lower-level details are likely spread among
In Figure 4, we see rough progressions of the remaining sentences.
introductory-type sentences to conclusory-type Locating sentences with the high-level details
sentences between sentences 1 and 10 and sen- of a document is reminiscent of the task of extrac-
tences 11 and 15. This may indicate that the ar- tive summarization, where significant research has
ticle is actually composed of two coherent subsec- been performed (Nenkova et al., 2011; Nenkova
tions, which means that the incoherence between and McKeown, 2012). It is thus natural to ask
sentences 10 and 11 is expected and natural. By how well a simple PPD-based approach performs
156
Model (lead baseline source) ROUGE-1 ROUGE-2 ROUGE-L
Lead-3 (Nallapati et al., 2017) 39.2 15.7 35.5
Lead-3 (See et al., 2017) 40.3 17.7 36.6
Lead-3 (Ours) 35.8 15.9 33.5
SummaRuNNer (Nallapati et al., 2017) ((Nallapati et al., 2017)) 39.6 16.2 35.3
Pointer-generator (See et al., 2017) ((See et al., 2017)) 39.5 17.3 36.4
RL (Paulus et al., 2017) ((Nallapati et al., 2017)) 41.2 15.8 39.1
TextRank (Mihalcea and Tarau, 2004) (ours) 26.2 11.1 24.3
Luhn (Luhn, 1958) (ours) 26.4 11.2 24.5
SumBasic (Nenkova and Vanderwende, 2005) (ours) 27.8 10.4 26.0
LexRank (Erkan and Radev, 2004) (ours) 28.4 11.6 26.3
PPDs (ours) 30.1 12.6 28.2
Table 3: ROUGE scores on the CNN/DailyMail summarization task. Our PPD-based heuristic outper-
forms the suite of established heuristic summarizers. However, the higher performance of the deep-
learning models demonstrates that training explicitly for summarization is beneficial.
at summarization. To answer this question, the uated using full-length ROUGE scores on the non-
summarization algorithm we will use is: select the anonymized CNN/DailyMail dataset.
n sentences with the highest P P D(s ∈ S)0 value, Table 3 contains the the comparison between
where S is the article being extractively summa- our PPD-based summarizer and several estab-
rized down to n sentences. For the article in Fig- lished heuristic summarizers. We observe that
ure 4, sentences 1, 11, and 3 would be chosen since our model has ROUGE scores superior to the
they have the highest first-quantile probabilities. other heuristic approaches by a margin of ap-
This heuristic is conceptually similar to the Lead proximately 2 points for ROUGE-1 and -L and 1
heuristic, where sentences that actually occur at point for ROUGE-2. In contrast, the deep-learning
the start of the document are chosen to be in the approaches trained explicitly for summarization
summary. Despite its simplicity, the Lead heuris- achieve even higher scores, suggesting that there is
tic often achieves near state-of-the-art results (See more to a good summary than the sentences sim-
et al., 2017). ply being introductory-like.
We experiment on the non-anonymized
CNN/DailyMail dataset (Hermann et al., 2015) 5 Related Work
and evaluate with full-length ROUGE-1, -2, and Extensive research has been done on text coher-
-L F1 scores (Lin and Hovy, 2003). For the ence, motivated by downstream utility of coher-
neural position model, we choose four promising ence models. In addition to the applications we
sets of hyperparameters identified during the demonstrate in Section 4, established applications
hyperparameter search for the sentence ordering include determining the readability of a text (co-
task in Section 3 and train each sentence position herent texts are easier to read) (Barzilay and La-
model on 10K of the 277K training articles (which pata, 2008), refinement of multi-document sum-
provides our sentence position model with over maries (Barzilay and Elhadad, 2002), and essay
270K sentences to train on). Test results are scoring (Farag et al., 2018).
reported for the model with the highest validation Traditional methods to coherence modelling
score. The final hyperparameters chosen for this utilize established theory and handcrafted linguis-
sentence location model are: Q = 10, epochs = 10, tic features (Grosz et al., 1995; Lapata, 2003). The
layer dropouts = (0.4, 0.2), layer widths = (512, Entity Grid model (Lapata and Barzilay, 2005;
64). Barzilay and Lapata, 2008) is an influential tradi-
We compare our PPD-based approach to other tional approach which works by first constructing
heuristic approaches3 . For completeness, we a sentence × discourse entities (noun phrases) oc-
also include results of deep-learning based ap- currence matrix, keeping track of the syntactic role
proaches and their associated Lead baselines eval- of each entity in each sentence. Sentence tran-
3
Implementations provided by Sumy library, available at sition probabilities are then calculated using this
https://pypi.python.org/pypi/sumy. representation and used as a feature vector as in-
157
put to a SVM classifier trained to rank sentences effort to improve this situation, we present a self-
on coherence. supervised approach to learning sentence embed-
Newer methods utilizing neural networks and dings, which we call PPDs, that rely on the con-
deep learning can be grouped together by whether nection between the meaning of a sentence and its
they indirectly or directly produce an ordering location in a text. We implement the new sentence
given an unordered set of sentences. embedding technique with a recurrent neural net-
Indirect ordering. Approaches in the indi- work trained to map a sentence to a discrete distri-
rect case include Window network (Li and Hovy, bution indicating where in the text the sentence is
2014), Pairwise Ranking Model (Chen et al., likely located. These PPDs have the useful prop-
2016), the deep coherence model from (Cui et al., erty that a high probability in a given quantile indi-
2017), and the discriminative model from (Li and cates that the sentence is typical of sentences that
Jurafsky, 2017). These approaches are trained to would occur at the corresponding location in the
take a set of sentences (anywhere from two (Chen text.
et al., 2016) or three (Li and Hovy, 2014) to the We demonstrate how these PPDs can be applied
whole text (Cui et al., 2017; Li and Jurafsky, to coherence tasks with algorithms simple enough
2017)) and predict whether the component sen- such that they can be visually performed by users
tences are already in a coherent order. A final or- while achieving near state-of-the-art, outperform-
dering of sentences is constructed by maximizing ing more complex and specialized systems. We
coherence of sentence subsequences. also demonstrate how PPDs can be used to ob-
Direct ordering. Approaches in the direct case tain various insights into data, including how to
include (Gong et al., 2016; Logeswaran et al., go about improving the writing, how to identify
2018; Cui et al., 2018). These model are trained potential subsections, and how to locate specific
to take a set of sentences, encode them using some types of information, such as introductory or sum-
technique, and with a recurrent neural network mary information. As a proof-of-concept, we ad-
decoder, output the order in which the sentences ditionally show that despite PPDs not being de-
would coherently occur. signed for the task, they can be used to create a
Models in these two groups all use similar high- heuristic summarizer which outperforms compa-
level architectures: a recurrent or convolutional rable heuristic summarizers.
sentence encoder, an optional paragraph encoder,
In future work, it would be valuable to evaluate
and then either predicting coherence from that en-
our approach on texts from a wider array of do-
coding or iteratively reconstructing the ordering
mains and with different sources of incoherence.
of the sentences. The PPD-based approaches de-
In particular, examining raw texts identified by hu-
scribed in Section 2 take a novel route of directly
mans as lacking coherence could be performed,
predicting location information of each sentence.
to determine how well our model correlates with
Our approaches are thus similar to the direct ap-
human judgment. Exploring how the algorithms
proaches in that position information is directly
utilizing PPDs may be refined for improved per-
obtained (here, in the PPDs), however the posi-
formance on the wide variety of coherence-related
tion information produced by our model is much
tasks may also prove fruitful. We are also in-
more rich than simply the index of the sentence in
terested in examining how PPDs may assist with
the new ordering. With the set of indirect order-
other NLP tasks such as text generation or author
ing approaches, our model approach to coherence
identification.
modelling shares the property that induction of an
ordering upon the sentences is only done after ex-
amining all of the sentence embeddings and ex- Acknowledgments
plicitly arranging them in the most coherent fash-
ion.
We acknowledge the support of the Natural
6 Conclusions Sciences and Engineering Research Council of
Canada (NSERC) through the Discovery Grants
The ability to facilitate deeper understanding of Program. NSERC invests annually over $1 billion
texts is an important, but recently ignored, prop- in people, discovery and innovation.
erty for coherence modelling approaches. In an
158
References Mirella Lapata. 2003. Probabilistic text structuring:
Experiments with sentence ordering. In Proceed-
Regina Barzilay and Noemie Elhadad. 2002. Infer- ings of the 41st Annual Meeting on Association
ring strategies for sentence ordering in multidocu- for Computational Linguistics-Volume 1. Associa-
ment news summarization. Journal of Artificial In- tion for Computational Linguistics, pages 545–552.
telligence Research .
Mirella Lapata and Regina Barzilay. 2005. Automatic
Regina Barzilay and Mirella Lapata. 2008. Modeling evaluation of text coherence: Models and represen-
local coherence: An entity-based approach. Compu- tations. In IJCAI. volume 5, pages 1085–1090.
tational Linguistics 34(1):1–34.
Jiwei Li and Eduard Hovy. 2014. A model of co-
Xinchi Chen, Xipeng Qiu, and Xuanjing Huang.
herence based on distributed sentence representa-
2016. Neural sentence ordering. arXiv preprint
tion. In Proceedings of the 2014 Conference on
arXiv:1607.06952 .
Empirical Methods in Natural Language Processing
François Chollet et al. 2015. Keras. https:// (EMNLP). pages 2039–2048.
github.com/keras-team/keras.
Jiwei Li and Dan Jurafsky. 2017. Neural net models
Baiyun Cui, Yingming Li, Ming Chen, and Zhongfei of open-domain discourse coherence. In Proceed-
Zhang. 2018. Deep attentive sentence order- ings of the 2017 Conference on Empirical Methods
ing network. In Proceedings of the 2018 in Natural Language Processing. pages 198–209.
Conference on Empirical Methods in Nat-
ural Language Processing. Association for Chin-Yew Lin and Eduard Hovy. 2003. Auto-
Computational Linguistics, pages 4340–4349. matic evaluation of summaries using n-gram co-
http://aclweb.org/anthology/D18-1465. occurrence statistics. In Proceedings of the 2003
Conference of the North American Chapter of the
Baiyun Cui, Yingming Li, Yaqing Zhang, and Zhongfei Association for Computational Linguistics on Hu-
Zhang. 2017. Text coherence analysis based on deep man Language Technology-Volume 1. Association
neural network. In Proceedings of the 2017 ACM for Computational Linguistics, pages 71–78.
on Conference on Information and Knowledge Man-
agement. ACM, pages 2027–2030. Lajanugen Logeswaran, Honglak Lee, and Dragomir
Radev. 2018. Sentence ordering and coherence
Günes Erkan and Dragomir R Radev. 2004. Lexrank: modeling using recurrent neural networks. In
Graph-based lexical centrality as salience in text Thirty-Second AAAI Conference on Artificial Intel-
summarization. Journal of Artificial Intelligence ligence.
Research 22:457–479.
Hans Peter Luhn. 1958. The automatic creation of lit-
Youmna Farag, Helen Yannakoudakis, and Ted erature abstracts. IBM Journal of research and de-
Briscoe. 2018. Neural automated essay scoring and velopment 2(2):159–165.
coherence modeling for adversarially crafted input.
arXiv preprint arXiv:1804.06898 . Rada Mihalcea and Paul Tarau. 2004. Textrank: Bring-
ing order into text. In Proceedings of the 2004 con-
Jingjing Gong, Xinchi Chen, Xipeng Qiu, and Xu- ference on empirical methods in natural language
anjing Huang. 2016. End-to-end neural sentence processing.
ordering using pointer network. arXiv preprint
arXiv:1611.04953 . Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017.
Summarunner: A recurrent neural network based se-
Barbara J Grosz, Scott Weinstein, and Aravind K Joshi. quence model for extractive summarization of docu-
1995. Centering: A framework for modeling the lo- ments. In AAAI. pages 3075–3081.
cal coherence of discourse. Computational linguis-
tics 21(2):203–225. Ani Nenkova and Kathleen McKeown. 2012. A sur-
vey of text summarization techniques. In Mining
Karl Moritz Hermann, Tomas Kocisky, Edward text data, Springer, pages 43–76.
Grefenstette, Lasse Espeholt, Will Kay, Mustafa Su-
leyman, and Phil Blunsom. 2015. Teaching ma- Ani Nenkova, Kathleen McKeown, et al. 2011. Auto-
chines to read and comprehend. In Advances in Neu- matic summarization. Foundations and Trends
R in
ral Information Processing Systems. pages 1693– Information Retrieval 5(2–3):103–233.
1701.
Ani Nenkova and Lucy Vanderwende. 2005. The im-
Armand Joulin, Edouard Grave, Piotr Bojanowski, and pact of frequency on summarization. Microsoft Re-
Tomas Mikolov. 2016. Bag of tricks for efficient text search, Redmond, Washington, Tech. Rep. MSR-TR-
classification. arXiv preprint arXiv:1607.01759 . 2005 101.
Diederik P Kingma and Jimmy Ba. 2014. Adam: A Romain Paulus, Caiming Xiong, and Richard Socher.
method for stochastic optimization. arXiv preprint 2017. A deep reinforced model for abstractive sum-
arXiv:1412.6980 . marization. arXiv preprint arXiv:1705.04304 .
159
Mike Schuster and Kuldip K Paliwal. 1997. Bidirec-
tional recurrent neural networks. IEEE Transactions
on Signal Processing 45(11):2673–2681.
Abigail See, Peter J Liu, and Christopher D Manning.
2017. Get to the point: Summarization with pointer-
generator networks. In Proceedings of the 55th An-
nual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers). volume 1,
pages 1073–1083.
Simone Teufel et al. 1999. Argumentative zoning: In-
formation extraction from scientific text. Ph.D. the-
sis, Citeseer.
160
Risk Factors Extraction from Clinical Texts based on Linked Open Data
161
Proceedings of Recent Advances in Natural Language Processing, pages 161–167,
Varna, Bulgaria, Sep 2–4 2019.
semantic information provided by medical ontolo- Information Extraction (IE) refers to the auto-
gies and other resources in English, like scien- matic extraction of concepts, entities and events
tific publications and encyclopedic data. Several as well as their relations and associated attributes
data mining experiments were run on datasets con- from free text. A recent review of clinical IE ap-
taining outpatient records of diabetic patients in plications (Wang et al., 2018) notes the increas-
Bulgarian linked to encyclopedic extracts and Life ing interest to NLP but lists only 25 IE systems
Sciences LOD in English. The results show that which were used multiple times, outside the labs
LOD infuse some relations that are not found by where they were created. Isolated attempts ex-
standard text mining techniques of clinical narra- ist to apply IE in the context of EHR process-
tives, and thus enable the discovery of associations ing in frameworks for semantic search, for in-
hinting to further risk factors for diabetes mellitus. stance SemEHR deployed to identify contextu-
alized mentions of biomedical concepts within
2 Related Work EHRs in a number of UK hospitals (Wu et al.,
Mining of inter-related collections of clinical texts 2018). We mention the following research pro-
and LOD is still rare. On the one hand, with totypes as experimental developments, based on
hundreds of open biomedical ontologies and nu- some sort of IE: (Shi et al., 2017) reports about
merous biomedical datasets made available as a system extracting textual medical knowledge
LOD, there is a salient opportunity to integrate from heterogeneous sources in order to integrate
clinical and biomedical data to better interpret it into knowledge graphs; (Hassanpour and Lan-
patient-related texts and to uncover associations glotz, 2016) describes a machine learning system
of biomedical interest. On the other hand, such that annotates radiology reports and extracts con-
mining experiments require significant efforts to cepts according to a model covering most clini-
make clinical data interoperable with standardized cally significant contents in radiology; (Jackson
health terminologies, biomedical ontologies and et al., 2018) presents the information extraction
growing LOD repositories. One of the earliest pa- and retrieval architecture CogStack, deployed in
pers in this direction is (Pathak et al., 2013) which the King’s College Hospital. CogStack has func-
describes how patient EHRs data at Mayo Clinic tionality to transform records into de-identified
are represented as Resource Description Frame- text documents and applies generic clinical IE
work (RDF) in order to identify potential drug- pipelines to derive additional structured data from
drug interactions for widely prescribed cardiovas- free texts.
cular and gastroenterology drugs. Some drug-drug Most of the successful systems listed above
interactions of interest were identified which sug- work for clinical narratives in English. All ma-
gest lack of consensus on practice guidelines and jor resources, ontologies and terminology classifi-
recommendations. The authors of (Odgers and cations like UMLS4 and MESH5 are available in
Dumontier, 2015) describe how they transformed English. The comprehensive ontology SNOMED
a de-identified version of the STRIDE3 EHRs CT6 was developed initially in English and then
into a semantic clinical data warehouse contain- translated to other languages. Progress in biomed-
ing among others annotated clinical notes. They ical NLP for languages other than English will cat-
showed the feasibility of using semantic web tech- alyze the development of tools in the respective
nologies to directly exploit existing biomedical languages and will enable access to medical data
ontologies and LOD. As far as NLP is concerned, presented in a variety of languages (Névéol et al.,
an open-source tool (NegEx) is used in the EHR 2018). In Europe, the European commission sup-
transformation to recognize negated terms. The ports the development of multilingual platforms
integrated search in EHR data and LOD is not yet like SEMCARE which performs queries on un-
considered as a popular trend in the secondary use structured medical data in English, German, and
of clinical narratives (Meystre et al., 2017) and is Dutch (López-García et al., 2016).
still an emerging direction of research mostly due Using Big Data (nowadays - millions of EHRs)
to the complex data preparation. to advance medical research and health care prac-
3 4
Stanford Translational Research Integrated Database https://www.nlm.nih.gov/research/
Environment including a repository for EHR data, umls/
5
https://www.ncbi.nlm.nih.gov/pmc/ https://meshb.nlm.nih.gov/search
6
articles/PMC2815452/ http://www.snomed.org/
162
tices is now on the rise (Kessel and Combs, 2016). data related to diagnoses, risk factors and symp-
Core NLP components are already embedded in toms, the following open datasets are selected:
general clinical platforms similar to CogStack
(Jackson et al., 2018). Development of high qual- • Wikidata8 - contains multilingual encyclo-
ity corpora and terminology is a key factor for pedic information. Wikidata is a trusted
NLP progress in smaller languages. Here we em- resource, providing multilingual terminolo-
ploy English terminology in data mining tasks gies, their association with MESH codes, and
concerning EHRs in Bulgarian language. complex relations between diagnoses, risk
factors, and symptoms. Currently Wikidata
3 Materials contains descriptions of 5,227 items included
in ICD–10 and 10,517 descriptions of items
The datasets used in this study are a blend between included in ICD–10–CT. The main problem
LOD and clinical texts in Bulgarian language that is that many duplicated entities exist. For in-
belong to the Repository underpinning the Bulgar- stance, for ICD–10 code I20 there are two
ian Diabetes Register. items "angina pectoris (Q180762)" and "is-
The Register was automatically generated in chaemic heart disease (Q1444550)". Using
2015 from 260 million pseudonymized outpatient SPARQL9 queries, from Wikidata we collect
records (ORs) provided by the National Health In- for a given diagnosis all risk factors related
surance Fund (NHIF) for the period 2011–2014 to it as well as the associated MESH codes.
for more than 5 million citizens yearly, more than From the list of risk factors that is originally
7 million citizens in total (Boytcheva et al., 2017). in English we produce also a list in Bulgarian
Updated twice with data about 2015 and 2016, to- for the corresponding terms.
day the Register is maintained by the University
Specialized Hospital for Active Treatment of Di- • PubMed10 – the largest collection of scien-
abetes (USHATE) - Medical University Sofia. At tific publications in the area of biomedicine
present the Repository of ORs, which underpins and life sciences. From Pubmed we auto-
the Register, contains about 262 million records. matically extract publication abstracts and re-
These are reimbursement requests submitted by lated MESH terms via advanced queries11
General Practitioners and Specialists from Ambu- through API. The search is limited to 10,000
latory Care after every contact with a patient. The abstracts in order to keep balance between the
average number of patients with Diabetes Mellitus amounts of clinical narratives and texts of sci-
Type 2 (T2DM) per year is about 450,000. entific publications.
In the primary database, from where we ex- • Wikipedia – from Wikipedia we extract auto-
tract our datasets, the ORs are stored as semi- matically Wikipedia pages’ summaries for a
structured files with predefined XML-format. Ad- specified query via MediaWiki RESTful web
ministrative information is structured: visit date service API12 . The information in Wikipedia
and time; pseudonymized personal data and visit- is encyclopedic and more broader, thus the
related information, demographic data etc. All di- semantic information there is too vague and
agnoses are given by ICD–107 codes and location shallow, in contrast to PubMed abstracts.
names are specified in Bulgarian according to a
standard nomenclature. However much informa- • MESH ontology – this ontology is chosen be-
tion is provided as free text: anamnesis (case his- cause both Pubmed publications and Wiki-
tory, previous treatments, often family history, risk data contain references to it. In addition a
factors), patient status (summary of patient state, mapping between MESH and SNOMED CT
height, weight, body-mass index, blood pressure is available.
etc.), clinical tests (values of clinical examinations 8
https://www.wikidata.org/wiki/
and lab data listed in arbitrary order) as well as Wikidata:Main_Page
9
https://www.w3.org/TR/
prescribed treatment (codes of drugs reimbursed rdf-sparql-query/
by NHIF, free text descriptions of other drugs). 10
https://www.ncbi.nlm.nih.gov/pubmed/
11
To enhance clinical information with semantic https://www.ncbi.nlm.nih.gov/pmc/
articles/PMC2651214/
7 12
http://apps.who.int/classifications/ https://www.mediawiki.org/wiki/API:
icd10/browse/2016/en#/ Tutorial
163
Figure 1: Pipeline for identification of patients at risk
164
4.2 Semantic Model 5 Experiments and Results
Semantic Knowledge Graphs are used recently as
The diagnosis with ICD–10 code I20 "Ischaemic
powerful representation of entities and relations
heart disease" is chosen for experiments because
between them (Paulheim, 2017). Often knowledge
the Diabetes Register contains a plenty of clinical
graphs are generated automatically from semi-
descriptions about this case. Patients with T2DM
structured resources or from the documents under-
are at higher risk for developing I20 which is one
pinning various ontologies, through terms/words
of the T2DM complications. Sets of symptoms
they contain, by a combination of linguistic and
and risk factors for I20, both in Bulgarian and En-
statistical methods (Grainger et al., 2016).
glish, and English MeSH codes are automatically
In our experiment, all data are interlinked in a
extracted by Wikidata queries. Seven symptoms
Semantic Knowledge graph via MESH codes and
are extracted: angina pectoris, nausea, dyspnea,
ICD–10 codes. Wikidata is the mediator between
lightheadedness, unstable angina, neck pain, fa-
all resources providing cross-lingual ontology in-
tigue. Only for three of them there are labels in
formation and mapping between MESH and ICD-
Bulgarian, and for five of them there are MESH
10 codes. Term mappings from MeSH to ICD-10
codes. In addition, there are 18 risk factors and
are 1,535 and to ICD-10-CM are 2,127. In ad-
for 11 of them labels in Bulgarian exist. Multi-
dition Wikidata provides multilingual vocabulary
ple MeSH codes are associated to some risk fac-
for symptoms and diseases, and Pubmed publica-
tors but there diagnoses without associated MeSH
tions are annotated by Mesh codes.
code. The final cardinality of the generated term
For each symptom and risk factor, related to
set is |RE| = 24, |RB| = 14 and |M | = 34.
the selected diagnosis D, the system identifies
The Wikipedia API extracts 87 documents for
the most significant words related to D from the
a query with the RB terms in Bulgarian and we
Wikipedia and Pubmed datasets respectively, us-
limit the set to the top 5 related documents. They
ing p-value as a measure for their significance.
contain 14,600 tokens from 3,627 types. Al-
The knowledge graph is enriched with relations
though only the top 5 most related documents are
between these terms. The main relation between
taken into consideration, some of the extracted
clinical texts and other resources is based on ICD-
Wikipedia pages are not directly related to the
10 codes and some symptoms and risk factors that
symptoms and risk factors, as they discuss e.g.
are presented in the anamnesis section of ORs.
herbs and medications for treatment. But some
4.3 Predictive Model very related symptoms are included: for example
for "nausea" the Wikipedia page about "vomiting"
Two types of clinical texts are used as training
is extracted, and for "smoking" pages about "to-
datasets - for patients that have the diagnosis D,
bacco" and "pipe" are found. In addition, some
and for patients that do not have D. After text pre-
barely related pages are extracted – mainly about
processing, semantic hashing of all clinical texts
some famous people, who suffer from the diseases
in both datasets is done for predefined size of the
in question and have related symptoms. Unfor-
hash. Two predictive models are applied:
tunately the information in Wikipedia categories
• Based on the ORs information only "Medical conditions" and "Diseases" for Bulgar-
ian is too limited. For "Medical conditions" there
• Based on the ORs information enhanced with
are only 41 pages, and for many diseases the arti-
semantic data for symptoms and risk factors.
cles in the Bulgarian Wikipedia are stubs or some
In this case the vector space is extended; not
pages are not tagged in the respective categories.
only the dimensions of semantic hash vectors
For the query with RE terms 146 documents
are used, but also additional dimensions for
are identified that contain 40,099 tokens from
all symptoms, risk factors and the most sig-
3,803 types. The extracted Wikipedia pages in
nificant terms related to them.
English are also sometimes noisy and unrelated
Several machine learning techniques were used to mainly due to the ambiguity e.g. pages about
train the predictive model, including Naïve Bayes "Nausea(novel)" and "Nausea(band)", or "Insom-
(NB) (McCallum et al., 1998), kNN, Tree, Lo- nia (2002 film)" are extracted too. Other pages,
gistic Regression and Artificial Neural Networks. indirectly related to the risk factors, contain infor-
(Dreiseitl and Ohno-Machado, 2002). mation about Health organizations for treatment
165
of the respective diseases, about diagnostic proce- Model F1 P R
kNN 0.796 0.795 0.801
dures, medication for the treatment etc. Tree 0.778 0.776 0.783
Using the Pubmed advanced search, the gener- Neural Network 0.705 0.713 0.735
NB 0.588 0.615 0.701
ated query with RE terms identified 67,103 re- Logistic Regression 0.581 0.640 0.703
lated scientific papers, from which we retrieved
a subset of 2,000 abstracts only. The MeSH Table 1: Baseline: Performance of employed ML
headings only contain 104,363 tokens from 2,509 algorithm using semantic hashing over ORs only.
types. Both MeSH heading and Pubmed abstracts
Model F1 P R
contain 546,772 tokens from 12,684 types. kNN 0.819 0.752 0.899
Despite the imperfection of all texts extracted Tree 0.893 0.858 0.932
from Wikipedia and PubMed, the most significant Neural Network 0.743 0.746 0.760
NB 0.823 0.704 0.989
terms related to the predefined subsets of symp- Logistic Regression 0.825 0.703 0.999
toms and risk factors are sound and correct. Their
identification is based on bag–of–words and cal- Table 2: Performance of employed ML algorithm
culation of p-value (p ≤ 0.01) and False Dis- using semantic hashing of ORs enhanced with se-
covery Rate (F D ≤ 0.2). For example, for mantic model data.
тютюнопушене (tobacco smoking) the follow-
ing words are identified as relevant: тютюнопу- 6 Conclusion and Further Work
шене (tobacco smoking), дим (smoke), пуше-
не (smoking), тютюнев (tobacco), практику- The proposed approach shows how clinical texts
вам (practice), пристрастяване (addiction), can be enhanced with additional information about
пушач (smoker), цигара (cigar). These words the diseases, their symptoms and risk factors. The
were selected among 3,588 words in the text of re- experimental results show promising improve-
lated Wikipedia pages in Bulgarian. From those ment of the risk factors prediction accuracy. Still
words 754 were filtered as relevant, and the final there is a problem with Latin medical terminology
selection contain 9 words as most significant. that is often used in the Bulgarian clinical texts.
Two subsets of ORs (Anamnesis section) are ex- Another issue is the imperfection of the additional
tracted from the Repository behind the Diabetes terms provided by the LOD resources, due to
Register: S1 for 36,580 patients with diagnosis many ambiguous terms included there. As future
I20 and S2 for 86,000 patients without diagnosis work we plan to apply word sense disambigua-
I20. All clinical texts are preprocessed. The total tion to the texts extracted from open resources and
number of tokens in S1 and S2 is 123,258 from more precise methods for constructing the rela-
25,086 types. For experiments are used kNN (5 tions in the semantic knowledge graphs. As fu-
neighbours, Mahalanobis metric), Tree (Pruning: ture work we are planning to do deep analysis of
at least 2 instances in leaves, at least 5 instances in the individual contribution of each new term added
internal nodes, maximum depth 100; Stop splitting to the clinical texts. Another direction for future
when majority reaches 95%), Neural Network (30 work is to use some transfer learning methods like
hidden layers, Rectified Linear Activation Func- UMLfit (Howard and Ruder, 2018), BERT (De-
tion (ReLu) (Nair and Hinton, 2010), stochastic vlin et al., 2018) and XLnet (Yang et al., 2019) to
gradient-based optimizer Adam (Kingma and Ba, train models for word embedding on clinical texts
2014)), α = 0.0004, Max iterations 200, replica- in Bulgarian.
ble training) , NB and Logistic Regression (Ridge
L2). The baseline results of the prediction model
Acknowledgments
based on ORs only are presented in Table 1. This research is partially funded by the Bulgarian
The results of prediction models trained with Ministry of Education and Science, grant DO1-
enhanced LOD data (Table 2) show improvement 200/2018 ’Electronic health care in Bulgaria’ (e-
of the overall performance of all algorithms on Zdrave) and the Bulgarian National Science Fund,
this task, especially NB and Logistic Regression, grant DN-02/4-2016 ’Specialized Data Mining
when the clinical texts are enriched with additional Methods Based on Semantic Attributes’ (IZIDA).
information provided by open data resources and We are grateful to anonymous reviewers for useful
medical terminologies. comments and suggestions.
166
References SM Meystre, Christian Lovis, T Bürkle, G Tognola,
A Budrionis, and CU Lehmann. 2017. Clinical data
Svetla Boytcheva, Galia Angelova, Zhivko Angelov, reuse or secondary use: current status and potential
and Dimitar Tcharaktchiev. 2017. Integrating data future progress. Yearbook of medical informatics
analysis tools for better treatment of diabetic pa- 26(01):38–52.
tients. CEUR Workshop Proceedings 2022:229–
236. Vinod Nair and Geoffrey E Hinton. 2010. Rectified
linear units improve restricted boltzmann machines.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and In Proceedings of the 27th international conference
Kristina Toutanova. 2018. Bert: Pre-training of deep on machine learning (ICML-10). pages 807–814.
bidirectional transformers for language understand-
ing. arXiv preprint arXiv:1810.04805 . Aurélie Névéol, Hercules Dalianis, Sumithra Velupil-
lai, Guergana Savova, and Pierre Zweigenbaum.
Stephan Dreiseitl and Lucila Ohno-Machado. 2002. 2018. Clinical natural language processing in lan-
Logistic regression and artificial neural network guages other than english: opportunities and chal-
classification models: a methodology review. Jour- lenges. Journal of biomedical semantics 9(1):12.
nal of biomedical informatics 35(5-6):352–359.
David J Odgers and Michel Dumontier. 2015. Min-
Trey Grainger, Khalifeh AlJadda, Mohammed Ko- ing electronic health records using linked data.
rayem, and Andries Smith. 2016. The semantic AMIA Summits on Translational Science Proceed-
knowledge graph: A compact, auto-generated model ings 2015:217.
for real-time traversal and ranking of any relation-
Jyotishman Pathak, Richard C Kiefer, and Christo-
ship within a domain. In 2016 IEEE International
pher G Chute. 2013. Using linked data for
Conference on Data Science and Advanced Analyt-
mining drug-drug interactions in electronic health
ics (DSAA). IEEE, pages 420–429.
records. Studies in health technology and informat-
ics 192:682.
Saeed Hassanpour and Curtis Langlotz. 2016. Infor-
mation extraction from multi-institutional radiology Heiko Paulheim. 2017. Knowledge graph refinement:
reports. Artificial Intelligence in Medicine 66:29– A survey of approaches and evaluation methods. Se-
39. mantic web 8(3):489–508.
Jeremy Howard and Sebastian Ruder. 2018. Universal Martin F Porter. 2006. An algorithm for suffix strip-
language model fine-tuning for text classification. ping. Program .
arXiv preprint arXiv:1801.06146 .
Longxiang Shi, Shijian Li, Xiaoran Yang, Jiaheng Qi,
Richard Jackson, Ismail Kartoglu, Clive Stringer, et al. Gang Pan, and Binbin Zhou. 2017. Semantic health
2018. Cogstack - experiences of deploying inte- knowledge graph: Semantic integration of hetero-
grated information retrieval and extraction services geneous medical knowledge and services. BioMed
in a large national health service foundation trust research international 2017.
hospital. BMC Medical Informatics and Decision
Makingvolume 18. Yanshan Wang, Liwei Wang, Majid Rastegar-Mojarad,
Sungrim Moon, Feichen Shen, Naveed Afzal, Sijia
Kerstin A Kessel and Stephanie E Combs. 2016. Re- Liu, Yuqun Zeng, Saeed Mehrabi, Sunghwan Sohn,
view of developments in electronic, clinical data col- et al. 2018. Clinical information extraction appli-
lection, and documentation systems over the last cations: a literature review. Journal of biomedical
decade–are we ready for big data in routine health informatics 77:34–49.
care? Frontiers in oncology 6:75.
Honghan Wu, Giulia Toti, Katherine I Morley, Zina M
Ibrahim, Amos Folarin, Richard Jackson, Ismail
Diederik P Kingma and Jimmy Ba. 2014. Adam: A
Kartoglu, Asha Agrawal, Clive Stringer, Darren
method for stochastic optimization. arXiv preprint
Gale, et al. 2018. Semehr: A general-purpose se-
arXiv:1412.6980 .
mantic search system to surface semantic data from
clinical notes for tailored care, trial recruitment, and
Pablo López-García, Markus Kreuzthaler, Stefan
clinical research. Journal of the American Medical
Schulz, Daniel Scherr, Philipp Daumke, Kornél G
Informatics Association 25(5):530–537.
Markó, Jan A Kors, Erik M van Mulligen, Xinkai
Wang, Hanney Gonna, et al. 2016. Semcare: Mul- Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Car-
tilingual semantic search in semi-structured clinical bonell, Ruslan Salakhutdinov, and Quoc V Le.
data. In eHealth. pages 93–99. 2019. Xlnet: Generalized autoregressive pretrain-
ing for language understanding. arXiv preprint
Andrew McCallum, Kamal Nigam, et al. 1998. A com- arXiv:1906.08237 .
parison of event models for naive bayes text classi-
fication. In AAAI-98 workshop on learning for text
categorization. Citeseer, volume 752 (1), pages 41–
48.
167
Parallel Sentence Retrieval From Comparable Corpora for Biomedical
Text Simplification
168
Proceedings of Recent Advances in Natural Language Processing, pages 168–177,
Varna, Bulgaria, Sep 2–4 2019.
derstand medical and health information is an im- 2006; Nelken and Shieber, 2006; Zhu et al.,
portant issue, which motivates our work. 2010);
In order to perform biomedical text simplifi-
cation, we propose to collect parallel sentences, • knowledge-based methods which exploit ex-
which align difficult and simple information, as ternal resources, such as WordNet (Miller
they provide crucial and necessary indicators for et al., 1993) or PPDB (Ganitkevitch et al.,
automatic systems for text simplification. Indeed, 2013). The features exploited can be: over-
such pairs of sentences contain cues on transfor- lap with external resources, distance between
mations which are suitable for the simplification, the synsets, intersection of synsets, seman-
such as lexical substitutes and syntactic modifica- tic similarity of resource graphs, presence of
tions. Yet, this kind of resources is seldom avail- synonyms, hyperonyms or antonyms (Mihal-
able, especially in languages other than English. cea et al., 2006; Fernando and Stevenson,
As a matter of fact, it is easier to access compa- 2008; Lai and Hockenmaier, 2014);
rable corpora: they cover the same topics but are
differentiated by their registers (documents cre- • syntax-based methods which exploit the syn-
ated for medical professionals and documents cre- tactic modelling of sentences. The fea-
ated for patients). More precisely, we can exploit tures often exploited are: syntactic cate-
an existing monolingual comparable corpus with gories, syntactic overlap, syntactic dependen-
medical documents in French (Grabar and Cardon, cies and constituents, predicat-argument rela-
2018). The purpose of our work is to detect and tions, edition distance between syntactic trees
align parallel sentences from this comparable cor- (Wan et al., 2006; Severyn et al., 2013; Tai
pus. We also propose to test what is the impact of et al., 2015; Tsubaki et al., 2016);
imbalance on categorization results: imbalance of
categories is indeed the natural characteristics in • corpus-based methods which exploit distri-
textual data. butional methods, latent semantic analysis
The existing work on searching parallel sen- (LSA), topics modelling, word embeddings,
tences in monolingual comparable corpora indi- etc. (Barzilay and Elhadad, 2003; Guo and
cates that the main difficulty is that such sentences Diab, 2012; Zhao et al., 2014; Kiros et al.,
may show low lexical overlap but be neverthe- 2015; He et al., 2015; Mueller and Thyagara-
less parallel. Recently, this task gained in pop- jan, 2016).
ularity in general-language domain thanks to the
semantic text similarity (STS) initiative. Dedi- There has been work for detection of para-
cated SemEval competitions have been proposed phrases in French comparable biomedical corpora
for several years (Agirre et al., 2013, 2015, 2016). (Deléger and Zweigenbaum, 2009), but there is no
The objective, for a given pair of sentences, is work on building a corpus for text simplification
to predict whether they are semantically similar in the biomedical domain. Our work is positioned
and to assign a similarity score going from 0 in this area.
(independent semantics) to 5 (semantic equiva- In what follows, we first present the linguistic
lence). This task is usually explored in general- material used, and the methods proposed. We then
language corpora (Coster and Kauchak, 2011; present and discuss the results obtained, and con-
Hwang et al., 2015; Kajiwara and Komachi, 2016; clude with directions of future work.
Brunato et al., 2016). Among the exploited meth-
ods, we can notice: 2 Method
We use the CLEAR comparable medical cor-
• lexicon-based methods which rely on simi-
pus (Grabar and Cardon, 2018) available online1
larity of subwords or words from the pro-
which contains three comparable sub-corpora in
cessed texts or on machine translation (Mad-
French. Documents within these sub-corpora are
nani et al., 2012). The features exploited can
contrasted by the degree of technicality of the in-
be: lexical overlap, sentence length, string
formation they contain with typically specialized
edition distance, numbers, named entities,
the longest common substring (Clough et al., 1
http://natalia.grabar.free.fr/
2002; Zhang and Patrick, 2005; Qiu et al., resources.php#clear
169
and simplified versions of a given text. These cor- Those three corpora have different degrees of
pora cover three genres: drug information, sum- parallelism: Wikipedia and Vikidia articles are
maries of scientific articles, and encyclopedia arti- written independently from each other, drug infor-
cles. We also exploit a reference dataset with sen- mation documents are related to the same drugs
tences manually aligned by two annotators. but the types of information presented for experts
and laypeople vary, while simplified summaries
2.1 Comparable Corpora from the Scientific corpus are created starting from
Table 1 indicates the size of the comparable cor- the expert summaries.
pus in French: number of documents, number
2.2 Reference Data
of words (occurrences and lemmas) in special-
ized and simplified versions. This information The reference data with aligned sentence pairs,
is detailed for each sub-corpus: drug information which associate technical and simplified contents,
(Drugs), summaries of scientific articles (Scient.), are created manually. We have randomly selected
and encyclopedia articles (Encyc.). 2x14 Encyclopedia articles, 2x12 Drugs docu-
The Drugs corpus contains drug information ments, and 2x13 Scientific summaries. The sen-
such as provided to health professionals and pa- tence alignment is done by two annotators follow-
tients. Indeed, two distinct sets of documents ex- ing these guidelines:
ist, each of which contains common and specific
information. This corpus is built from the public 1. exclude identical sentences or sentences with
drug database2 of the French Health ministry. Spe- only punctuation and stopword difference ;
cialized versions of documents provide more word 2. include sentence pairs with morphological
occurrences than simplified versions. variations (e.g. Ne pas dépasser la posolo-
The Scientific corpus contains summaries of gie recommandée. and Ne dépassez pas la
meta-reviews of high evidence health-related ar- posologie recommandée. – both examples
ticles, such as proposed by the Cochrane collabo- can be translated by Do not take more than
ration (Sackett et al., 1996). These reviews have the recommended dose.);
been first intended for health professionals but re-
cently the collaborators started to create simplified 3. exclude sentence pairs with overlapping se-
versions of the reviews (Plain language summary) mantics, when each sentence brings own con-
so that they can be read and understood by the tent, in addition to the common semantics;
whole population. This corpus has been built from
the online library of the Cochrane collaboration3 . 4. include sentence pairs in which one sentence
Here again, specialized version of summaries is is included in the other, which enables many-
larger than the simplified version, although the dif- to-one matching (e.g. C’est un organe fait de
ference is not very important. tissus membraneux et musculaires, d’environ
The Encyclopedia corpus contains encyclo- 10 à 15 mm de long, qui pend à la partie
pedia articles from Wikipedia4 and Vikidia5 . moyenne du voile du palais. and Elle est con-
Wikipedia articles are considered as technical stituée d’ un tissu membraneux et musculaire.
texts while Vikidia articles are considered as their – It is an organ made of membranous and
simplified versions (they are created for children muscular tissues, approximately 10 to 15 mm
from 8 to 13 year old). Similarly to the works long, that hangs from the medium part of the
done in English, we associate Vikidia with Simple soft palate. and It is made of a membranous
Wikipedia6 . Only articles indexed in the medical and muscular tissue.);
portal are exploited in this work. From Table 1, we
5. include sentence pairs with equivalent se-
can see that specialized versions (from Wikipedia)
mantics – other than semantic intersection
are also longer than simplified versions.
and inclusion (e.g. Les médicaments in-
2
http://base-donnees-publique. hibant le péristaltisme sont contre-indiqués
medicaments.gouv.fr/ dans cette situation. and Dans ce cas, ne
3
http://www.cochranelibrary.com/
4
https://fr.wikipedia.org
prenez pas de médicaments destinés à blo-
5
https://fr.vikidia.org quer ou ralentir le transit intestinal. – Drugs
6
http://simple.wikipedia.org that inhibit peristalsis are contraindicated in
170
corpus # docs # occsp # occsimpl # lemmassp # lemmassimpl
Drugs 11,800x2 52,313,126 33,682,889 43,515 25,725
Scient. 3,815x2 2,840,003 1,515,051 11,558 7,567
Encyc. 575x2 2,293,078 197,672 19,287 3,117
Table 1: Size of the three source corpora. Column headers: number of documents, total number of occurrences
(specialized and simple), total number of unique lemmas (specialized and simple)
Table 2: Size of the reference data with consensual alignment of sentences. Column headers: number of docu-
ments, sentences and word occurrences for each subset, alignment rate
that situation. and In that case, do not take ment of abdominal pain, global assessment (over-
drugs intended for blocking or slowing down all relief of IBS symptoms) or symptom score. In
the intestinal transit.). the Encylopedia corpus such notions are replaced
by simpler words, or removed. Finally, in all cor-
The judgement on semantic closeness may vary pora, we observe frequent substitutions by syn-
according to the annotators. For this reason, the onyms, like {nutrition, food}, {enteral, directly in
alignments provided by each annotator undergo the stomach}, or {hypersensitivity, allergy}. No-
consensus discussions. This alignment process tice that with such substitutions, lexical similarity
provides a set of 663 aligned sentence pairs. The between sentences is reduced.
inter-annotator agreement is 0.76 (Cohen, 1960).
The documents are pre-processed. They are
It is computed within the two sets of sentences
segmented into sentences using strong punctuation
proposed for alignment by the two annotators.
(i.e. .?!;:). We removed, from each subcorpus, the
Table 2 shows the details of the manually
sentences that are found in at least half of the doc-
aligned set of sentences. Because the three cor-
uments of a given corpus. Those sentences are typ-
pora vary in their capacity to provide parallel sen-
ically legal notices, section titles, and remainders
tences, we compute their alignment rate. The
from the conversion of the HTML versions of the
alignment rate for a given corpus is the number of
documents. The lines that contain no alphabetic
sentences that are part of an aligned pair relative to
characters have also been removed. That reduces
the total number of sentences. As expected, only
the total number of possible pairs for each docu-
a tiny fraction of all possible pairs corresponds to
ment pair approximately from 940,000 to 590,000.
aligned sentences. We can observe that the Sci-
entific corpus is the most parallel with the highest 2.3 Automatic Detection and Alignment of
alignment rate of sentences, while the two other Parallel Sentences
corpora (Drugs and Encylopedia) contain propor-
Automatic detection and alignment of parallel sen-
tionally less parallel sentences. Sentences from
tences is the main step of our work. The unity
simplified documents in the Scientific and drugs
processed is a pair of sentences. The objective is
corpora are longer than sentences from special-
to categorize the pairs of sentences in one of the
ized documents because they often add explana-
two categories:
tions for technical notions, like in this example:
We considered studies involving bulking agents (a • alignment: the sentences are parallel and can
fibre supplement), antispasmodics (smooth muscle be aligned;
relaxants) or antidepressants (drugs used to treat
depression that can also change pain perceptions) • non-alignment: the sentences are non-
that used outcome measures including improve- parallel and cannot be aligned.
171
The reference data provide 663 positive exam- sub-word level, some sequences of characters
ples (parallel sentence pairs). In order to perform may be meaningful for the alignment of sen-
the automatic categorization, we also need nega- tences if they are shared by them;
tive examples, which are obtained by randomly
6. Word-based similarity measure exploits three
pairing all sentences from all the document pairs
scores (cosine, Dice and Jaccard). This fea-
except the sentence pairs that are already found to
ture provides a more sophisticated indica-
be parallel. Approximately 590,000 non-parallel
tion on word overlap between two sentences.
sentences pairs are created in this way. That high
Weight assigned to each word is set to 1;
degree of imbalance is the main challenge in our
work and we address it in the experimental design 7. Character-based minimal edit distance (Lev-
(sec 2.4). enshtein, 1966). This is a classical acception
For the automatic alignment of parallel sen- of edit distance. It takes into account basic
tences, we first use a binary classification model edit operations (insertion, deletion and sub-
that relies on the random forests algorithm stitution) at the level of characters. The cost
(Breiman, 2001). The implementation we use is of each operation is set to 1;
the one that is available in scikit-learn (Pedregosa
et al., 2011). Our goal is to propose features that 8. Word-based minimal edit distance (Leven-
can work on textual data in different languages shtein, 1966). This feature is computed with
and registers. We use several features which are words as units within sentence. It takes into
mainly lexicon-based and corpus-based, so that account the same three edit operations with
they can be easily applied to textual data in other the same cost set to 1. This feature permits to
corpora, speacialized areas and languages or trans- compute the cost of lexical transformation of
posed on them. The features are computed on one sentence into another;
word forms (occurrences). The features are the 9. WAVG. This features uses word embeddings.
following: The word vectors of each sentence are aver-
aged, and the similarity score is calculated by
1. Number of common non-stopwords. This
comparing the two resulting sentence vectors
feature permits to compute the basic lexical
(Stajner et al., 2018);
overlap between specialized and simplified
versions of sentences (Barzilay and Elhadad, 10. CWASA. This feature is the continuous word
2003). It concentrates on non-lexical content alignment-based similarity analysis, as de-
of sentences; scribed in (Franco-Salvador et al., 2016).
2. Percentage of words from one sentence in- For the last two features, we trained the em-
cluded in the other sentence, computed in beddings on the CLEAR corpus using word2vec
both directions. This features represents pos- (Mikolov et al., 2013), and the scores are com-
sible lexical and semantic inclusion relations puted using the CATS tool (Stajner et al., 2018).
between the sentences; 2.4 Experimental Design
3. Sentence length difference between special- The set with manually aligned pairs is divided into
ized and simplified sentences. This feature three subsets:
assumes that simplification may imply stable • equivalence: 238 pairs with equivalent se-
association with the sentence length; mantics,
4. Average length difference in words between • tech in simp: 237 pairs with inclusion where
specialized and simplified sentences. This the content of technical sentence is fully in-
feature is similar to the previous one but takes cluded in simplified sentence, and simplified
into account average difference in sentence sentence provides additional content,
length;
• simp in tech: 112 pairs with inclusion where
5. Total number of common bigrams and tri- the content of simplified sentence is fully in-
grams. This feature is computed on charac- cluded in technical sentence, and technical
ter ngrams. The assumption is that, at the sentence provides additional content.
172
(a) equivalence, test subsets (b) inclusion, technical in simple, (c) inclusion, simple in technical,
test subsets test subsets
(d) equivalence, real data (e) inclusion, technical in simple, (f) inclusion, simple in technical,
real data real data
Figure 1: Precision, Recall and F-1 for the various experiments and subsets
For each subset, we perform two sets of experi- the global score from the observations is that when
ments: the data are imbalanced (negative class is grow-
ing progressively), misclassifying the positive data
1. We train and test the model with balanced has little influence over the global scores, which
data (we randomly select as many non- thus always appear to be high (metrics above
aligned pairs as aligned pairs), and then we 0.99).
progressively increase the number of non-
Finally, we apply the best model for equivalent
aligned pairs until we reach a ratio of 3000:1,
pairs on another 30 randomly selected documents
which is close to the real data (∼4000:1).
and evaluate the output.
2. Then, for each ratio, we apply the obtained
model to the whole dataset and evaluate the 3 Presentation and Discussion of Results
results. Note that the training data is included
We present the results in Figure 1: The x axis rep-
in the whole dataset, we proceed this way be-
resents the growing of imbalance (the first posi-
cause of the low volume of available data.
tion is 1 and corresponds to balanced data), while
As there is some degree of variability coming the y axis represents the values of Precision, Re-
with the subset of non-aligned pairs that are ran- call and F-measure. The results for the three sub-
domly selected for the imbalance ratio, every sin- sets are presented: equivalence (Figures 1(a) and
gle one of those experiments has been performed 1(d)), inclusion of technical sentence in simple
fifty times: the results that are presented corre- sentence (Figures 1(b) and 1(e)), and inclusion
spond to the mean values over the fifty runs. of simple sentence in technical sentence (Figures
1(c) and 1(f)). Besides, Figures 1(a), 1(b) and 1(c)
2.5 Evaluation present the results obtained by training and test-
For evaluating the results, in each experiment we ing the model on the dataset with the same im-
divide the indicated datasets in two parts: two balance ratio (first set of experiments described in
thirds for training and one third for testing. The section 2.4). As for Figures 1(d), 1(e) and 1(f),
metrics we use are Recall, Precision and F1 scores. they present the results obtained by the models
As we are primarily focused on detection of the mentioned above that are applied on the whole set
aligned pairs, we only report scores for that class. of manually annotated data (second set of experi-
Another reason to exclude the negative class and ments described in section 2.4).
173
equivalence simp. in tech. tech. in simp. intersection false positives
nb. of pairs 56 10 4 9 1
ratio 70% 12.5% 5% 11.25% 1.25%
Table 3: Breakdown by pair types of the output of the model trained on equivalent pairs with an imbalance ratio of
1200:1 and applied to 30 randomly chosen pairs of documents
The most visible conclusion we can draw from again the equivalence dataset resists better. An in-
those experiments is that equivalent pairs (Figures teresting fact is that, when the model is learned
1(a) and 1(d)) are easier to classify than inclusion on a substantial degree of imbalance, the Precision
pairs (the rest of the Figures). Values of both, Pre- score is high when that model is applied to the real
cision and Recall, are higher on the equivalence data, which has a ratio of about 4,000:1. This is
dataset at different imbalance points. For instance, interesting because it shows that the model is par-
with training on the equally balanced dataset (po- ticularly good at discriminating counter-examples.
sition 1 on Figure 1(a)), the scores for Precision The recall value is also high, but since two thirds
(0.98) and Recall (0.95) are higher than the scores of the real data examples have been used for train-
obtained by the technical in simple dataset (0.96 ing, that good score should be considered cau-
Precision and 0.94 Recall) and the simple in tech- tiously. We are planning to evaluate the models on
nical dataset (0.95 Precision and 0.93 Recall) at a separate set of manually annotated documents.
the same point. For the application to the real data, This is still a good result, as during the tests that
for ratio 1200:1 – the point where Precision and we performed with other classification algorithms,
Recall meet for equivalent pairs, see Figure 1(d) the models did not successfully recognize the ex-
– we obtain 0.81 Precision and 0.81 Recall. At amples they had seen during training.
that same ratio, for the technical in simple pairs For further evaluation, we randomly selected 30
the scores are 0.65 Precision and 0.73 Recall, and pairs of documents to evaluate the performances
for the simple in technical pairs Precision is 0.73 of the models. We used the model that was trained
and Recall is 0.70. This result is positive because at a ratio of 1200:1 on equivalent pairs. In terms
the equivalence dataset usually provides the main of precision, the model shows 98.75% on all the
and the most complete information on transforma- sentence pairs aligned (80 sentence pairs), includ-
tions required for the simplification. As for the ing equivalence, inclusions and intersection. Ta-
inclusion relations, they cover a large variety of ble 3 shows the breakdown of this output in terms
situations which do not necessarily correspond to of pair types: 70% (56 pairs) are equivalent pairs,
the searched information. This is illustrated by the 29% (23 pairs) are examples of inclusion (10 sim-
unstability of the curves in Figures 1(b) and 1(c), ple in technical, 4 technical in simple) and inter-
whereas they are smooth in Figure 1(a). The neg- section (9), and one pair contains two unrelated
ative examples subset seems to have a quite high sentences. Those results show that we have a
influence on the results, which indicates that it is model that can be used to automatically generate
more difficult to draw a clear separation between a parallel corpus with reduced noise, from highly
positive and negative examples. We need to design imbalanced comparable corpus, for text simplifi-
additional processing steps to be able to classify cation purposes.
those pairs in a more efficient way.
4 Conclusion and Future Work
We can also observe from Figures 1(a), 1(b)
and 1(c) that the use of balanced data provides We addressed the task of detection and alignment
very high results, both for Precision and Recall, of parallel sentences from a monolingual compa-
which are very close to the reference data (> 0.90 rable French corpus. The comparable aspect is on
performance). This is true for the three subsets the technicality axis, as the corpus contrasts tech-
tested (equivalence and inclusions). These good nical and simplified versions of information on the
results in an artifical setting cannot be applied to same subjects. We use the CLEAR corpus, that is
the real dataset, as is indicated by the starting point related to the biomedical area.
in Figures 1(d), 1(e) and 1(f). Yet, the imbalance Several experiments were performed. We di-
has greater effect on the inclusion datasets, while vide the task in three subtasks – equivalent pairs,
174
and inclusion on both directions – and make obser- Leo Breiman. 2001. Random forests. Machine learn-
vations on the effect of imbalance during training ing, 45(1):5–32.
on the performance on the real data. We show that Dominique Brunato, Andrea Cimino, Felice
increasing the imbalance during training increases Dell’Orletta, and Giulia Venturi. 2016. PaCCSS-IT:
the Precision of the model while still maintaining A parallel corpus of complex-simple sentences for
a stable value for Recall. We also find that the task automatic text simplification. In Proceedings of the
2016 Conference on Empirical Methods in Natural
is easier to perform on sentence pairs that have the Language Processing, pages 351–361, Austin,
same meaning, than on sentence pairs where one Texas. Association for Computational Linguistics.
is included in the other.
We will use that model to generate a corpus of Ping Chen, John Rochford, David N. Kennedy, Sous-
san Djamasbi, Peter Fay, and Will Scott. 2016. Au-
parallel sentences in order to work on the devel- tomatic text simplification for people with intellec-
opment of methods for biomedical text simplifi- tual disabilities. In AIST, pages 1–9.
cation in French. We will also perform experi-
ments on the general language. Another task we Paul Clough, Robert Gaizauskas, Scott S.L. Piao, and
Yorick Wilks. 2002. METER: Measuring text reuse.
will explore addresses the question on how that In ACL, pages 152–159.
model performs with the cross-lingual transfer of
descriptors and models. Jacob Cohen. 1960. A coefficient of agreement for
nominal scales. Educational and Psychological
Measurement, 20(1):37–46.
5 Acknowledgements
William Coster and David Kauchak. 2011. Simple En-
This work was funded by the French National glish Wikipedia: A new text simplification task. In
Agency for Research (ANR) as part of the CLEAR Proceedings of the 49th Annual Meeting of the Asso-
project (Communication, Literacy, Education, Ac- ciation for Computational Linguistics: Human Lan-
cessibility, Readability), ANR-17-CE19-0016-01. guage Technologies, pages 665–669, Portland, Ore-
gon, USA. Association for Computational Linguis-
The authors would like to thank the reviewers for tics.
their helpful comments.
Louise Deléger and Pierre Zweigenbaum. 2009. Ex-
tracting lay paraphrases of specialized expressions
from monolingual comparable medical corpora. In
References Proceedings of the 2nd Workshop on Building and
Eneko Agirre, Carmen Banea, Claire Cardie, Daniel Using Comparable Corpora: from Parallel to Non-
Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei parallel Corpora (BUCC), pages 2–10, Singapore.
Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Association for Computational Linguistics.
Mihalcea, German Rigau, Larraitz Uria, and Janyce
Wiebe. 2015. SemEval-2015 task 2: Semantic tex- Samuel Fernando and Mark Stevenson. 2008. A se-
tual similarity, english, spanish and pilot on inter- mantic similarity approach to paraphrase detection.
pretability. In SemEval 2015, pages 252–263. In Comp Ling UK, pages 1–7.
175
William Hwang, Hannaneh Hajishirzi, Mari Ostendorf, George A. Miller, Richard Beckwith, Christiane Fell-
and Wei Wu. 2015. Aligning sentences from stan- baum, Derek Gross, and Katherine Miller. 1993. In-
dard Wikipedia to simple Wikipedia. In Proceed- troduction to wordnet: An on-line lexical database.
ings of the 2015 Conference of the North Ameri- Technical report, WordNet.
can Chapter of the Association for Computational
Linguistics: Human Language Technologies, pages Jonas Mueller and Aditya Thyagarajan. 2016. Siamese
211–217, Denver, Colorado. Association for Com- recurrent architectures for learning sentence similar-
putational Linguistics. ity. In AAAI Conference on Artificial Intelligence,
pages 2786–2792.
Tomoyuki Kajiwara and Mamoru Komachi. 2016.
Building a monolingual parallel corpus for text sim- Rani Nelken and Stuart M. Shieber. 2006. Towards ro-
plification using sentence similarity based on align- bust context-sensitive sentence alignment for mono-
ment between word embeddings. In Proceedings lingual corpora. In EACL, pages 161–168.
of COLING 2016, the 26th International Confer-
ence on Computational Linguistics: Technical Pa- Gustavo H. Paetzold and Lucia Specia. 2016. Bench-
pers, pages 1147–1158, Osaka, Japan. The COLING marking lexical simplification systems. In LREC,
2016 Organizing Committee. pages 3074–3080.
Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov, F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
Richard S. Zemel, Antonio Torralba, Raquel Urta- B. Thirion, O. Grisel, M. Blondel, P. Pretten-
sun, and Sanja Fidler. 2015. Skip-thought vectors. hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-
In Neural Information Processing Systems (NIPS), sos, D. Cournapeau, M. Brucher, M. Perrot, and
pages 3294–3302. E. Duchesnay. 2011. Scikit-learn: Machine learning
Philipp Koehn. 2005. Europarl: A Parallel Corpus for in Python. Journal of Machine Learning Research,
Statistical Machine Translation. In Conference Pro- 12:2825–2830.
ceedings: the tenth Machine Translation Summit,
pages 79–86, Phuket, Thailand. AAMT, AAMT. Long Qiu, Min-Yen Kan, and Tat-Seng Chua. 2006.
Paraphrase recognition via dissimilarity significance
Alice Lai and Julia Hockenmaier. 2014. Illinois-LH: A classification. In Empirical Methods in Natural
denotational and distributional approach to seman- Language Processing, pages 18–26, Sydney, Aus-
tics. In Workshop on Semantic Evaluation (SemEval tralia.
2014), pages 239–334, Dublin, Ireland.
ED Rudd. 2013. Needed action in health literacy. J
Gondy Leroy, David Kauchak, and Obay Mouradi. Health Psychol, 18(8):1004–10.
2013. A user-study measuring the effects of lexical
simplification and coherence enhancement on per- David L. Sackett, William M. C. Rosenberg, Jeffrey A.
ceived and actual text difficulty. Int J Med Inform, MuirGray, R. Brian Haynes, and W. Scott Richard-
82(8):717–730. son. 1996. Evidence based medicine: what it is and
what it isn’t. BMJ, 312(7023):71–2.
Vladimir I. Levenshtein. 1966. Binary codes capable
of correcting deletions, insertions and reversals. So- Aliaksei Severyn, Massimo Nicosia, and Alessandro
viet physics. Doklady, 707(10). Moschitti. 2013. Learning semantic textual similar-
ity with structural representations. In Annual Meet-
Nitin Madnani, Joel Tetreault, and Martin Chodorow. ing of the Association for Computational Linguis-
2012. Re-examining machine translation metrics tics, pages 714–718.
for paraphrase identification. In NAACL-HLT, pages
182–190. Sanja Stajner, Marc Franco-Salvador, Simone Paolo
A Mcgray. 2005. Promoting health literacy. J of Am Ponzetto, and Paolo Rosso. 2018. Cats: A tool
Med Infor Ass, 12:152–163. for customised alignment of text simplification cor-
pora. In Proceedings of the 11th Language Re-
Rada Mihalcea, Courtney Corley, and Carlo Strap- sources and Evaluation Conference, LREC 2018,
parava. 2006. Corpus-based and knowledge-based Miyazaki, Japan, May 7-12.
measures of text semantic similarity. In AAAI, pages
1–6. Kai Sheng Tai, Richard Socher, and Christopher D.
Manning. 2015. Improved semantic representations
Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S. from tree-structured long short-term memory net-
Corrado, and Jeffrey Dean. 2013. Distributed rep- works. In Annual Meeting of the Association for
resentations of words and phrases and their com- Computational Linguistics, pages 1556–1566, Bei-
positionality. In Advances in Neural Information jing, China.
Processing Systems 26: 27th Annual Conference on
Neural Information Processing Systems 2013. Pro- Masashi Tsubaki, Kevin Duh, Masashi Shimbo, and
ceedings of a meeting held December 5-8, 2013, Yuji Matsumoto. 2016. Non-linear similarity learn-
Lake Tahoe, Nevada, United States., pages 3111– ing for compositionality. In AAAI Conference on Ar-
3119. tificial Intelligence, pages 2828–2834.
176
Tu Thanh Vu, Giang Binh Tran, and Son Bao
Pham. 2014. Learning to simplify children stories
with limited data. In Intelligent Information and
Database Systems, pages 31–41.
Stephen Wan, Mark Dras, Robert Dale, and Cecile
Paris. 2006. Using dependency-based features to
take the para-farce out of paraphrase. In Aus-
tralasian Language Technology Workshop, pages
131–138.
Yitao Zhang and Jon Patrick. 2005. Paraphrase iden-
tification by text canonicalization. In Australasian
Language Technology Workshop, pages 160–166.
Jiang Zhao, Tian Tian Zhu, and Man Lan. 2014.
ECNU: One stone two birds: Ensemble of heteroge-
nous measures for semantic relatedness and textual
entailment. In Workshop on Semantic Evaluation
(SemEval 2014), pages 271–277.
Zhemin Zhu, Delphine Bernhard, and Iryna Gurevych.
2010. A monolingual tree-based translation model
for sentence simplification. In COLING 2010, pages
1353–1361.
177
Classifying Author Intention for Writer Feedback in Related Work
178
Proceedings of Recent Advances in Natural Language Processing, pages 178–187,
Varna, Bulgaria, Sep 2–4 2019.
tion of author intentions and show that the novel lated Work is (Angrosh et al., 2010). This work
features we introduce contribute significantly to focuses on sentences in terms of their ability to
the classifier performance, improving on perfor- support intelligent information retrieval in digital
mance of existing writer feedback tools. library services. While aspects of this work and
the previous works on citation function are rele-
2 Related Work vant, what is missing is the need to identify where
an author talks about their own work in context to
Automating Author Intentions – Previous mod- other work, showing why it is different or how it
els of author intentions in research articles have fills a gap. As discussed in the Introduction, one
been successfully automated. One of the first and of the problems with poor writing in Related Work
widely used is Teufel (1999) who proposed Ar- is bland statements that provide lists of citations.
gument Zoning (AZ) which labels sentences with Cited works should be ones that have implications
zones representing the rhetoric purpose (author for the author’s work (Maxwell, 2006). To pro-
intent) within the global context of a document vide such feedback, we must capture this context
e.g. background, aim or conclusion. Further work in addition to citation function.
has applied this schema to biology papers (Mizuta
Recognising Author Intentions – Specific
and Collier, 2004), with a modified, finer grained
phrasing has been shown to function in structuring
approach applied to papers on chemistry (Teufel
discourse by guiding readers through a text (Hy-
et al., 2009). Liakata et al. (2012) took a different
land, 2012) and can be found to align to sections,
approach to labelling author intentions, studying
such as the Introduction or Results. Most pre-
the conceptual structure of biology articles treat-
vious work in automating author intentions have
ing the article as an investigation. Fisas et al.
utilised these patterns as part of their feature set.
(2015) develop a schema based on both Liakata
The early work of Teufel (1999) (in the domain
and Teufel’s work to represent scientific concepts
of computational linguistics) uses cue phrases and
that appear in computer graphics articles. These
lexical patterns that involve parts of speech and ci-
works have successfully identified author inten-
tation markers as features. Jurgens et al. (2018)
tions, but they differ from our work by seeking
shows how applying bootstrapping to Tuefel’s lex-
intentions in a global context across a whole ar-
icon improves citation function recognition.
ticle. For example, AZ was developed to support
summarisation and information access. The au- Verbs have been shown to have a role in under-
thor intentions that these activities would be asso- standing citation function by determining rhetori-
ciated with are rarely found in a Related Work sec- cal and semantic levels. Verbs used to report can
tion and are unlikely to be helpful in writing feed- show positive and negative aspects of evaluation
back for this section. They are nevertheless useful in cited works and differentiate between intentions
in supporting writing feedback on Abstracts and e.g in Angrosh et al. (2010) they use reporting
summaries of PhDs (Feltrim et al., 2006). verbs that describe (examine, propose), refer to
Related Work does have in common with other an outcome (develop, show) or show a strength
sections the fact that it should contain citations. (improve). Citation forms (Swales, 1990) of in-
Understanding the motivations or function of a tegral and non-integral have been shown to be a
citation can help determine an author intention contributing feature to author intention recogni-
(Teufel et al., 2006). Work on citation function has tion, with studies of novice writers showing that
been an area of research for several decades (We- they use a limited range of citation types (Thomp-
instock, 1971; Oppenheim and Renn, 1978; Teufel son and Tribble, 2001). Our approach also uses
et al., 2006; Angrosh et al., 2012), with more re- linguistic patterns, verb types and citation types to
cent work considering how this recognition can be support building our feature set, and we do this
automated. Jurgens et al. (2018) investigates the within one domain, computational linguistics (cf.
framing of citations and how this can be used to Section 4 ).
study the evolution of a field. Teufel et al. (2006) Automated Assessment of Writing – Exist-
work on automated recognition of citation func- ing, academic writing tools have focused on iden-
tion and show a strong relationship between func- tifying author intentions, such as those described
tion and sentiment. One work that specifically by Swales (1990), that can be found in an Intro-
looks at context identification of sentences in Re- duction (Cotos and Pendar, 2016; Anthony and
179
V. Lashkia, 2003; Abel, 2018). The Criterion on- lates to the cited work or background in general.
line writing service, focuses on automated persua- The sentence label schema we use can be found
sive essay evaluation and uses recognition of dis- in Table 1, and we indicate which of the qualities
course elements based on aspects such as support- each label falls into.
ing ideas, introductions and conclusion (Burstein
et al., 2003, 2004). Several other works have fo-
cused on identifying argument components and re- 3.2 Annotated Dataset
lations and how these relate to essay scores (Ghosh
et al., 2016; Song et al., 2014). Recognizing argu- The annotated dataset in (Casey et al., 2019) is
ment components in this case focuses on premises composed of papers from the ACL anthology
and claims largely based on the Toulmin model of (Bird et al., 2008) that have been pre-annotated
argumentation (Toulmin, 2003) which is a differ- for citations and co-reference to the author’s own
ent approach to ours. In addition, all this work work by (Schäfer et al., 2012). We use 94 pa-
focuses on feedback for persuasive essays which pers with Related Work sections after removing
will differ in linguistic practices found in scien- one due to OCR issues. All papers were confer-
tific papers and from the author intention structure ence papers 6-8 pages in length. The authors re-
of a Related Work. Overall, whilst aspects may be port annotator agreement, based on Cohen Kappa
relevant in general, these methods would not facil- (Cohen, 1960) at 0.77, which increased to 0.85 fol-
itate the kind of content feedback that would help lowing a round of discussion.
a writer with Related Work.
180
Quality Label Description
Background BG-NE Description of the state of the field, describing/listing known methods or
common knowledge. No evidence i.e. citation is not included
BG-EP As above but evidence provided i.e.citation included
BG-EV Positive, shortcoming, problem or gap in the field
Cited Works CW-D Describes cited work, this could be specific details, or very high level details or
nothing more than a reference for further information
A-CW-U Author’s work uses/builds/similar to a cited work
CW-EV Positive, shortcoming, problem or gap about the cited work
Gap and Contribution A-D Author describes their work with no linguistic marking to other’s work or
being different
A-GAP Author specifically says they address a gap or highlights the novelty of their
work
A-CW-D Author’s highlights how their work is different to cited work
TXT Sentence provides information about what will be discussed in the next section
181
will help to discriminate between background adapt it manually using patterns we observe
sentences with citation evidence and citation in Related Work. For example, one aspect
description sentences. We also take a count we consider is contrasts that occur at the be-
of type 1 and 2 citations and a separate count ginning of a sentence and those that happen
of type 3 citations. mid sentence, creating lexical expressions to
capture these. We also produce finer grained
• N-grams Work based on a much larger cor- lexicon patterns for discourse connectives
pus than ours show that n-grams contribute as these are indicative of a continuation
significantly to the performance of their clas- sentence. Within those patterns we include
sifier. Liakata et al. (2012) show a 40% citation types and co-references as described
contribution and Cotos and Pendar (2016) above.
work is mainly based on n-gram features
of 650 Introductions. While our corpus is • Sentiment We use our adapted version of
much smaller (Related Work from 94 arti- Teufel’s list to identify positive and negative
cles), we nevertheless include binary values words (e.g advantageous - positive adjective,
for bigrams and trigrams occurring with a fre- inaccurate - negative adjective ). In addition,
quency of ≥5. We do not remove stop words. we use the 82 polar phrases found in (Athar,
2014). We parse each sentence and count the
• Co-referencing Features Often discussion positive and negative words.
about a work or the author’s work will be
carried out over several sentences. The later • Counts Counts of sentence words, nouns, ad-
sentences can have co-references to the orig- verbs, discourse connectives.
inal citation such as ‘this paper’ ‘this model’
However, as Teufel (1999) shows, determin- • Subject We assign a sentence subject label
ing what she calls agents (e.g US AGENT - before assigning a sentence label to decide if
‘our paper’), these co-reference phrases can the sentence is about a citation, background
be ambiguous. For example does ‘this paper’ or field information, author’s work, or a com-
mean the previously cited paper or it is ref- bination of author’s work and cited work.
erencing the author’s work. We take a differ- This subject feature is based on rules of sen-
ent approach and use the annotations in our tence and previous sentence features e.g our
corpus for (i) references to the authors own finer grained approach to discourse connec-
work, (ii) cited work. In addition, we man- tives in conjunction with co-reference mark-
ually mark co-reference to multiple works in ers help us to understand subject.
background sentences e.g ’previous work has
5 Experimental Setup and Evaluation
been done in the area of ..’ and co-reference
to previously cited work e.g. ’these previ- 5.1 Baseline
ously mentioned works above’
We provide two baselines, one with n-gram fea-
• Verb Features We use part of speech (POS) tures only and one with all features based on the
tags to identify verbs, treating the six possi- majority class.
ble VB tags (VB, VBD, VBG, VPN, VBP, 5.2 Evaluation
VBZ) as binary features of being present or
not in a sentence. Having substituted the co- Our work is similar to other automated classifica-
references, described above, in a sentence we tions but not directly comparable as schemas and
then parse for dependency extracting subject experimental settings differ. Our results are more
and object verb pairs in every sentence. comparable to the works of (Teufel, 1999; Jur-
gens et al., 2018; Teufel and Kan, 2011) as we
• Linguistic Patterns Teufel (1999) makes use the same pattern list from (Teufel, 1999) as
available a list of patterns containing a starting point. These works use Naive Bayes,
cue phrases/words, lexical categories, Random Forest and Maximum Entropy as classi-
constrained by PoS tags, developed on fier methods. We report their published Macro F1
computational linguistic literature. Like scores, range of F1 scores for labels and the num-
(Jurgens et al., 2018) we use this list and ber of labels in the schema for comparison (Table
182
Features Prec Recall F1 Acc% 6 Results
ALL .69 (.005) .7 (.004) .7 (.005) 70 (.48)
Cotos(2016) .686 .549 .61 72.9% 6.1 Classifier Performance
Teufel(2011) .478 .376 .4142 66.8%
We compare our results to those mentioned in Sec-
Table 3: Classifier Performance, Mean scores after tion 5, Table 4. Comparing F1 scores overall, we
10 iterations, Variance in brackets outperform the other systems by a reasonable mar-
gin. In addition, the range of F1 scores for our la-
bels are also similar to other systems. We outper-
System Macro F1/Range Label No. form the work of Cotos and Pendar (2016), who
(Teufel and Kan, 2011) 0.41 (0.19-0.81) 8
(Jurgens et al., 2018) 0.53 6 looks at classification for Introduction feedback,
(Teufel, 1999) 0.68 (0.28-0.86) 12 despite their work being based on a bigger anno-
(Cotos and Pendar, 2016) 0.61 (0.36-0.85) 17
tated corpus. We significantly outperform both our
Our Work
-all feat 0.70 (.25 -0.88) 10 baselines of n-grams and majority class. We re-
- no novel feat 0.54 (.15-.87) run our classification (no novel feat) removing the
Baseline
Ngram(B,T) 0.39 (.02-.68)
manual additions we added to the original pattern
Majority 0.57 (-) list of Teufel (1999), removing co-references and
subject labels. This results in lower performance,
Table 4: Classifier Comparison, * significant 0.01 significant (p <0.01) than our all features and our
majority baseline.
183
Features BG-EV BG-NE BG-EP CW-EV CW-D A-CW-U A-D TXT A-CW-D A-GAP
ALL .39 .72 .73 .53 .84 .48 .47 .88 .63 .25
Feat-(LOO)
-subject .33 .62 .71 .51 .81 .49 .41 .85 .64 .22
-n-grams .33 .70 .70 .53 .84 (.50) .39 .83 .62 .25
-verbtense .35 .71 .72 .51 .84 .48 .46 .88 .66 .32
-sentiment .34 .71 .71 .50 .84 .46 .43 .67 .61 (.28)
-counts (.40) .72 .73 .52 .84 (.50) .46 .87 (.64) .26
-Tot cit count .38 .71 (.74) (.54) (.85) .49 (.48) .88 .64 .26
-paragraph (.40) .71 .73 (.54) .84 .49 .47 .87 .62 .22
Features BG-EV BG-NE BG-EP CW-EV CW-D A-CW-U A-D TXT A-DIFF A-GAP
ALL .39 .72 .73 .53 .84 .48 .47 .88 .63 .25
Feat-(SF)
-patterns .30 .54 (.74) .41 .77 (.57) (.48) .80 (.65) .26
-subject - .58 - - .80 - .45 .75 .46 -
-sub+patt+dep .31 .72 .73 .47 .83 (.55) .46 .84 .63 (.27)
-n-grams .11 .31 .21 .24 .62 .23 .04 .68 .39 .02
tributes here. Neither of our evaluations labels by (Liakata et al., 2012) to contribute to sentence
(BG-EV, CW-EV) perform as well as expected. classification. Our gold subject label was deter-
These two labels are merged from the annotation mined from the annotated label. We see that de-
schema, positive and shortcoming/problem into termining this label accurately has an almost 15%
one evaluation label. These original labels are increase in the performance of our classifier and
both quite different linguistically and we speculate an increase in F1 score for all label categories. In-
that this might prove difficult for the classifier. cluding a previous label also increases the classi-
Total citation counts and counts of adverbs, fier performance, but this increase was not signifi-
words, nouns and discourse connectives seem to cant.
actually make the performance of the classifier
worse on many of the labels, although not sig- 7 Discussion and Conclusions
nificantly so. There is an overlap in total citation
We use a manually annotated data-set designed
counts with the count of our citation types perhaps
for support of writer feedback of a Related Work
indicating this feature could be omitted.
section and show that we can outperform exist-
We note that the features we add to the pat- ing similar methods. We describe our feature
tern list, dependencies and subject label show very set, proposing some novel features such as co-
close to performance of the All features model. reference specific to Related Works, citation types
We observe better performance on the rare label and include these in our adapted pattern set. We
author gap (A-GAP) with just these features alone. show the introduction of our features over and
Most categories are negatively impacted by the above the original pattern features (Teufel, 1999)
removal of the subject label with the exception was a contributing factor to the performance of our
of author uses/build/similar to cited work (A-CW- classifier. This highlights the importance of under-
U) and authors work differs from cited work (A- standing the author intentions of interest and look-
CW-D). As a single feature we see that sub- ing for patterns that are specific to these. This ma-
ject is important to the classifier performance and jor contribution of patterns though is also a limita-
contributes to several of the labels - background tion in that this is built on a study of patterns that
with no evidence (BG-NE), cited work description occur within the computational linguistic domain
(CW-D), author description (A-D) and author and and how it would perform in another domain re-
cited work differ (A-CW-D). Leaving out subject mains to be investigated. Recent work of (Asadi
label was the only feature to cause a drop in clas- et al., 2019) show that using WordNet roots for
sifier performance that was significant. In Table 6 Nouns, e.g where nouns are taken to their more
and Table 7 experiments from using a gold subject general form (e.g., mm and cm become quantity,
label and using a history feature of previous label is a useful feature for author intention identifica-
are presented. History label was previously shown tion. The application of WordNet is one possible
184
Features BG-EV BG-NE BG-EP CW-EV CW-D A-CW-U A-D TXT A-D A-G
ALL .39 .72 .73 .53 .84 .48 .47 .88 .63 .25
+Plabel .50 .60 .70 .60 .86 .51 .46 .63 .61 .27
+GoldSubject .61 .86 .88 .67 .94 .68 .72 1 .88 .4
Table 6: Mean F-Measures for Labels All features and All with Gold Subject and Previous label
Features Prec Recall F1 Acc% Sophie Kitto Kirsty Knight Simon Buckingham
ALL .69 .7 .7 70 Shum Simon Abel. 2018. Designing personalised,
+Plabel .71 .72 .71 71.72
+GoldSubject .84 .85 .84 84.6
automated feedback to develop students research
writing skills. In Proceedings of 2018 Australasian
Table 7: Classifier Performance, Mean scores after Society for Computers in Learning in Tertiary Edu-
cation. pages 15–24.
10 iterations
Claire Aitchison, Janice Catterall, Pauline Ross, and
Shelley Burgin. 2012. ’Tough love and tears’: learn-
avenue that may assist in transitioning our pattern ing doctoral writing in the sciences. Higher Ed-
list to another domain. ucation Research & Development 31(4):435–447.
In future work, we intend to investigate aug- https://doi.org/10.1080/07294360.2011.559195.
menting our pattern set further. Jurgens et al. Mandya A. Angrosh, Stephen Cranefield, and Nigel
(2018) implement a bootstrapping pattern that Stanger. 2010. Context identification of sen-
identifies over four times the manually curated tences in related work sections using a condi-
patterns, identifying new patterns that apply in a tional random field: Towards intelligent digi-
tal libraries. In Proceedings of the 10th An-
citing sentence, the preceding or following sen-
nual Joint Conference on Digital Libraries. ACM,
tence. Bootstrapping to expand seed cue phrases New York, NY, USA, JCDL ’10, pages 293–302.
based on rhetorical relations (Abdalla and Teufel, https://doi.org/10.1145/1816123.1816168.
2006) has also been successful. Incorporating
Mandya A. Angrosh, Stephen Cranefield, and Nigel
more information from a preceding or following
Stanger. 2012. A citation centric annotation scheme
sentence we believe could help classify sentences for scientific articles. In Proceedings of the Aus-
where there is no linguistic clue as to the subject tralasian Language Technology Association Work-
e.g. those that carry on describing a cited work shop 2012. Dunedin, New Zealand, pages 5–14.
but their is no co-reference to signal the subject. https://www.aclweb.org/anthology/U12-1003.
Understanding sentence subject is important, cur- Laurence Anthony and George V. Lashkia.
rently it contributes to the classifier performance 2003. Mover: A Machine Learning Tool
but we show an almost 15% increase in perfor- to Assist in the Reading and Writing of
Technical Papers. Professional Communi-
mance that could occur using a gold sentence sub-
cation, IEEE Transactions on 46:185 – 193.
ject label. Having a way to improve our cur- https://doi.org/10.1109/TPC.2003.816789.
rent implementation of sentence subject assign-
ment would be beneficial. Nasrin Asadi, Kambiz Badie, and Maryam Tayefeh
Mahmoudi. 2019. Automatic zone identifi-
Our overall intention for this work is to support cation in scientific papers via fusion tech-
writer feedback and so we intend to investigate niques. Scientometrics 119(2):845–862.
how well our current level of automatic recogni- https://doi.org/10.1007/s11192-019-03060-9.
tion of author intentions can support feedback and
Awais Athar. 2014. Sentiment analysis of scien-
how useful this is to novice writers. tific citations. Technical Report UCAM-CL-TR-
856, University of Cambridge, Computer Labora-
tory. https://www.cl.cam.ac.uk/techreports/UCAM-
References CL-TR-856.pdf.
Rashid M. Abdalla and Simone Teufel. 2006. A Steven Bird, Robert Dale, Bonnie Dorr, Bryan Gibson,
bootstrapping approach to unsupervised detec- Mark Joseph, Min-Yen Kan, Dongwon Lee, Brett
tion of cue phrase variants. In Proceedings Powley, Dragomir Radev, and Yee Fan Tan. 2008.
of the 21st International Conference on Compu- The ACL Anthology Reference Corpus: A Refer-
tational Linguistics and the 44th Annual Meet- ence Dataset for Bibliographic Research in Compu-
ing of the Association for Computational Linguis- tational Linguistics. In LREC 2008.
tics. Association for Computational Linguistics,
Stroudsburg, PA, USA, ACL-44, pages 921–928. Leo Breiman. 2001. Random forests. Machine learn-
https://doi.org/10.3115/1220175.1220291. ing 45(1):5–32.
185
J Burstein, D Marcu, and K Knight. 2003. Finding the Workshop on Biomedical Natural Language Pro-
WRITE stuff: Automatic identification of discourse cessing. Association for Computational Linguistics,
structure in student essays. IEEE Intelligent Systems pages 99–107.
.
Ken Hyland. 2008. As can be seen: Lexical bundles
Jill Burstein, Martin Chodorow, and Claudia Leacock. and disciplinary variation. English for Specific Pur-
2004. Automated Essay Evaluation: The Criterion poses 27(1):4–21.
Online Writing Service. AI Magazine 25:27–36.
Ken Hyland. 2012. Bundles in academic discourse.
Arlene J Casey, Bonnie Webber, and Dorota Głowacka. Annual Review of Applied Linguistics 32:150–169.
2019. A Framework for Annotating ‘Related https://doi.org/10.1017/S0267190512000037.
Works’, to Support Feedback to Novice Writers. In
LAW ’13: Proceedings of the Linguistic Annotation David Jurgens, Srijan Kumar, Raine Hoover, Dan
Workshop. Association for Computational Linguis- McFarland, and Dan Jurafsky. 2018. Measuring
tics, Stroudsburg, PA, USA. the Evolution of a Scientific Field through Citation
Frames. Transactions of the Association of Compu-
Chih-Chung Chang and Chih-Jen Lin. 2011. LIB- tational Linguistics 6:391–406.
SVM: A library for support vector machines. ACM
Transactions on Intelligent Systems and Technology Barbara Kamler and Pat Thomson. 2006.
2:27:1–27:27. Helping doctoral students write: Ped-
agogies for supervision. Routledge.
Arman Cohan and Nazli Goharian. 2015. Scientific https://doi.org/10.4324/9780203969816.
article summarization using citation-context and ar-
Maria Liakata, Shyamasree Saha, Simon Dobnik,
ticle’s discourse structure. In Proceedings of the
Colin Batchelor, and Dietrich Rebholz-Schuhmann.
2015 Conference on Empirical Methods in Natu-
2012. Automatic recognition of conceptualization
ral Language Processing. Association for Computa-
zones in scientific articles and two life science ap-
tional Linguistics, Lisbon, Portugal, pages 390–400.
plications. Bioinformatics 28(7):991–1000.
https://doi.org/10.18653/v1/D15-1045.
Joseph A. Maxwell. 2006. Literature Reviews
Jacob Cohen. 1960. A coefficient of agreement for of, and for, Educational Research: A Commen-
nominal scales. Educational and psychological tary on Boote and Beile’s “Scholars Before Re-
measurement 20(1):37–46. searchers”. Educational Researcher 35(9):28–31.
https://doi.org/10.3102/0013189X035009028.
Elena Cotos and Nick Pendar. 2016. Discourse classi-
fication into rhetorical functions for awe feedback. Yoko Mizuta and Nigel Collier. 2004. Zone iden-
calico journal 33(1):92–116. tification in biology articles as a basis for infor-
mation extraction. In Proceedings of the interna-
Valéria D Feltrim, Simone Teufel, Maria Graças V das tional joint workshop on natural language process-
Nunes, and Sandra M Aluı́sio. 2006. Argumenta- ing in biomedicine and its applications. Associa-
tive zoning applied to critiquing novices scientific tion for Computational Linguistics, pages 29–35.
abstracts. In Computing Attitude and Affect in Text: https://www.aclweb.org/anthology/W04-1205.
Theory and Applications, Springer, pages 233–246.
Claude Nadeau and Yoshua Bengio. 1999. Inference
Beatriz Fisas, Horacio Saggion, and Francesco Ron- for the Generalization Error. In Proceedings
zano. 2015. On the discoursive structure of of the 12th International Conference on Neu-
computer graphics research papers. In Pro- ral Information Processing Systems. MIT Press,
ceedings of the 9th linguistic annotation work- Cambridge, MA, USA, NIPS’99, pages 307–313.
shop. Association for Computational Linguis- http://dl.acm.org/citation.cfm?id=3009657.3009701.
tics, Denver, Colorado, USA, pages 42–51.
https://doi.org/10.3115/v1/W15-1605. Charles Oppenheim and Susan P Renn. 1978. Highly
cited old papers and the reasons why they continue
Debanjan Ghosh, Aquila Khanam, Yubo Han, and to be cited. Journal of the American Society for In-
Smaranda Muresan. 2016. Coarse-grained argu- formation Science 29(5):225–231.
mentation features for scoring persuasive essays.
In Proceedings of the 54th Annual Meeting of Brian Paltridge and Sue Starfield. 2007. Thesis and
the Association for Computational Linguistics (Vol- Dissertation Writing in a Second Language. Rout-
ume 2: Short Papers). Association for Computa- ledge. https://doi.org/10.4324/9780203960813.
tional Linguistics, Berlin, Germany, pages 549–554.
https://doi.org/10.18653/v1/P16-2089. Ulrich Schäfer, Christian Spurk, and Jörg Stef-
fen. 2012. A fully coreference-annotated
Yufan Guo, Anna Korhonen, Maria Liakata, corpus of scholarly papers from the ACL an-
Ilona Silins Karolinska, Lin Sun, and Ulla Ste- thology. In Proceedings of COLING 2012:
nius. 2010. Identifying the information structure Posters. The COLING 2012 Organizing Com-
of scientific abstracts: an investigation of three mittee, Mumbai, India, pages 1059–1070.
different schemes. In Proceedings of the 2010 https://www.aclweb.org/anthology/C12-2103.
186
Hagit Shatkay, Fengxia Pan, Andrey Rzhetsky, and
W John Wilbur. 2008. Multi-dimensional clas-
sification of biomedical text: toward automated,
practical provision of high-utility text to di-
verse users. Bioinformatics 24(18):2086–2093.
https://doi.org/10.1093/bioinformatics/btn381.
Yi Song, Michael Heilman, Beata Beigman Kle-
banov, and Paul Deane. 2014. Applying ar-
gumentation schemes for essay scoring. In
Proceedings of the First Workshop on Argu-
mentation Mining. Association for Computational
Linguistics, Baltimore, Maryland, pages 69–78.
https://doi.org/10.3115/v1/W14-2110.
Marc Sumner, Eibe Frank, and Mark A Hall.
2005. Speeding up logistic model tree in-
duction. PKDD LNCS 3721:675–683.
https://hdl.handle.net/10289/1446.
John M Swales. 1990. Genre Analysis: English in aca-
demic and research settings. Cambridge University
Press.
Simone Teufel. 1999. Argumentative zoning: Infor-
mation extraction from scientific text. Ph.D. thesis,
University of Edinburgh.
Simone Teufel and Minyen Kan. 2011. Robust Ar-
gumentative Zoning for Sensemaking in Scholarly
Documents. In In Advanced Language Technologies
for Digital Libraries. Springer, pages 154–170.
Simone Teufel and Marc Moens. 2002. Sum-
marizing scientific articles: Experiments
with relevance and rhetorical status. Com-
putational Linguistics 28(4):409–445.
https://doi.org/10.1162/089120102762671936.
Simone Teufel, Advaith Siddharthan, and Colin Batch-
elor. 2009. Towards domain-independent argu-
mentative zoning: Evidence from chemistry and
computational linguistics. In Proceedings of the
2009 Conference on Empirical Methods in Natu-
ral Language Processing. Association for Compu-
tational Linguistics, Singapore, pages 1493–1502.
https://www.aclweb.org/anthology/D09-1155.
Simone Teufel, Advaith Siddharthan, and Dan Tid-
har. 2006. Automatic classification of citation
function. In Proceedings of the 2006 Con-
ference on Empirical Methods in Natural Lan-
guage Processing. Association for Computational
Linguistics, Sydney, Australia, pages 103–110.
https://www.aclweb.org/anthology/W06-1613.
Paul Thompson and Chris Tribble. 2001. Looking at
citations: Using corpora in english for academic pur-
poses. Language Learning Technology 5(3):91 –
105.
Stephen E Toulmin. 2003. The Uses of Argument.
Cambridge University Press.
Melvin Weinstock. 1971. Citation indexes. encyclo-
pedia of library and information science. volume 5.
eds. a. kent & h. lancour.
187
Sparse Victory – A Large Scale Systematic Comparison of
Count-Based and Prediction-Based Vectorizers for Text Classification
188
Proceedings of Recent Advances in Natural Language Processing, pages 188–197,
Varna, Bulgaria, Sep 2–4 2019.
(Peters et. al, 2018) use masked language modeling the present study. 4. Our benchmark contains 73
and textual entailment tasks to generate context- datasets in comparison to GLUE which has only
sensitive character-level representations. In the 10, thereby making it more diverse and
same vein, Flair embeddings (Akbik et. al, 2018) challenging. Finally, we have made our source
leverage the internal states of a trained character code++, datasets** (including train and test splits),
language model to produce a novel type of word result files and all other necessary information
embedding which the authors allude to as publicly available so that, researchers can
contextual string embedding. Moving from reproduce our results and further the progress in the
individual word representations to document and field. While not central to the study we have also
phrase level representation, we observe a less carried out an interpretation analysis on the
spectacular retinue of research work. Notable predictions of these vectorizers by using model
among these are Skip-Thought (Kiros et al, 2015) agnostic, locally interpretable explanations
and InferSent (Conneau et al, 2017). Recently (Riberio et. al, 2016), the results are not included
proposed Universal Sentence Encoder (Cer et. al, in the paper, however interested readers are
2018) which uses multi-task transfer learning encouraged to refer to Appendix A for more details.
based on the transformer architecture (Vaswani et. The paper is organized as follows – Section 1
al, 2017) to deliver promising results on several introduces the paper and gives an overview of the
natural language inference tasks. prior research work. Section-2 provides details of
In light of these prolific advances made in the our datasets and the models used. Section-3
field of text vectorization, it becomes necessary to elucidates the approach we have taken for our
evaluate the different algorithms on downstream experiments. Section-4 presents the results of our
tasks and juxtapose their performance with the experiments, including an extensive analysis.
traditional non-neural counterparts. Existing Section-5 concludes the paper and provides useful
evaluations (Baroni et. al, 2014) have only focused future research directions.
on the semantic aspect of these representations
while ignoring tasks like text classification. Even 2 Data & Model Details
when comparisons are made on benchmarks
similar to the GLUE benchmark (Wang et. al, We have collected the datasets from a variety of
2018), they are almost always made with state of online sites like Kaggle, Crowdflower (now known
art deep neural network based classifiers, the non- as FigureEight), DataTurks, UCI repository and
neural classifiers like Random Forests, SVMs and others. They have been grouped into 8 categories
GradientBoosting are left out. To the best of our for ease of analysis, these are – emotion, sentiment,
knowledge there is no existing research which reviews, medical, general classification, news,
comprehensively evaluates the performance of spam-fake-hate-ironic and other. The general
modern text vectorizers on text classification tasks, classification category set includes things like
it is this research gap which we want to bridge in gender classification, website categorization
the present study. The main contributions of the weather and disaster detection from tweets etc. The
paper are the following – 1. We have collected, other category set includes a set of language tasks
curated and standardized a set of 73 different like natural language inference, duplicate question
datasets which cover all aspects of text detection, objectivity-subjectivity analysis which
classification in particular and language modeling have been recast in a classification framework to
in general. 2. We have extensively analyzed the promote uniformity. Details about the metadata of
performance of neural vectorizers like Word2Vec, each category is present in table 1. All the end tasks
GloVe, FastText, ELMo and Flair on these datasets are different text categorizations ranging from
across many dimensions like dataset-size, class classification of sentiments, emotions, news
imbalance, classification metrics and juxtaposed it articles, reviews, gender, hate speech detection etc.
with their count-based non-neural alternatives like All the datasets have been standardized in a
Feature Hashing and Tf-Idf. 3. We have also common format, this format contains only two
reported results on the performance of traditional fields one for the text data other for the class label.
ML classifiers, since our main aim is to study the Refer to Appendix B for necessary details about the
efficacy of vectorization algorithms we haven’t data standardization process. As can be inferred
included any neural network based classifiers in from table 1, all the categories contain more than a
Table 2. Mean Performance Metrics on a category basis for all Vectorizers (dataset size less than 10K)
General
Classification 47.4/41.9/58.9 48.5/43.1/59.9 48.8/42.9/59.6 41.8/37.6/54.7 60.4/56.7/68.5 57.4/52.0/65.1 52.3/46.1/63.1
(6)
Reviews (1) 35.9/24.3/56.6 33.9/24.3/56.6 34.7/24.4/56.6 30.9/23.1/54.9 44.1/33.2/60.9 43.4/29.6/58.7 36.1/25.2/55.4
Spam-Fake-
Ironic- 61.4/51.8/76.7 63.4/53.2/77.6 63.5/53.0/77.4 58.5/47.4/74.8 61.5/54.5/79.0 60.7/51.4/76.9 67.0/54.2/78.3
Hate(4)
News (1) 15.9/9.2/50.5 15.7/9.0/49.8 16.2/9.5/51.7 14.8/9.0/46.6 37.1/29.3/75.4 44.9/36.7/74.6 23.3/16.7/59.2
Table 3. Mean Performance Metrics on a category basis for all Vectorizers (dataset size between 10K and 50K)
Table 4. Mean Performance Metrics on a category basis for all Vectorizers (dataset size greater than 50K)
192
Category Name RandomForest GradientBoost AdaBoost Logit Regression SVM (Linear)
(Pr./Rec./Acc.) (Pr./Rec./Acc.) (Pr./Rec./Acc) (Pr./Rec./Acc.) (Pr./Rec./Acc.)
Sentiment (10) 46.5/40.2/60.8 44.7/39.3/60.6 38.4/37.4/57.7 41.1/38.4/60.9 42.7/39.4/60.7
General
Classification 58.7/51.7/66.0 58.2/52.0/66.2 50.2/45.9/59.3 53.8/50.3/66.4 57.1/51.2/65.8
(8)
Other (5) 61.6/57.1/68.6 60.1/55.6/68.0 58.0/54.8/66.2 54.0/51.8/68.2 58.2/55.0/67.6
Table 5. Mean Performance Metrics on a category basis for all Classifiers (dataset size less than 10K)
General
Classification 54.9/46.3/61.9 51.6/46.8/62.1 43.8/41.7/56.6 49.9/46.2/62.1 53.0/47.5/63.0
(6)
Reviews (1) 39.5/25.8/56.9 36.9/25.7/57.2 35.2/26.8/56.6 35.4/26.7/58.0 38.8/27.4/58.1
Table 6. Mean Performance Metrics on a category basis for all Classifiers (dataset size between 10K and 50K)
General
Classification 41.5/34.2/51.3 39.5/32.8/51.9 38.5/31.4/45.9 37.0/33.5/50.9 23.2/22.2/32.5
(3)
Other (2) 55.0/52.0/61.3 53.9/47.4/58.8 48.8/45.2/54.7 46.2/43.7/53.6 48.4/44.4/55.6
Table 7. Mean Performance Metrics on a category basis for all Classifiers (dataset size greater than 50K)
193
Category GloVe FastText Word2Vec ELMo Tf-Idf FeatureHash Flair
Name (Pr./Rec./Acc.) (Pr./Rec./Acc.) (Pr./Rec./Acc.) (Pr./Rec./Acc.) (Pr./Rec./Acc.) (Pr./Rec./Acc.) (Pr./Rec./Acc.)
Sentiment (4) 69.0/69.0/69.0 69.7/69.7/69.7 69.4/69.3/69.3 62.1/61.7/61.6 74.9/74.1/74.1 71.7/71.4/71.4 69.0/68.2/68.2
General
Classification 59.4/57.1/65.0 60.5/58.5/66.1 59.6/57.9/65.6 51.4/52.0/60.8 63.4/60.9/67.3 59.2/56.5/63.7 61.2/61.2/67.5
(7)
Other (5) 65.1/61.0/71.7 65.5/61.1/71.7 64.8/61.2/71.8 57.6/56.4/69.8 65.2/58.4/71.8 62.2/57.3/71.5 64.1/57.9/72.6
Reviews (1) 80.3/54.8/91.9 64.9/54.2/91.4 79.2/54.3/91.6 69.0/54.5/91.7 82.6/58.1/92.0 73.8/63.1/91.5 84.4/59.5/92.1
Spam-Fake-
Ironic-Hate 75.3/69.1/80.2 76.7/70.7/80.9 76.5/70.4/80.8 69.6/63.2/77.7 80.8/75.0/83.9 76.2/70.8/80.9 77.2/72.8/81.7
(8)
Medical (2) 72.2/65.9/84.6 69.1/65.3/87.0 72.5/65.6/86.9 63.4/62.0/85.9 69.4/64.4/83.3 71.1/64.0/85.4 64.5/58.9/93.7
News (3) 64.4/64.1/64.5 62.9/62.6/63.2 63.5/63.3/63.7 43.3/41.6/42.4 80.9/80.1/80.3 71.7/71.5/71.6 83.8/83.6/83.8
Table 8. Mean Performance Metrics on a category basis for all Vectorizers (for imbalance measure between 0.0 and 1.03)
Table 9. Mean Performance Metrics on a category basis for all Vectorizers (for imbalance measure between 1.03 and 4.46)
General
Classification 32.4/19.4/52.1 32.8/19.7/51.8 35.9/20.6/53.1 27.8/17.8/50.8 55.7/47.5/67.7 52.6/40.4/63.5 40.7/23.4/53.8
(3)
Reviews (4) 30.0/19.3/56.2 29.3/21.0/60.3 27.6/20.8/60.1 24.5/20.3/59.6 36.4/27.8/65.8 33.7/24.7/63.3 31.7/25.0/61.4
Spam-Fake-
Ironic-Hate 47.1/31.7/89.3 53.0/32.5/90.2 52.6/33.4/90.1 41.0/28.0/90.6 32.5/16.5/85.6 41.0/19.8/87.0 58.8/34.9/91.1
(1)
Medical (3) 35.6/31.0/64.5 34.2/31.3/65.4 34.6/30.5/65.4 33.4/28.4/62.7 46.5/38.4/68.5 36.8/33.1/63.9 36.6/25.3/66.0
Table 10. Mean Performance Metrics on a category basis for all Vectorizers (for imbalance measure between 4.46 and ∞)
194
Figure 6: Heatmap(accuracy) Figure 7: Heatmap(accuracy) Figure 8: Performance (f1-score) Figure 9: Performance (f1-score)
for a news classification dataset for a inference dataset for imbalance range [1.03, 4.46) for imbalance range [4.46, ∞)
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Alex Wang, Amanpreet Singh, Julian Michael, Felix
Corrado, and Jeff Dean. "Distributed Hill, Omer Levy, and Samuel R. Bowman. "Glue: A
representations of words and phrases and their multi-task benchmark and analysis platform for
compositionality." In Advances in neural natural language understanding." arXiv preprint
information processing systems, pp. 3111-3119. arXiv:1804.07461 (2018).
2013. Marco Tulio Ribeiro, Sameer Singh, and Carlos
Jeffrey Pennington, Richard Socher, and Christopher Guestrin. "Why should i trust you?: Explaining the
Manning. "Glove: Global vectors for word predictions of any classifier." In Proceedings of the
representation." In Proceedings of the 2014 22nd ACM SIGKDD international conference on
conference on empirical methods in natural knowledge discovery and data mining, pp. 1135-
1144. ACM, 2016.
196
Austin Appleby. Murmurhash3 64-bit finalizer.
Version 19/02/15. https://code. google.
com/p/smhasher/wiki/MurmurHash3.
Jeremy Howard and Sebastian Ruder. "Universal
Language Model Fine-tuning for Text
Classification." In Proceedings of the 56th Annual
Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), vol. 1, pp.
328-339. 2018.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. "Bert: Pre-training of deep
bidirectional transformers for language
understanding." arXivpreprint
arXiv:1810.04805 (2018).
Bryan McCann, James Bradbury, Caiming Xiong, and
Richard Socher. "Learned in translation:
Contextualized word vectors." In Advances in
Neural Information Processing Systems, pp. 6294-
6305. 2017.
Ajay Patel, Alexander Sands, Chris Callison-Burch,
and Marianna Apidianaki. "Magnitude: A Fast,
Efficient Universal Vector Embedding Utility
Package." In Proceedings of the 2018 Conference
on Empirical Methods in Natural Language
Processing: System Demonstrations, pp. 120-126.
2018.
197
A Fine-Grained Annotated Multi-Dialectal Arabic Corpus
Anis Charfi Wajdi Zaghouani
Information Systems College of Humanities and Social Sciences
Carnegie Mellon University in Qatar Hamad Bin Khalifa University Qatar
acharfi@andrew.cmu.edu wzaghouani@hbku.edu.qa
198
Proceedings of Recent Advances in Natural Language Processing, pages 198–204,
Varna, Bulgaria, Sep 2–4 2019.
2 Related Work and Iraqi (Alsarsour et al., 2018). The annotation
of this corpus was done through crowd sourcing.
Most research on Arabic NLP has focused on Bouamor et al. (2014) presented a multi-dialectal
Modern Standard Arabic (MSA) (Habash, 2010). parallel corpus with 2,000 sentences translated to
There are many parallel and monolingual anno- MSA, Tunisian, Jordanian, Palestinian and Syr-
tated data collections with syntactic and semantic ian Arabic. Later on, MADAR was developed
information such as the different iterations of Penn as a Multi-dialectal large scale corpus that pro-
Arabic Probanks (Palmer et al., 2008; Zaghouani vides parallel translation for 25 Arabic city di-
et al., 2010, 2012) and treebanks (Maamouri et al., alects (Bouamor et al., 2018). All these efforts on
2010). Based on such resources, various tools creating Dialectal Arabic corpora either targeted
were developed for syntactic parsing and morpho- some specific dialects only or did not provide the
logical analysis (Habash, 2010). necessary annotation for author profiling such as
Even though there are relatively many resources annotation about age and gender.
for MSA, Dialectal Arabic (DA) lags behind in
terms of available resources. There have been 3 Methodology
some limited efforts toward creating resources for
the most popular dialects such as the Egyptian In the following, we report on the methodology
and Levantine dialects which were presented in and process followed for the creation, annotation,
(Habash et al., 2013; Diab and Habash, 2007; and validation of our corpus.
Pasha et al., 2014). For example, Habash et al. 3.1 Corpus Overview
created resources for morphological analysis of
Egyptian dialect (Habash et al., 2013). For their Twitter is an increasingly popular social media
work on machine translation, Zbib et al. (2012) platform among the Arabic speaking people who
created Levantine-English and Egyptian-English tend to frequently use their Arabic dialects when
parallel corpora using crowd sourcing. Khalifa sharing their stories and opinions. Therefore, we
et al. (2018) created a morphologically annotated focused on Twitter to collect the data for our cor-
data corpus of Emirati Dialect. pus. ARAP-Tweet 2.0 was developed in the con-
Khalifa et al. (2016) created a corpus of text of the ARAP research project1 and it includes
100M words covering various Arabic dialects. about 5.3 million Arabic tweets of approximately
Other related projects worth to be mentioned are: 3000 users from the following 15 Arabic speak-
the Egyptian Arabic Treebank (Maamouri et al., ing regions: Qatar, Kuwait, United Arab Emirate
2014); the Levantine Arabic Treebank (Maamouri (UAE), Kingdom of Saudi Arabia (KSA), Oman,
et al., 2006), The Curras Palestinian Arabic anno- Yemen, Iraq, Tunisia, Algeria, Libya, Morocco,
tated corpus with more than 70,000 words of vari- Lebanon-Syria, Palestine-Jordan, Egypt, and Su-
ous genres (Jarrar et al., 2014). dan. To the best of our knowledge, there is no
other corpus that covers as many Arab regions as
Furthermore, AlShargi et al. (2016) created
ours. For every region, we collected the tweets of
a Yemeni (Sanaa Dialect) dataset and also a
198 users that were equally-balanced over gender
Moroccan Arabic corpus, while Al-Twairesh et
and three age groups: under 25, 25 until 34, and
al. (2018) built SUAR, a Najdi Arabic cor-
35 and up. Compared to the first version of our
pus annotated with the morphological analyzer
corpus (Zaghouani and Charfi, 2018a) the number
MADAMIRA (Pasha et al., 2014). Finally, Voss
of users per region doubled in this second version.
et al. (2014) presented a Moroccan Arabic corpus
annotated for code-switching (French, Berber and 3.2 Users
Morrocan Arabic).
We used Twitter as our source for finding the
Moreover, there have been some efforts towards
users. In the beginning, we got a number of ac-
creating Dialectal Arabic corpora by either trans-
counts by the geographical location of tweets and
lating existing corpora to dialects or by crowd
by searching on Twitter using specific words. We
sourcing annotation for data collected through var-
collected the users who posted tweets in a specific
ious sources such as microblogs (e.g., Twitter).
geographical location as defined in the tweet’s ob-
Along these lines, DART was developed as a Twit-
ject. As this information was not available for all
ter based data set of dialectal Arabic covering
five Arabic dialects: Egyptian, Levantine, Gulf, 1
arap.qatar.cmu.edu
199
tweets, we searched for users who had tweets that tives used in his tweets would not have that letter.
include some unique words, which are specific to a The annotators used this Arabic rule to determine
certain Arab region. Table 1 shows some of these the user’s gender when it was not possible to do so
unique words from every region covered by our using the username or the profile photo. It is note-
corpus along with a tweet example for each word. worthy that users for which the gender could not
After that, we found further accounts using the determined were not included in the corpus and
followers of the initially identified users. Once the were replaced by other users for which the gender
users were found, we hired experienced annotators could be determined as explained above.
to manually annotate their age, gender, and dialect.
3.3.2 Age Annotation
The annotators followed well-defined annotation
guidelines, which are based on an extended ver- We annotated the age of the selected Twitter users
sion of the guidelines published in (Zaghouani and using three age groups: under 25, 25 until 34, and
Charfi, 2018b). We continuously monitored the 35 and above. In many cases, we were able to
annotators’ work and updated the guidelines when find the exact ages whereas in some other cases we
needed. For each region, we retrieved the tweets were able to determine the correct age group with-
of the users from their Twitter profiles. The se- out the exact age. The annotators went through the
lected users had to have at least 100 Arabic tweets following steps to determine the age of a user:
and at most 3200 tweets. Moreover, the tweets had Getting the exact age: Several Twitter users in-
to be the original tweets, which means that we did clude their birth year as a part of their usernames.
not include retweets in our corpus. The average For example, in the username Omda1981m the
number of words per tweet is 10. birth year is most likely 1981, which indicates that
this user is 38 years old. Some other users put their
3.3 Annotation date of birth in their Twitter biography or in their
other social media accounts such as Facebook, In-
In the following, we report on the annotation of
stagram, or LinkedIn. Other users put their exact
the users with respect to gender, age and dialect.
age in some of their tweets and we were able to
3.3.1 Gender Annotation search for those tweets using some relevant key-
words in different languages as shown in Table 2.
The gender was manually annotated for users and
Getting the estimated age: There were sev-
the number of users is balanced with respect to
eral cases in which we were not able to deter-
the gender across all regions, which means that
mine the exact age even after following the above-
we have 99 male users and 99 female users for
mentioned steps. In these cases, we opted to deter-
each region. The annotation was done based on
mine the approximate ages, so that we could an-
guidelines and criteria that were used to distin-
notate the user with the correct age group. This
guish male and female users. For every user, the
was done by searching for the user on other social
annotator determined the gender by analyzing the
media networks such as LinkedIn and Facebook.
username, profile picture, and indicative words.
We often found either their exact ages or other age
The annotator checked the username if it is de-
related hints such as the year of their graduation
noting a male or female name. Some users put
from university, which helped us in making an ed-
their real photos in their Twitter profile, which
ucated guess towards the user’s age. Furthermore,
also helped in determining their gender. In some
as a last resort for some few users, two annotators
cases, the username and the profile picture did not
guessed the age separately using the Twitter photo.
give sufficient information to identify their gender.
Then, we used Microsoft’s online tool2 that pre-
Therefore, the annotators were looking for some
dicts the age by analyzing the facial features of a
indicative words in the user’ s tweets. For exam-
user through machine learning. Only users were
ple, the adjectives that describe feminine in Ara-
bic usually end with é (Taa’ Marbootah). There-
included for which the age group guessed by Mi-
crosoft’s age determination tool.
fore, finding words that end with the Arabic let-
(Feeling cold),
ter Taa’ Marbootah such as éK@XQK 3.3.3 Dialect Annotation
.
éKAª
K (Feeling sleepy), or éKAªk
. (Feeling hungry) As mentioned above, we selected the users for our
corpus by searching for tweets with dialect spe-
in the tweets of a user indicates that the writer is a
female. However, if the writer is a male, the adjec- 2
www.how-old.net
200
Region Unique Word User’s Tweet
Lebanon-Syria @YJ
ë ù
ªÓ ©Ê£ úÎK
@YJ
ë ð É«ñ « IÊÔ «
Tunisia AQK . ú
æñ JË@ É ¯B@ ð YË B@ ù®J.K
ÈðX AQK . ¨AJÓ ú溻 ¨@ñK
@ I ¯ X
Iraq éK
@ñë@
B@ ©Ó h. XAÖ ß újJë ñ» @ éK @ñë@ iJm
é<Ë@ð ø @
.
KSA éÊ®m k.
éÊ®mk. éJ
¯ð à@
éJ
¯ ¨ññÖÏ @ ú
Gñ¯Y
Palestine-Jordan Aë ®Ó úæ
@ Aë
éJÓ ¨ðQ AªJ.£
Qatar ú
GAg@ PX@ AÓ . é@ P
YË@ úGAg@ úæJ ªË AK@
.
Kuwait É¢J£ é«AÔ . AK
ÕºJ
® É¢J£
KñºK ðQ®Ó ñËX ú
æK@ XA«
g á
Sudan Èð P Èð P ø
@ ú
Í IK
ñJK
P ÉÒªë AÓ úGA K
UAE H. ñK
AÓ H. ñK
AÓ éJ
«@P l×. ðX éJ
¯@ QËAK
ð YË@ éJ«@P AK éJ
¯@ ñËAK á ÒÊ
Morocco AK. @X ú
¯A Ë@
H. @ñm.Ì '@ IJ
®Ë AK. @X H. AJºËA¯ áK
A¿
@ ÉKA
J» AÖß
X I J»
Algeria ð@ P ø
ñÓ èðQK
YKð èñjÊ ð ð@ P ðYJ
ùÔ« YËðð AK
ñkð AK@ AJ»
Libya èPYîD
ú
ÎJ.K
P ú
¯ ù
ÖÞ
@ èPYîD
Ê«
Egypt
ú
æ¯ñËX
ñ» AK@ ú
æ¯ñËX ú
GñJ.
Oman éJ.® éJ® ùªÊ¢ éK@ Qåk Ð ø XA« áK Q K@
ú
æJ
¯Y .
.
Table 1: Examples of unique words from each dialectal region with tweets that include these words.
Table 2: Examples of keywords from different languages used to determine the users age.
cific words and then retrieving the users of those gramming language with the tweepy.py3 library to
tweets and their followers. In addition, once we interact with the Twitter API and retrieve each
retrieved the user’s tweets, the annotator(s) man- user’s timeline (tweets). Due to restrictions by
ually went through them and checked that at least Twitter, we were only allowed to retrieve at most
half of the tweets are in the user’s respective di- 3,200 of the users most recent Tweets. The num-
alect. When the users’ tweets were not mostly ber of tweets of every user in our corpus ranges
in their originally identified dialect, the annotators from a required minimum of 100 tweets to a max-
were instructed to replace them with other users imum of 3,200 tweets. After collecting the maxi-
from the same dialect, the same gender, and the mum possible number of tweets for each user, we
same age group. filtered them. Specifically, we removed the non-
Arabic tweets, retweets, and short tweets that con-
3.4 Tweets Retrieval tain less than 3 words. Eventually, we kept only
the Twitter accounts who had at least 100 origi-
Twitter offers an Application Programming Inter- nal users tweets that are in Arabic and that have
face (API) that allows developers to interact with
Twitter’s web services. We used the Python pro- 3
http://docs.tweepy.org/en/v3.5.0/index.html
201
three words at least. At the end of this process, curred for example when the age of the user was
we were able to have a well-balanced multi-dialect determined mainly by using their Twitter profile
corpus of 198 users for 15 regions spanning 17 photo.
Arab countries. The corpus is equally-balanced Moreover, the annotators found initially some
with respect to gender and age. difficulties with determining the gender for some
users for instance when the profile did not in-
4 Verification and Evaluation clude a real name nor a profile photo. For the lat-
ter cases, the accounts were replaced by accounts
Our main objective was to produce a multi-
from the same gender and age group for which the
dialectal corpus of tweets with a very good qual-
age could be determined with higher confidence.
ity of annotation so that it can be used to promote
For all accounts, a second round of annotation was
Arabic NLP research including research on author
performed by a different annotator. At the end
profiling. This is a major improvement compared
of this phase, we were able to have two annota-
to the previous version of the corpus. To ensure a
tions from different contributors for each user ac-
good quality annotation, we performed two rounds
count. Eventually, variations were checked and re-
of annotation for each Arab region, which means
solved by a member of the research team together
that each region was annotated by two different
with the annotators by reviewing and discussing
annotators. Next, we report on the verification and
the provided justifications for a certain annotation
evaluation process for the annotation of dialect,
of age, gender, or dialect.
age, and gender. Moreover, we present the eval-
uation results for each annotation category. 4.3 Annotation Evaluation
4.1 Verification of Dialect Annotation In order to evaluate the quality of the annotation,
we used the inter-annotator agreement measures.
The annotation was done by experienced Arab an-
We were able to have two rounds of annotation by
notators who were asked to go through each file of
different annotators for each user and this applies
retrieved users tweets and check the dialect used
for the gender, dialect and age group for the whole
in these tweets. This step was necessary to verify
corpus unlike in the previous version (Zaghouani
the overall dialect of all tweets for every user, and
and Charfi, 2018a) in which the inter-annotator
not only the last tweets in a users timeline. Based
agreement was based on a 10 % subset of the cor-
on this dialect verification, we did the following:
pus data.
Removing the users who post tweets in multi-
We measured the inter-annotator agreement us-
ple dialects: many Arab users do not live in their
ing Cohens Kappa, which is a statistical measure
original countries, and consequently, their dialects
of the degree of agreement for a data point (gen-
are affected by the dialects of the countries they
der, dialect, and age in our case), that was labeled
live in. We opted to drop such users from our cor-
by two annotators, over what would be expected
pus and we replaced them by users from the same
by chance. We obtained substantial results for
gender and age group and which used one dialect
the agreement with average Kappa values of 0.99,
only.
0.92, and 0.88 respectively for the annotation of
Removing the users who have many tweets in
gender, dialect, and age. The exact Kappa values
MSA: some Arab users post few tweets in their
per region are shown in Table 3.
own dialects, and they write their tweets mostly
in MSA. After checking all retrieved tweets of a
5 Challenges
given user, we included in our corpus only those
who have more than half of their tweets written in In the following, we report on some challenges
their dialect. that we faced when developing this corpus:
First, we encountered some difficulties in find-
4.2 Verification of Age and Gender ing user accounts for some age groups for cer-
Annotation tain regions. This was the case for Sudani, Iraqi,
Although the age was determined manually, the and Gulf females whose age is 35 or above. For
corpus included initially some users for which nei- some Arab regions, female Twitter users tend to
ther the exact age nor the age group could be de- hide their real name and also avoid putting their
termined with high confidence. These cases oc- photos on social media because of the local cul-
202
Region Gender Age Dialect As a result, when we retrieved again the users or
Lebanon-Syria 1 0.924 1 their tweets we noticed that some profiles were ei-
Tunisia 1 0.742 1 ther made private or even closed. In such cases, we
Iraq 1 0.696 0.9 were no longer able to access the users profile and
Saudi Arabia 1 0.856 0.9 tweets. We addressed this problem by periodically
Palestine-Jordan 0.989 0.856 0.9 checking the users’ accounts and replacing any ac-
Qatar 0.969 0.901 0.9 counts that were deactivated or made private.
Kuwait 0.979 0.886 0.9
Sudan 1 0.734 0.8
6 Conclusion
UAE 0.979 0.848 0.8 We presented in this paper a significantly
Yemen 1 0.984 1 improved version of a large-scale manually-
Morocco 1 0.954 0.9 annotated and fine-grained multi-dialectal corpus
Algeria 0.959 0.931 0.9 of Arabic tweets, which is compsed of 5 millions
Libya 0.989 0.924 1 dialectal Arabic tweets of about 3000 Twitter users
Egypt 1 0.969 1 from 17 Arab countries. To the best of our knowl-
Oman 0.989 0.931 0.9 edge, our corpus is the most comprehensive di-
Overall 0.99 0.88 0.92 alectal corpus in terms of coverage for so many
Arab dialects and the data volume. Moreover, our
Table 3: Kappa values per region for gender, age, corpus is balanced with respect to dialect, gen-
and dialect annotation. der, and age. The corpus was annotated manually
based on well-defined annotation guidelines and
it was fully evaluated using the inter-annotation
ture and norms. This was an issue for us as we agreement measures.
sometimes depend on the Twitter profile photos to The corpus was already and it can further be
guess the age especially if the username is not a used to promote research on Arabic NLP including
real name. Second, we notice that older Arabs of- author profiling (Rosso et al., 2018), authorship at-
ten write in Modern Standard Arabic (MSA) on tribution, dialect identification, sentiment analysis
social media, and this made the task of finding di- in dialectal Arabic, and bots detection in dialectal
alect users above 35 years much harder. Third, Arabic.
we faced an issue with Maghrebi users who often
tweet in French more than Arabic. Acknowledgments
To address these issues, we put more effort into We thank our dedicated annotators: Hoda Ibrahim,
the user selection and we replaced any users who Safaa Ellaham, Syrine Guediche for their hard
did not fulfill the criteria explained above. work on the corpus annotation. This publication
Fourth, we noticed that some users use more was made possible by NPRP 9-175-1-033 from
than one Arabic dialect to write their tweets on the Qatar National Research Fund (a member of
Twitter. For example, in Arab regions with high Qatar Foundation). The findings achieved herein
immigration rates such as Gulf countries, the di- are solely the responsibility of the authors.
alect of the residents is sometimes affected by
the language of their host country. Consequently,
References
these users tweet in multiple dialects and they
might mix dialects even in one same tweet. Other Faisal Al-Shargi, Aidan Kaplan, Ramy Eskander, Nizar
Habash, and Owen Rambow. 2016. Morphologi-
users had similar issues because their parents cally annotated corpora and morphological analyz-
come from different Arab countries. To address ers for moroccan and sanaani yemeni arabic. In
this issues, all tweets were manually reviewed and 10th Language Resources and Evaluation Confer-
users tweeting in two dialects or more were re- ence (LREC 2016).
placed by mono-dialectal users. Nora Al-Twairesh, Rawan Al-Matham, Nora Madi,
Fifth, we worked on regions sequentially, i.e., Nada Almugren, Al-Hanouf Al-Aljmi, Shahad Al-
shalan, Raghad Alshalan, Nafla Alrumayyan, Shams
we selected a large set of accounts from Twitter Al-Manea, Sumayah Bawazeer, et al. 2018. Suar:
for one region, annotated them, and then retrieved Towards building a corpus for the saudi dialect. Pro-
their tweets. Then, we worked on the next region. cedia computer science 142:72–82.
203
Israa Alsarsour, Esraa Mohamed, Reem Suwaileh, and Hamdy Mubarak and Kareem Darwish. 2014. Using
Tamer Elsayed. 2018. Dart: A large dataset of di- twitter to collect a multi-dialectal corpus of arabic.
alectal arabic tweets. In Proceedings of the Eleventh In Proceedings of the EMNLP 2014 Workshop on
International Conference on Language Resources Arabic Natural Language Processing (ANLP). pages
and Evaluation (LREC-2018). 1–7.
Houda Bouamor, Nizar Habash, and Kemal Oflazer. Martha Palmer, Olga Babko-Malaya, Ann Bies,
2014. A multidialectal parallel corpus of arabic. In Mona T Diab, Mohamed Maamouri, Aous Man-
LREC. pages 1240–1245. souri, and Wajdi Zaghouani. 2008. A pilot arabic
propbank. In LREC.
Houda Bouamor, Nizar Habash, Mohammad Salameh,
Wajdi Zaghouani, Owen Rambow, Dana Abdul- Arfath Pasha, Mohamed Al-Badrashiny, Mona T Diab,
rahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Ahmed El Kholy, Ramy Eskander, Nizar Habash,
Alexander Erdmann, et al. 2018. The madar arabic Manoj Pooleery, Owen Rambow, and Ryan Roth.
dialect corpus and lexicon. In Proceedings of the 2014. Madamira: A fast, comprehensive tool for
Eleventh International Conference on Language Re- morphological analysis and disambiguation of ara-
sources and Evaluation (LREC-2018). bic. In LREC. volume 14, pages 1094–1101.
Mona Diab and Nizar Habash. 2007. Arabic dialect Paolo Rosso, Francisco M. Rangel Pardo, Bilal
processing tutorial. In Proceedings of the Human Ghanem, and Anis Charfi. 2018. ARAP: arabic
Language Technology Conference of the NAACL, author profiling project for cyber-security. Proce-
Companion Volume: Tutorial Abstracts. Association samiento del Lenguaje Natural 61:135–138.
for Computational Linguistics, pages 5–6.
Clare R Voss, Stephen Tratz, Jamal Laoudi, and Dou-
Nizar Habash, Ryan Roth, Owen Rambow, Ramy Es- glas M Briesch. 2014. Finding romanized arabic di-
kander, and Nadi Tomeh. 2013. Morphological alect in code-mixed tweets. In LREC. pages 2249–
analysis and disambiguation for dialectal arabic. In 2253.
Proceedings of the 2013 Conference of the North
American Chapter of the Association for Computa- Wajdi Zaghouani and Anis Charfi. 2018a. Arap-tweet:
tional Linguistics: Human Language Technologies. A large multi-dialect twitter corpus for gender, age
pages 426–432. and language variety identification. arXiv preprint
arXiv:1808.07674 .
Nizar Y Habash. 2010. Introduction to arabic natural
language processing. Synthesis Lectures on Human Wajdi Zaghouani and Anis Charfi. 2018b. Guidelines
Language Technologies 3(1):1–187. and annotation framework for arabic author profil-
ing. arXiv preprint arXiv:1808.07678 .
Mustafa Jarrar, Nizar Habash, Diyam Fuad Akra, and
Nasser Zalmout. 2014. Building a corpus for pales- Wajdi Zaghouani, Mona Diab, Aous Mansouri, Sameer
tinian arabic: a preliminary study . Pradhan, and Martha Palmer. 2010. The revised ara-
bic propbank. In Proceedings of the fourth linguis-
Salam Khalifa, Nizar Habash, Dana Abdulrahim, and tic annotation workshop. Association for Computa-
Sara Hassan. 2016. A large scale corpus of gulf ara- tional Linguistics, pages 222–226.
bic. arXiv preprint arXiv:1609.02960 .
Wajdi Zaghouani, Abdelati Hawwari, and Mona Diab.
Salam Khalifa, Nizar Habash, Fadhl Eryani, Ossama 2012. A pilot propbank annotation for quranic ara-
Obeid, Dana Abdulrahim, and Meera Al Kaabi. bic. In Proceedings of the NAACL-HLT 2012 Work-
2018. A morphologically annotated corpus of emi- shop on Computational Linguistics for Literature.
rati arabic. In Proceedings of the Eleventh Interna- pages 78–83.
tional Conference on Language Resources and Eval-
uation (LREC-2018). Rabih Zbib, Erika Malchiodi, Jacob Devlin, David
Stallard, Spyros Matsoukas, Richard Schwartz, John
Mohamed Maamouri, Ann Bies, Tim Buckwalter, Makhoul, Omar F Zaidan, and Chris Callison-
Mona T Diab, Nizar Habash, Owen Rambow, and Burch. 2012. Machine translation of arabic dialects.
Dalila Tabessi. 2006. Developing and using a pilot In Proceedings of the NAACL 2012. pages 49–59.
dialectal arabic treebank. In LREC. pages 443–448.
Mohamed Maamouri, Ann Bies, Seth Kulick, Michael
Ciul, Nizar Habash, and Ramy Eskander. 2014. De-
veloping an egyptian arabic treebank: Impact of di-
alectal morphology on annotation and tool develop-
ment. In LREC. pages 2348–2354.
Mohamed Maamouri, Ann Bies, Seth Kulick, Wajdi
Zaghouani, David Graff, and Michael Ciul. 2010.
From speech to trees: Applying treebank annotation
to arabic broadcast news. In LREC.
204
Personality-dependent Neural Text Summarization
205
Proceedings of Recent Advances in Natural Language Processing, pages 205–212,
Varna, Bulgaria, Sep 2–4 2019.
reports two experiments comparing the proposed RNN computes a sequence of outputs (y1 , ..., yt )
models against a number of alternatives, and Sec- mapped onto sequences using the following equa-
tion 5 presents final remarks and future work. tion (Sundermeyer et al., 2012):
206
need to pay regard to contextual dependency when generates yj based on the current hidden state sj ,
modelling phrases and sentences. and then updates the hidden state sj+1 based on sj
Bidirectional GRUs (Bi-GRUs) are applied to and yj . Formally, the attention decoder is defined
a wide range of tasks to scan and learn both left- by original equation proposed in Cho et al. (2014):
to-right and right-to-left dependencies, which can
capture complementary types of information from s1 = tanh(W (s) bm )
its inputs. The left and right hidden represen- p(yj = w|x, y1:j−1 ) α exp(U [sj , cj ])
tations produced by GRUs can be linearly com- sj+1 = GRU ([φ(out) (yj ), cj ], sj )
bined (θ) to form a final representation (Goodfel- where i {1, ..., m}, j {1, ..., m} and the context
low et al., 2016): ht = h← →
t θ ht . vector cj , is the result of general attention (Luong
et al., 2015). The matrices W (s) , W (α) , U and
2.2 Attention Mechanism the embedding function φ(out) are decoder param-
Sequence-to-sequence architectures have been eters.
successfully applied to a wide range of tasks, in-
cluding machine translation and natural text gen-
3 Current Work
eration (Cho et al., 2014; Sutskever et al., 2014a) Our basic model is generally inspired from the
and, accordingly, have been subject to a great deal architecture in Cho et al. (2014), with an added
of extensions and improvements. Among these, personality embedding layer. As in many other
the use of more context-aware sequence genera- sequence-to-sequence models with attention, our
tion methods (Cho et al., 2014) and the use of at- model takes as an input a sentence, and produces
tention mechanism to score and select words that as an output a set of words that summarizes the
best describe the intended output are discussed be- given input. The actual rendering of this output as
low. structured text is presently not addressed.
In natural language generation, attention mod- The proposed architecture is illustrated in Fig-
els as introduced in Cho et al. (2014) and ure 1, which is adapted from Cho et al. (2014),
Sutskever et al. (2014a) are intended to generalize and further discussed below.
the text generation task so as to handle sequence In this example, B B B B represent the input
pairs with different sizes of inputs and outputs. sequence from the target sequence Z X, and C
This approach, subsequently called sequence-to- is the personality embedding representation. The
sequence with attention mechanism, applies a five main components of the architecture are as
mapping strategy from a variable-length sentence follows.
to another variable-length sentence. This mapping
strategy is a scoring system over the contextual (A) a bidirectional GRU that maps words to per-
information from the input sequence (Cho et al., sonality types
2014), making a set of attention weights. (B) a word embedding layer
Attention-based models (Sutskever et al.,
2014b; Luong et al., 2015) are sequence-to- (C) a personality embedding layer
sequence networks that employ an encoder to rep-
resent the text utterance and an attention-based de- (D) an attention mechanism
coder that generates the response, one token at a (E) a bidirectional GRU that outputs word encod-
time. More specifically, neural text summarization ings
can be viewed as a sequence-to-sequence prob-
lem (Sutskever et al., 2014a), where a sequence The input bidirectional GRU (A) produces a
of input language tokens x = x1 , ..., xm describ- word-to-personality compositional representation
ing the input text are mapped onto a sequence of of each word. This serves two main purposes:
output language tokens y1 , ..., yn describing the combining the composite sequences of words and
target text output. The encoder is a GRU unit personality information, and combining attention
(Cho et al., 2014) that converts x, ..., xm into a se- weights over sequences in our decoder model.
quence of context-sensitive embeddings b1 , ..., bm . The word embeddings layer (B) produces a typ-
A general-attention decoder generates output to- ical word-level representation of each input word.
kens one at a time. At each time step j, the decoder In the present work, we make use of both random
207
`1 (θ, D(c), D(x, y))
X
=− logP (Y |X, < ki , vi >)
(X,Y ) Dc ∪ Dpr
X
=− logP f r (Y |X)
(X,Y ) Dc
4 Evaluation
We envisaged two experiments on neural text sum-
marization based on the model described in the
previous section. The first experiment aims to
assess whether a general or a dot product atten-
Figure 1: Model architecture tion mechanism is more suitable to the task. The
second experiment focuses on our main research
question, that is, on whether the use of personality
and pre-trained word embeddings. The latter are
information does improve summarization results.
Skip-gram 300 word embeddings taken from Hart-
As in many (or most) sequence-to-sequence ap-
mann et al. (2017).
proaches to text generation, our work focuses on
Word embeddings are complemented with in- the selection of text segments to compose an ab-
duced personality embeddings (C) for each tar- stract summary, but it does not address the actual
get author. The role of this layer is twofold. rendering of the final output text, which would
First, it is intended to learn the probability normally require additional post-processing. Each
P (Y |X, personality), that is, the personality of the two experiments is discussed in turn in the
representation of each author for each word in the following sections, but first we describe the dataset
vocabulary. Second, this layer is also intended taken as their basis.
to decide which profile value should be selected
(from the corpus gold standard annotation) in or- 4.1 Data
der to generate a summary. We make use of the text and caption portions of
The attention mechanism (D) attempts to learn the b5 corpus in Ramos et al. (2018), called b5-text
a general representation from the most important and b5-caption. The corpus conveys 1510 multi-
parts of the input text at each time step. To this and single-sentence image description pairs, all of
end, the experiments described in the next section which labelled with Big Five personality informa-
will consider two score function alternatives: gen- tion about their authors. Table 1 summarizes the
eral attention and dot product. corpus descriptive statistics.
Finally, the output bidirectional GRU (E) com- The corpus was elicited from a crowd sourcing
bines the attention weight representations, and task in which participants were requested to pro-
produces a final encoding for each word. A loss vide both long and short descriptions for 10 stim-
function describe the overall generation probabil- ulus images taken from GAPED, a database of im-
ity, and it is intended to optimize the above param- ages classified by valence and normative signif-
eters. This function is described as follows. icance designed to elicit various reactions (Dan-
208
Table 1: Corpus descriptive statistics.
Data Words Average Types Average
text 84463 559.4 37210 246.4
caption 4896 32.4 4121 27.3
209
Table 3: 10-fold cross validation BLEU scores for Table 4: 10-fold cross validation BLEU scores
text summarization using dot product (sDot) and for text summarization with (sPers) and without
general (sGen) attention. the best result is high- (sBase) personality information. The best result is
lighted. highlighted.
Model BLEU Model BLEU
sGen 13.88 sBase 14.21
sDot 13.63 sPers 14.58
parameters were initialized by sampling from a All optimization, training and other basic proce-
uniform distribution U (−sqrt(3/n), sqrt(3/n)), dures are the same as in the previous experiment.
where n is the parameter dimension. Results are presented in Table 3.
We performed 10-fold cross-validation over our We notice that personality-dependent summa-
corpus data, and we compared the output sum- rization as provided by sPers outperforms standard
maries produced by both models using BLEU2 . summarization (i.e., with no access to personal-
Results are presented in Table 3. ity information) as provided by sBase. Although
From these results, we notice that the attention the difference is once again small (which may be
mechanism based on the general function in sGen explained by the limited size of our dataset), this
outperforms the use of dot function in sDot. Al- outcome offers support to our main research hy-
though the difference is small, the use of a gen- pothesis by illustrating that the use of author per-
eralized network to learn how to align the con- sonality information may improve summarization
textual information is superior to simply concate- accuracy.
nating contextual information obtained from the
global weights. Based on these results, the gen- 4.4 Selected Examples
eral attention strategy will be our choice for the As a means to illustrate the kinds of output that
next experiment. may be produced by our models, Table 5 presents
a number of examples taken from the original cor-
4.3 Experiment 2: Personality-dependent pus summaries, and the corresponding summaries
Summarization obtained from the same input by making use of the
Our second and main experiment assesses the use sBase baseline and by the personality-dependent
of personality information in text summarization. sPers models. For ease of illustration, the exam-
To this end, two models are considered: the full ples are informally grouped into three error cate-
personality-aware model presented in Section 3, gories (small, moderate and large) according to the
hereby called sPers, and a simplified baseline ver- distance between the corpus summaries and their
sion of the same architecture without access to per- sPers counterparts, and are presented in both orig-
sonality information, hereby called sBase. In do- inal (Portuguese) and translated (English) forms.
ing so, our goal is to show that summaries pro-
duced by sPers resemble the human-made texts (as 5 Final Remarks
seen in the corpus) more closely than those pro- This paper addressed the use of Big Five person-
duced by sBase. ality information about the target author to gener-
Both sPers and sBase make use of pre-trained ate personalized summaries in neural sequence-to-
skip-gram 300 word embeddings for the Brazil- sequence text summarization. The model - con-
ian Portuguese language taken from Hartmann sisting of two bidirectional GRUs, word embed-
et al. (2017). Both models also make use of dings and attention mechanism - was evaluated in
encoder/decoder randomized word embedding of two versions, namely, with and without an addi-
size 300, and two encoder/decoder hidden units of tional personality embedding layer. Initial results
size 600 with general attention. suggest that having access to personality informa-
2
We are aware that, although popular in machine transla- tion does lead to more accurate (or human-like)
tion and text generation, BLEU may not be the ideal metrics text summaries.
for the present task (Liu et al., 2011; Song et al., 2013), and
that it may not co-relate well with, e.g., human judgments The use of personality information is of course
(Reiter and Belz, 2009). only one among many possible personalization
210
Table 5: Selected examples taken from the corpus, baseline (sBase) and personality-dependent (sPers)
summarization models, grouped by distance (small, moderate or large) between sPers and the expected
(corpus) summary in original Portuguese (Pt) and translated English (En).
Error Model Summary (Pt) Summary (En)
corpus homem na cerca man by fence
small sBase homem idoso elderly man
sPers homem na cerca man by fence
corpus pessoas pedindo ajuda people asking for help
moderate sBase pessoas esperando people waiting
sPers pessoas aguardam atendimento people waiting for help
corpus menino com um balde de terra boy with a bucket full of soil
large sBase crianca com balde child with bucket
sPers crianca com balde de terra child with bucket full of soil
strategies for text summarization. In particular, Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bah-
we notice that the increasing availability of text danau, and Yoshua Bengio. 2014. On the properties
of neural machine translation: Encoder–decoder ap-
corpora labelled with author demographics in gen-
proaches. In Proceedings of SSST-8, Eighth Work-
eral (e.g., gender, age, education information etc.) shop on Syntax, Semantics and Structure in Statisti-
may in principle support a broad range of speaker- cal Translation. Association for Computational Lin-
dependent summarization models. Thus, as fu- guistics, Doha, Qatar, pages 103–111.
ture work we intend to extend the current approach
Elise S. Dan-Glauser and Klaus R. Scherer. 2011. The
along these lines, and provide additional summa- Geneva affective picture database (GAPED): a new
rization strategies that may represent more signifi- 730-picture database focusing on valence and nor-
cant gains over the standard, fixed-output summa- mative significance. Behavior Research Methods
rization approach. 43(2):468–477.
211
Chang Liu, Daniel Dahlmeier, and Hwee Tou Ng. David E. Rumelhart, Geoffrey Hinton, and Ronald J.
2011. Better evaluation metrics lead to better ma- Williams. 1986. Learning representations by
chine translation. In Proceedings of the Conference back propagating errors. Nature 323:533–536.
on Empirical Methods in Natural Language Pro- https://doi.org/10.1038/323533a0.
cessing. Association for Computational Linguistics,
pages 375–384. Alexander M. Rush, Sumit Chopra, and Jason We-
ston. 2015. A neural attention model for abstrac-
Thang Luong, Hieu Pham, and Christopher D. Man- tive sentence summarization. In Proceedings of the
ning. 2015. Effective approaches to attention-based 2015 Conference on Empirical Methods in Natu-
neural machine translation. In Proceedings of the ral Language Processing. Association for Computa-
2015 Conference on Empirical Methods in Natu- tional Linguistics, Lisbon, Portugal, pages 379–389.
ral Language Processing. Association for Compu- https://doi.org/10.18653/v1/D15-1044.
tational Linguistics, Lisbon, Portugal, pages 1412–
1421. https://doi.org/10.18653/v1/D15-1166. Abigail See, Peter J. Liu, and Christopher D. Manning.
2017. Get to the point: Summarization with pointer-
François Mairesse and Marilyn Walker. 2007. PER- generator networks. In Proceedings of the 55th An-
SONAGE: Personality generation for dialogue. In nual Meeting of the Association for Computational
45th Annual Meeting-Association For Computa- Linguistics (Volume 1: Long Papers). pages 1073–
tional Linguistics. Association for Computational 1083.
Linguistics (ACL), Sheffield, pages 496–503.
Xingyi Song, Trevor Cohn, and Lucia Specia. 2013.
Bleu deconstructed: Designing a better mt evalua-
Tomas Mikolov, Scott Wen-tau, and Geoffrey Zweig.
tion metric. International Journal of Computational
2013. Linguistic regularities in continuous space
Linguistics and Applications 4(2):29–44.
word representations. In Proc. of NAACL-HLT-
2013. Association for Computational Linguistics, Martin Sundermeyer, Ralf Schlüter, and Hermann Ney.
Atlanta, USA, pages 746–751. 2012. Lstm neural networks for language modeling.
In Thirteenth annual conference of the international
Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, speech communication association.
Caglar Gulcehre, and Bing Xiang. 2016. Ab-
stractive text summarization using sequence-to- Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014a.
sequence RNNs and beyond. In Proceedings of The Sequence to sequence learning with neural net-
20th SIGNLL Conference on Computational Natu- works. In Advances in neural information process-
ral Language Learning. Association for Computa- ing systems. pages 3104–3112.
tional Linguistics, Berlin, Germany, pages 280–290.
https://doi.org/10.18653/v1/K16-1028. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014b.
Sequence to sequence learning with neural net-
Kishore Papineni, Salim Roukos, Todd Ward, and Wei- works. In Advances in neural information process-
Jing Zhu. 2002. BLEU: a method for automatic ing systems. pages 3104–3112.
evaluation of machine translation. In Proceddings of
ACL-2002. Association for Computational Linguis- Xiaojun Wan and Jianwu Yang. 2006. Improved affin-
tics, Philadelphia, PA, USA, pages 311–318. ity graph based multi-document summarization. In
Proceedings of the human language technology con-
Ricelli Moreira Silva Ramos, Georges Basile Stavra- ference of the NAACL, Companion volume: Short
cas Neto, Barbara Barbosa Claudino Silva, papers. Association for Computational Linguistics,
Danielle Sampaio Monteiro, Ivandré Paraboni, and pages 181–184.
Rafael Felipe Sandroni Dias. 2018. Building a
corpus for personality-dependent natural language
understanding and generation. In 11th International
Conference on Language Resources and Evaluation
(LREC-2018). ELRA, Miyazaki, Japan, pages
1138–1145.
212
Self-Adaptation for Unsupervised Domain Adaptation
213
Proceedings of Recent Advances in Natural Language Processing, pages 213–222,
Varna, Bulgaria, Sep 2–4 2019.
main in UDA (McClosky et al., 2006; Reichart and learn a projected feature space in the target
Rappoport, 2007; Drury et al., 2011). Specifically, domain where the margin between the op-
the source domain’s labelled instances are used to posite pseudo-labelled nearest neighbours is
initialise the self-training process and during sub- maximised. We project labelled instances in
sequent iterations labels are inferred for the target the source domain and pseudo-labelled in-
domain’s unlabelled instances, which can be used stances in the target domain respectively us-
to train a classifier for the task of interest. ing Sprj and Tprj , and use those projected in-
So far in UDA projection-learning and self- stances to learn a classifier for the target task.
trained approaches have been explored separately.
An interesting research question we ask and an- As an evaluation task, we perform cross-domain
swer positively in this paper is whether can we im- sentiment classification on the Amazon multi-
prove the performance of projection-based meth- domain sentiment dataset (Blitzer et al., 2007).
ods in UDA using self-training? In particular, Although most prior work on UDA have used this
recent work on UDA (Morerio et al., 2018) has dataset as a standard evaluation benchmark, the
shown that minimising the entropy of a classifier evaluations have been limited to the four domains
on its predictions in the source and target domains books, dvds, electronic appliances and kitchen ap-
is equivalent to learning a projection space that pliances. We too report performances on those
maximises the correlation between source and tar- four domains for the ease of comparison against
get instances. Motivated by these developments, prior work. However, to reliably estimate the gen-
we propose Self-Adapt, a method that combines eralisability of the proposed method, we perform
the complementary strengths of projection-based an additional extensive evaluation using 16 other
methods and self-training methods for UDA. domains included in the original version of the
Our proposed method consists of three steps. Amazon multi-domain sentiment dataset.
Results from the cross-domain sentiment clas-
• First, using labelled instances from the
sification reveal several interesting facts. A base-
source domain we learn a projection (Sprj )
line that uses Sprj alone would still outperform a
that maximises the distance between each
baseline that uses a classifier trained using only
source domain labelled instance and its near-
the source domain’s labelled instances on the tar-
est neighbours with opposite labels. Intu-
get domain test instances, without performing any
itively, this process will learn a projected fea-
adaptations. This result shows that it is useful
ture space in the source domain where the
to consider the label distribution available in the
margin between the opposite labelled nearest
source domain to learn a projection even though it
neighbours is maximised, thereby minimis-
might be different to that in the target domain.
ing the risk of misclassifications. We project
On the other hand, training a classifier us-
the source domain’s labelled instances using
ing the pseudo-labelled target domain instances
Sprj for the purpose of training a classifier for
alone, without learning Tprj further improves per-
predicting the target task labels such as posi-
formance. This result shows that pseudo labels in-
tive/negative sentiment in cross-domain sen-
ferred for the target domain unlabelled instances
timent classification or part-of-speech tags in
can be used to overcome the issue of lack of la-
cross-domain part-of-speech tagging.
belled instances in the target domain.
• Second, we use the classifier trained in the Moreover, if we further use the pseudo-labelled
previous step to assign pseudo labels for the instances to learn Tprj , then we see a significant im-
(unlabelled) target domain instances. Differ- provement of performance across all domain pairs,
ent strategies can be used for this label infer- suggesting that UDA can benefit from both projec-
ence process such as selecting instances with tion learning and self-training.
the highest classifier confidence as in self- These experimental results support our claim
training or checking the agreement among that it is beneficial to combine projection-based
multiple classifier as in tri-training. and self-training-based UDA approaches. More-
• Third, we use the pseudo-labelled target do- over, our proposed method outperforms all self-
main instances to learn a projection for the training based domain adaptation methods such as
target domain (Tprj ) following the same pro- tri-training (Zhou and Li, 2005; Søgaard, 2010)
cedure used to learn Sprj . Specifically, we and is competitive against neural domain adapta-
214
tion methods (Louizos et al., 2015; Ganin et al., be sufficient and redundant views. Specifically,
2016; Saito et al., 2017; Ruder and Plank, 2018). in tri-training, as the name implies three separate
classifiers are trained using bootstrapped subsets
2 Related Work of instances sampled from the labelled instances.
If at least two out of the three classifiers agree
Self-training (Yarowsky, 1995) has been adapted upon a label for an unlabelled instance, that la-
to various cross-domain NLP tasks such as docu- bel is then assigned to the unlabelled instance.
ment classification (Raina et al., 2007), POS tag- Søgaard (2010) proposed a variation of tri-training
ging (McClosky et al., 2006; Reichart and Rap- (i.e. tri-training with diversification) that diversi-
poport, 2007) and sentiment classification (Drury fies the sampling process and reduces the num-
et al., 2011). Although different variants of self- ber of additional instances, where they require ex-
training algorithms have been proposed (Abney, actly two out of the three classifiers to agree upon
2007; Yu and Kübler, 2011) a common recipe can a label and the third classifier to disagree. It has
be recognised involving the following three steps: been shown that the classic tri-training algorithm
(a) Initialise the training dataset, L = SL , to the when applied to UDA acts as a strong baseline
labelled data in the source domain, and train a that outperforms even more complex SoTA neu-
classifier for the target task using L, (b) apply ral adaptation methods (Ruder and Plank, 2018).
the classifier trained in step (a) to the unlabelled As later shown in our experiments, the proposed
data in the target domain TU , and append the most Self-Adapt method consistently outperforms self-
confident predictions as identified by the classifier training, tri-training and tri-training with diversifi-
(e.g. higher than a pre-defined confidence thresh- cation across most of the domain pairs considered.
old τ ) to the labelled dataset L, (c) repeat the two
steps (a) and (b) until we cannot append additional Projection-based approaches for UDA learn a
high confidence predictions to L. (possibly lower-dimensional) projection where the
Another popular approach for inferring labels difference between the source and target feature
for the target domain is co-training (Blum and spaces is reduced. For example, Structural Cor-
Mitchell, 1998), where the availability of multi- respondence Learning (SCL) (Blitzer et al., 2006,
ple views of the feature space is assumed. In the 2007) learns a projection using a set of domain
simplest case where there are two views available invariant common features called pivots. Differ-
for the instances, a separate classifier is trained us- ent strategies have been proposed in the litera-
ing the source domain’s labelled instances that in- ture for finding pivots for different tasks such as
volve features from a particular view only. Next, the frequency of a feature in a domain for cross-
the two classifiers are used to predict pseudo la- domain POS tagging (Blitzer et al., 2006; Cui
bels for the target domain unlabelled instances. If et al., 2017a), mutual information (Blitzer et al.,
the two classifiers agree on the label for a particu- 2007) and pointwise mutual information (Bolle-
lar unlabelled instances, then that label is assigned gala et al., 2011, 2015) for cross-domain senti-
to that instance. Co-training has been applied to ment classification. Cui et al. (2017b) proposed
UDA (Yu and Kübler, 2011; Chen et al., 2011), a method for learning the appropriateness of a
where the feature spaces in the source and target feature as a pivot (pivothood) from the data dur-
domains were considered as the multiple views. ing training, without requiring any heuristics. Al-
The performance of co-training will depend on the though we use projections in the proposed method,
complementarity of the information captured by unlike prior work on projection-based UDA, we
the different feature spaces. Therefore, it is an im- do not require splitting the feature space into do-
portant to carefully design multiple feature spaces main independent and domain specific features.
when performing UDA. In contrast, our proposed Moreover, we learn two separate projections for
method does not require such multiple views and each of the source and target domain, which gives
does not require training multiple classifiers for us more flexibility to address the domain-specific
the purpose of assigning pseudo labels for the tar- constrains in the learnt projections.
get domain unlabelled instances, which makes the Neural adaptation methods have recently re-
proposed method easy to implement. ported SoTA for UDA. Louizos et al. (2015) pro-
Tri-training (Zhou and Li, 2005) relaxes the re- posed a Variational Fair Autoencoder (VFAE) to
quirement of co-training for the feature spaces to learn an invariant representation for a domain.
215
They used Maximum Mean Discrepancy (MMD) ing SL (Section 3.1), (b) pseudo labelling TU us-
(Gretton et al., 2006) for further promoting the in- ing a classifier trained on the projected SL (Sec-
variant projected feature space. Ganin et al. (2016) tion 3.2); (c) learning a target projection using the
proposed Domain Adversarial Neural Network pseudo-labelled target instances, and then learning
(DANN) to learn features that combine the dis- a classifier f for the target task (Section 3.3).
criminative power of a classifier and the domain-
invariance of the projection space to simultane- 3.1 Source Projection Learning (Sprj )
ously learn adaptable and discriminative projec- In UDA, the adaptation task does not vary be-
tions. Saito et al. (2017) proposed a deep tri- tween the source and target domains. Therefore,
training method with three neural networks, two we can use SL to learn a projection for the source
for pseudo labelling the target unlabelled data domain Sprj where the separation between an in-
and another one for learning discriminator using stance x ∈ SL and its opposite-labelled nearest
the inferred pseudo labels for the target domain. neighbours is maximised. Specifically, for an in-
Ruder and Plank (2018) proposed Multi-task Tri- stance x we represent the set of k of its nearest
training (MT-Tri) based on tri-training and Bi- neighbours NN(x, D, k) selected from a set D by
LSTM. They show that tri-training is a competi- a vector φ(x, D, k) as the sum of the word embed-
tive baseline and rivals more complex neural adap- dings of the neighbours given by (1).
tation methods. Although MT-Tri does not out- X
perform SoTA on cross-domain sentiment clas- φ(x, D, k) = θ(x, u)u. (1)
u∈NN(x,D,k)
sification tasks, their proposal reduces the time
and space complexity required by the classical tri- Here, the weight θ(x, u) is computed using the co-
training. sine similarity between u and x, and is normalised
P
As stated above, our proposed method Self- s.t. u∈NN(x,D,k) θ(x, u) = 1. Other similar-
Adapt, differs from the prior work discussed ity measures can also be used instead of cosine,
above in that it (a) does not require pivots, (b) does for example, Euclidean distance (Van Asch and
not require multiple feature views, (c) learns two Daelemans, 2016). Then, Sprj is defined by the
different projections for the source and target do- projection matrices A+ ∈ Rd×d and A− ∈ Rd×d
mains and (d) combines a projection and a self- and is learnt by maximising the objective OL given
training step in a non-iterative manner to improve by (2).
the performance in UDA.
X
OL (A+ , A− ) = A+ x − A− φ(x, SL− , k)2
3 Self-Adaptation (Self-Adapt)
2
+
x∈SL
X
+ A− x − A+ φ(x, SL+ , k)2 (2)
In UDA, we are given a set of positively (SL+ ) and 2
−
negatively (SL− ) labelled instances for a source do- x∈SL
main S (SL = SL+ ∪SL− ), and sets of unlabelled in- We initialise A+ and A− to the identify matrix
stances SU and TU respectively for the source and I ∈ Rd×d and apply Adam (Kingma and Ba, 2014)
target domain T . Given a dataset D, we are re- to find their optimal values denoted respectively
quired to learn a classifier f (x, y; D) that returns by A∗+ and A∗− . Finally, we project SL using the
the probability of a test instance x taking a label learnt Sprj to obtain a projected set of source do-
y. For simplicity, we consider the pairwise adap- main labelled instances SL∗ = A∗+ ◦ SL+ ∪ A∗− ◦ SL− .
tation from a single source to single target, and Here, we use the notation A ◦ D = {Ax|x ∈ D}
binary (y ∈ {−1, 1}) classification as the target to indicate the application of a projection matrix
task. However, self-adapt can be easily extended A ∈ Rd×d on elements x ∈ Rd in a dataset D.
to multi-domain and multi-class UDA.
We represent an instance (document/review) x 3.2 Pseudo Label Generation (PL)
by a bag-of-n-gram (BonG) embedding (Arora In UDA, we do not have labelled data for the target
et al., 2018), where we add the pre-trained d- domain. To overcome this issue, inspired by prior
dimensional word embeddings w ∈ Rd for the work on self-training approaches to UDA, we train
words w ∈ x to create a d-dimensional feature a classifier f (x, y; SL∗ ) on SL∗ first and then use
vector x ∈ Rd representing x. Self-adapt consists this classifier to assign pseudo labels for the tar-
of three steps: (a) learning a source projection us- get domain’s unlabelled data TU , if the classifier
216
Algorithm 1 Pseudo Label Generation lected from positively (TL0+ ) and negatively (TL0− )
pseudo-labelled instances. The objective OL 0 for
Input: source domain positively labelled data SL+ ,
source domain negatively labelled data this optimisation problem is given by (3).
SL− ,
X
source domain positive transformation ma- 0
OL (B+ , B− ) = B+ x − B− φ(x, TL0− , k)2
2
trix A+ , 0+
x∈TL
source domain negative transformation X
B− x − B+ φ(x, TL0+ , k)2 (3)
+
matrix A− , 0−
x∈TL
2
217
are labelled as negative. 3 star reviews are ignored overall performance in UDA using k = 1 for all
because of their ambiguity. In addition to the la- the steps. Specifically, we consider the following
belled reviews, the Amazon multi-domain dataset baselines.
contains a large number of unlabelled reviews for
each domain. We use the official balanced train NA: No-adaptation. Learn a classifier from SL
and test dataset splits, which has 800 (pos), 800 and simply use it to classify sentiment on target
(neg) training instances and 200 (pos), 200 (neg) domain test instances, without performing any do-
test instances for each domain. We name this main adaptation.
dataset as the Multi-domain Adaptation Dataset Sprj : Learn a source projection Sprj and apply it
(MAD). to project SL to obtain SL∗ = A∗+ ◦ SL+ ∪ A∗− ◦ SL− .
One issue that is often raised with MAD is that Train a sentiment classifier using SL∗ and use it to
it contains only four domains. In order to robustly classify target domain test instances.
evaluate the performance of an UDA method we
must evaluate on multiple domains. Therefore, in Sprj +PL: Use the classifier trained using SL∗ on
addition to MAD, we also evaluate on an extended target domain unlabelled data to create a pseudo-
dataset that contains 16 domains. We name this labelled dataset TL0 . Train a sentiment classifier on
dataset as the Extended Multi-domain Adaptation SL∗ ∪ TL0 and use it to classify target domain test
Dataset (EMAD). The reviews for the 16 domains instances.
contained in EMAD were also collected by Blitzer
et al. (2007), but were not used in the evaluations. Sprj +Tprj : This is the proposed method includ-
The same star-based procedure used in MAD is ing all three steps. A target projection Tprj is
used to label the reviews in EMAD. We randomly learnt using TL0 and is applied to obtain TL0∗ =
select 20% of the available labelled reviews as test B∗+ ◦ TL0+ ∪ B∗− ◦ TL0− . Finally, a sentiment classi-
data and construct a balanced training dataset from fier is trained using SL∗ ∪ TL0∗ and used to classify
the rest of the labelled reviews (i.e. for each do- target domain test instances.
main we have equal number of positive and nega- With all methods, we keep k = 1 in the nearest
tive labelled instances in the train neighbour feature representation in (1) for the ease
16
datasets). Like- of comparisons. Confidence threshold τ is tuned
wise in MAD, we generate 2 = 240 pair-wise
domain adaptation tasks from EMAD. in the range [0.5, 0.9] using cross-validation. From
Table 1 we see that Sprj consistently outperforms
We train an `2 regularised logistic regression
NA, showing that even without using any infor-
as the target (sentiment) classifier, in which we
mation from the target domain, it is still useful
tune the regularisation coefficient using validation
to learn source domain projections that discrim-
data. We randomly select 10% from the target
inates instances with opposite labels. When we
domain labelled data, which is separate from the
perform pseudo labelling on top of source projec-
train or test data. We tune regularisation coeffi-
tion learning (Sprj +PL) we see a slight but consis-
cient in [0.001, 0.01, 0.1, 1]. We use 300 dimen-
tent improvement in all domain-pairs. However,
sional pre-trained GloVe word embeddings (Pen-
when we use the pseudo labelled instances to learn
nington et al., 2014) to create BonG embeddings
a target projection (Sprj +Tprj ) we obtain the best
for uni and bigrams. We found that a maximum of
performance in all domain-pairs. Moreover, the
100 epochs to be sufficient to reach convergence in
obtained results are significantly better under the
all projection learning tasks for all domains. The
stricter p < 0.001 level over the NA baseline in 7
source code implementation of self-adapt will be
out of 12 domain-pairs.
made publicly available upon paper acceptance.
Table 3 shows the classification accuracy for the
EMAD. Due to the limited availability of space,
4.1 Effect of Projection Learning and
we show the average classification accuracy for
Pseudo-Labelling
adapting to the same target domain instead of
Our proposed method consists of 3 main steps as showing all 240 domain-pairs for EMAD. Like-
described in Section 3: source projection learn- wise in MAD, we see in EMAD we obtain the
ing, pseudo labelling and target projection learn- best results when we use both source and target
ing. Using MAD, in Table 1, we compare the rel- projections. Interestingly, we see that the pro-
ative effectiveness of these three steps towards the posed method adapting well even to the domains
218
S −T NA Sprj Sprj +PL Sprj +Tprj adapt on the same datasets, feature representations
B-D 73.50 74.25 74.50 76.25 and settings to conduct a fair comparison. All clas-
B-E 64.00 73.25** 73.50** 77.50** sical self-training methods were trained using the
B-K 68.50 75.25* 76.00* 78.75**
D-B 74.00 75.50 75.50 75.50 source domain labelled instances SL as seed data.
D-E 64.75 71.25* 71.50* 74.25** As discussed in Section 3.2, similar to Self-Adapt,
D-K 71.50 75.00 76.25 79.75**
E-B 67.25 74.50* 75.25** 76.25** we observed that the performance did not signifi-
E-D 66.75 67.75 68.00 69.50 cantly increase beyond the first iteration for any of
E-K 76.00 81.00 81.00 81.00
K-B 63.00 73.75** 73.75** 75.00**
the classical self-training methods in UDA. Con-
K-D 71.25 71.25 71.00 72.50 sequently, we compare all classical self-training
K-E 69.00 77.50** 78.00** 78.25** methods for their peak performance, obtained after
the first iteration. We tune the confidence thresh-
Table 1: Target domain test data classification old τ for each method using validation data and
accuracy for the different steps in the proposed found the optimal value of τ to fall in the range
method (k = 1). S − T denotes adapting from [0.6, 0.9]. k is a hyperparameter selected using
a source S to a target T domain. The best result validation dataset for Self-Adapt in comparison.
for each domain-pair is bolded. Statistically sig- Experimental results on MAD and EMAD are
nificant improvements over NA according to the shown respectively in Tables 2 and 4. From those
binomial exact test are shown by “*” and “**” re- Tables, we see that Self-Adapt for most of the
spectively at p = 0.01 and p = 0.001 levels. domain pairs performs similarly or slightly worse
with smaller numbers of unlabelled data such as than NA. Although Tri and Tri-D outperform NA
gourmet food (168 labelled and 267 unlabelled on all cases, we found that those two methods are
train instances). This is encouraging because it highly sensitive to the seed instances used to ini-
shows that the proposed method can overcome the tialise the pseudo-labelling process. We find the
lack of labelled instances via pseudo labelling and proposed Self-Adapt to outperform all classical
projection learning. self-training based methods in 11 out of 12 do-
main pairs in MAD and in all 16 target domains
4.2 Comparisons against Self-Training in EMAD, showing a strong and robust improve-
ment over classical self-training methods. This re-
Ruder and Plank (2018) evaluated classical sult shows that by combining source and target do-
general-purpose semi-supervised learning meth- main projections with self-training, we can obtain
ods proposed for inducing pseudo labels for un- superior performance in UDA in comparison to us-
labelled instances using a seed set of labelled in- ing classical self-training methods alone.
stances in the context of UDA. They found that
tri-training to outperform more complex neural S-T NA Self Tri Tri-D Self-Adapt
SoTA UDA methods. Considering the fact that B-D 73.50 74.25 74.75 77.25 77.25
B-E 64.00 65.00 73.25** 72.00** 78.50**
Self-Adapt is performing pseudo-labelling, sim- B-K 68.50 70.75 73.25 73.75 79.00**
D-B 74.00 73.00 77.00 76.00 81.00*
ilar to other self-training methods, it is interesting D-E 64.75 65.25 73.75** 70.50* 75.50**
to see how well it compares against classical self- D-K 71.50 71.75 75.75 74.00 79.75**
E-B 67.25 68.25 68.50 74.25* 76.25**
training methods for inducing labels (Yarowsky, E-D 66.75 66.25 68.25 73.50* 69.50
1995; Abney, 2007; Zhou and Li, 2005; Søgaard, E-K 76.00 76.50 82.50* 81.00 82.50*
K-B 63.00 63.00 70.00* 73.00** 75.00**
2010) when applied to UDA. Specifically, we K-D 71.25 71.00 71.25 72.00 72.75
K-E 69.00 68.75 73.00 75.50* 79.00**
consider the classical self-training (Yarowsky,
1995; Abney, 2007) (Self), Tri-training (Zhou and Table 2: Target domain test data classification ac-
Li, 2005) (Tri) and Tri-training with diversifica- curacy of classical self-training methods when ap-
tion (Søgaard, 2010) (Tri-D). For each of those plied to UDA.
methods, we use the labelled data in the source
domain as seeds and induce labels for the unla- 4.3 Comparisons against Neural UDA
belled data in the target domain. Table 2 reports In Table 5, we compare Self-Adapt against
the results on MAD. the following neural UDA methods on MAD:
We re-implement the classical self-training Variational Fair Autoencoder (Louizos et al.,
methods considered by Ruder and Plank (2018) 2015) (VFAE), Domain-adversarial Neural Net-
and evaluated them against the proposed self- works (Ganin et al., 2016) (DANN), Asymmet-
219
Target Domain NA Sprj Sprj +PL Sprj +Tprj Target Domain NA Self Tri Tri-D Self-Adapt
apparel 71.39 75.31 75.66 76.68* apparel 71.39 71.38 73.96 73.89 76.68
baby 68.00 71.06 71.37 72.63 baby 68.00 67.96 69.71 69.84 72.63
beauty 69.66 73.20 73.18 73.70 beauty 69.66 70.54 71.73 70.49 73.70
camera and photo 68.46 73.46 74.02 74.54* camera and photo 68.46 68.44 71.31 71.28 74.54*
computer and video games 61.78 67.38 67.85 68.11* computer and video games 61.78 62.28 64.65 64.93 68.11*
gourmet food 72.34 81.13** 82.89** 83.15** gourmet food 72.34 74.10 75.39 77.88 83.15**
grocery 71.01 76.90* 77.57 * 78.14* grocery 71.01 71.83 75.22 74.29 78.14*
health and personal care 65.85 68.38 68.67 69.24 health and personal care 65.85 66.08 67.10 67.07 69.24
jewelry and watches 68.90 77.86** 79.11** 79.79** jewelry and watches 68.90 71.04 76.67** 76.21** 79.79**
magazines 65.28 71.54* 71.45* 72.16* magazines 65.28 65.34 67.58 67.47 72.16*
music 66.27 70.69 70.89 71.41 music 66.27 66.22 68.82 68.27 71.41
outdoor living 71.97 76.77 78.01* 78.93* outdoor living 71.97 74.01 76.21 74.79 78.93*
software 65.72 69.56 69.60 70.31 software 65.72 65.47 67.36 67.53 70.31
sports and outdoors 66.56 70.18 70.67 71.02 sports and outdoors 66.56 66.87 69.22 68.05 71.02
toys and games 69.57 72.82 73.22 73.65 toys and games 69.57 70.03 71.76 72.32 73.65
video 64.19 67.59 68.37 68.97 video 64.19 64.88 66.34 67.49 68.97
Table 3: Average classification accuracy on each Table 4: Average classification accuracy on each
target domain in EMAD for the steps in the pro- target domain in EMAD of classical self-training
posed method (k = 1). based methods when applied to UDA.
ric Tri-training (Saito et al., 2017) (Asy-Tri), and stances that are determined to be confident by the
Multi-task Tri-training (Ruder and Plank, 2018) classifier more than a pre-defined threshold τ . The
(MT-Tri). We select these methods as they are the ability to operate on a lower-dimensional feature
current SoTA for UDA on MAD, and report the space and obviating the need to subsample unla-
results from the original publications in Table 5. belled data are properties of the proposed method
Although only in 3 out of 12 domain-pairs Self- that are attractive when applying UDA methods on
Adapt is obtaining the best performance, the dif- large datasets and across multiple domains.
ference of performance between DANN and Self-
Adapt is not statistically significant. Although S-T VFAE DANN Asy-Tri MT-Tri Self-Adapt
MT-Tri is outperforming Self-Adapt in 8 domain- B-D 79.90 78.40 80.70 81.47 77.25
pairs, it is noteworthy that MT-Tri is using a larger B-E
B-K
79.20
81.60
73.30
77.90
79.80
82.50
78.62
78.09
78.50
79.00
feature space than that of Self-Adapt. Specif- D-B 75.50 72.30 73.20 77.49 81.00**
D-E 78.60 75.40 77.00 79.66 75.50
ically, MT-Tri is using 5000 dimensional tf-idf D-K 82.20 78.30 82.50 81.23 79.75
weighted unigram and bigram vectors for repre- E-B 72.70 71.10 73.20 73.43 76.25
E-D 76.50 73.80 72.90 75.05 69.50
senting reviews, whereas we Self-Adapt uses a E-K 85.00 85.40 86.90 87.07 82.50
K-B 72.00 70.90 72.50 73.60 75.00
300 dimensional BonG representation computed K-D 73.30 74.00 74.90 77.41 72.75
K-E 83.80 84.30 84.60 86.06 79.00
using pre-trained GloVe vectors. Moreover, prior
work on neural UDA have not used the entire un-
Table 5: Classification accuracy compared with
labelled datasets and have sampled a smaller sub-
neural adaptation methods. The best result is
set due to computational feasibility. For exam-
bolded. Statistically significant improvements
ple, MT-Tri uses only 2000 unlabelled instances
over DANN according to the binomial exact test
for each domain despite the fact that the original
are shown by “*” and “**” respectively at p =
unlabelled datasets contain much larger numbers
0.01 and p = 0.001 levels.
of reviews. This is not only a waste of available
data but it is also non-obvious as how to subsam-
ple unlabelled data for training. Our preliminary 5 Conclusions
experiments revealed that the performance of neu-
ral UDA methods to be sensitive to the unlabelled We proposed Self-Adapt, an UDA method that
datasets used.3 On the other hand, Self-Adapt combines projection learning and self-training.
does not require sub-sampling of unlabelled data Our experimental results on two datasets for cross-
and uses all the available unlabelled data for UDA. domain sentiment classification show that projec-
During the pseudo-labelling step, Self-Adapt au- tion learning and self-training have complemen-
tomatically selects a subset of unlabelled target in- tary strengths and jointly contribute to improve
UDA performance. In future, we plan to apply
3
Unfortunately, the source code for MT-Tri was not avail- Self-Adapt to other UDA tasks in NLP such as
able for us to run this method with the same set of features
and unlabelled dataset that we used. Therefore, we report the cross-domain POS tagging and NER.
results from the original publication.
220
References Proceedings of Workshop on Robust Unsupervised
and Semisupervised Methods in Natural Language
Steven Abney. 2007. Semisupervised Learning for Processing, pages 9–16.
Computational Linguistics, 1st edition. Chapman &
Hall/CRC. Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan,
Pascal Germain, Hugo Larochelle, François Lavi-
Sanjeev Arora, Mikhail Khodak, Nikunj Saunshi, and olette, Mario Marchand, and Victor Lempitsky.
Kiran Vodrahalli. 2018. A compressed sensing view 2016. Domain-adversarial training of neural net-
of unsupervised text embeddings, bag-of-n-grams, works. Journal of Machine Learning Research,
and LSTMs. In International Conference on Learn- 17(59):1–35.
ing Representations.
Arthur Gretton, Karsten M Borgwardt, Malte Rasch,
John Blitzer, Mark Dredze, and Fernando Pereira. Bernhard Schölkopf, and Alex J Smola. 2006. A
2007. Biographies, bollywood, boom-boxes and kernel method for the two-sample-problem. In Ad-
blenders: Domain adaptation for sentiment classi- vances in neural information processing systems,
fication. In Proc. of ACL, pages 440–447. pages 513–520.
John Blitzer, Ryan McDonald, and Fernando Pereira. Diederik P. Kingma and Jimmy Ba. 2014. Adam:
2006. Domain adaptation with structural correspon- A method for stochastic optimization. CoRR,
dence learning. In Proc. of EMNLP, pages 120–128. abs/1412.6980.
Avrim Blum and Tom Mitchell. 1998. Combining la- Christos Louizos, Kevin Swersky, Yujia Li, Max
beled and unlabeled data with co-training. In Pro- Welling, and Richard S. Zemel. 2015. The varia-
ceedings of the eleventh annual conference on Com- tional fair autoencoder. CoRR, abs/1511.00830.
putational learning theory, pages 92–100. ACM.
David McClosky, Eugene Charniak, and Mark John-
Danushka Bollegala, Tingting Mu, and Yannis Gouler- son. 2006. Reranking and self-training for parser
mas. 2015. Cross-domain sentiment classifica- adaptation. In Proceedings of the 21st International
tion using sentiment sensitive embeddings. IEEE Conference on Computational Linguistics and the
Transactions on Knowledge and Data Engineering 44th annual meeting of the Association for Compu-
(TKDE), 28(2):398–410. tational Linguistics, pages 337–344. Association for
Computational Linguistics.
Danushka Bollegala, David Weir, and John Carroll.
2011. Using multiple sources to construct a senti- Pietro Morerio, Jacopo Cavazza, and Vittorio Murino.
ment sensitive thesaurus for cross-domain sentiment 2018. Minimal-entropy correlation alignment for
classification. In ACL/HLT’11, pages 132 – 141. unsupervised deep domain adaptation. In Interna-
tional Conference on Learning Representations.
Danushka Bollegala, David Weir, and John Carroll.
2013. Cross-domain sentiment classification using Sinno Jialin Pan, Xiaochuan Ni, Jian-Tao Sun, Qiang
a sentiment sensitive thesaurus. IEEE Transactions Yang, and Zheng Chen. 2010. Cross-domain senti-
on Knowledge and Data Engineering, 25(8):1719 – ment classification via spectral feature alignment. In
1731. Proc. of WWW, pages 751–760.
Minmim Chen, Kilian Q. Weinberger, and John Blitzer. Jeffery Pennington, Richard Socher, and Christo-
2011. Co-training for domain adaptation. In pher D. Manning. 2014. Glove: global vectors for
NIPS’11. word representation. In Proc. of Empirical Methods
in Natural Language Processing (EMNLP), pages
Xia Cui, Frans Coenen, and Danushka Bollegala. 1532–1543.
2017a. Effect of data imbalance on unsupervised
Rajat Raina, Alexis Battle, Honglak Lee, Benjamin
domain adaptation of part-of-speech tagging and
Packer, and Andrew Y. Ng. 2007. Self-taught learn-
pivot selection strategies. In Proc. of the Wokshop
ing: Transfer learning from unlabeled data. In
on Learning With Imbalanced Domains: Theory
ICML’07.
and Applications (LIDTA) at the European Confer-
ence on Machine Learning and Principles and Prac- Roi Reichart and Ari Rappoport. 2007. Self-training
tice of Knowledge Discovery in Databases (ECML- for enhancement and domain adaptation of statisti-
PKDD), pages 103–115. cal parsers trained on small datasets. In ACL 2007,
pages 616 – 623.
Xia Cui, Frans Coenen, and Danushka Bollegala.
2017b. TSP: Learning task-specific pivots for un- Sebastian Ruder and Barbara Plank. 2018. Strong
supervised domain adaptation. In Proc. of the baselines for neural semi-supervised learning under
European Conference on Machine Learning and domain shift. pages 1044–1054.
Principles and Practice of Knowledge Discovery in
Databases (ECML-PKDD), pages 754–771. Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada.
2017. Asymmetric tri-training for unsupervised do-
Brett Drury, Luı́s Torgo, and Jose Joao Almeida. 2011. main adaptation. In International Conference on
Guided self training for sentiment classification. In Machine Learning, pages 2988–2997.
221
Anders Søgaard. 2010. Simple semi-supervised train-
ing of part-of-speech taggers. In Proceedings of the
ACL 2010 Conference Short Papers, ACLShort ’10,
pages 205–208, Stroudsburg, PA, USA. Association
for Computational Linguistics.
Vincent Van Asch and Walter Daelemans. 2016. Pre-
dicting the effectiveness of self-training: Appli-
cation to sentiment classification. arXiv preprint
arXiv:1601.03288.
David Yarowsky. 1995. Unsupervised word sense dis-
ambiguation rivaling supervised methods. In Pro-
ceedings of the 33rd Annual Meeting on Association
for Computational Linguistics, ACL ’95, pages 189–
196, Stroudsburg, PA, USA. Association for Com-
putational Linguistics.
Ning Yu and Sandra Kübler. 2011. Filling the
gap: Semi-supervised learning for opinion detec-
tion across domains. In Proceedings of the Fif-
teenth Conference on Computational Natural Lan-
guage Learning, pages 200–209. Association for
Computational Linguistics.
Zhi-Hua Zhou and Ming Li. 2005. Tri-training: ex-
ploiting unlabeled data using three classifiers. IEEE
Transactions on Knowledge and Data Engineering,
17(11):1529–1541.
222
SPECULATION and NEGATION DETECTION in FRENCH
BIOMEDICAL CORPORA
Clément Dalloux and Vincent Claveau
Univ Rennes, Inria, CNRS, IRISA, F-35000 Rennes, France
name.surname@irisa.fr
Natalia Grabar
UMR 8163 STL CNRS, Université de Lille, France
natalia.grabar@univ-lille.fr
i.e., tokens within the sentence which are affected from bone-marrow homograft)
by the negation or speculation. In this paper, we 4. Lymphome non hodgkinien à cellules B ma-
present two French datasets annotated with nega- tures récidivant/réfractaire. (Relapsed/refrac-
tion and speculation cues and their scope. We also tory mature B-cell non-Hodgkin’s lymphoma.)
223
Proceedings of Recent Advances in Natural Language Processing, pages 223–232,
Varna, Bulgaria, Sep 2–4 2019.
5. Elle n’est soulagée que par la marche et doit 3. Elle aurait eu la pose à 2 reprises d’un dis-
donc écouter la télévision en faisant les cent positif intrautérin (DIU). (She would have had
pas dans son salon. (She is only relieved by walk- the pose of an intrauterine device (IUD) twice.)
ing and must therefore listen to the TV pacing in her
living room.)
Example (1) shows a typical occurrence of
pourrait (could). As in English, where can, could,
Examples (1-2) show the possible effect of fre- may, etc. express speculation, in French, pouvoir
quency adverbs, here toujours (always). In the can occur in many forms.
first sentence traitements locaux (chirurgie, radio- Example (2) shows the effect of reste à (re-
thérapie, radiofréquence, cryoablation), the con- mains to) combined with an infinitive verb, here
tent would be negated without toujours(always). In démontrer. Other infinitive verbs, such as expli-
the second sentence, with or without toujours, the quer (to explain) or vérifier (to verify), associated with
meaning of the sentence does not change, there- reste à trigger speculation.
fore, the scope of the negation remains the same. Example (3) shows how the conditional tense
Example (3) shows how the preposition en de- triggers speculation. In English, would triggers
hors de(apart from), can stop the scope of nega- speculation in this case, which is simpler to de-
tion. Many other prepositions such as à part, à tect. Indeed, in French, the conditional tense is
l’exception de or excepté, with more or less the expressed via suffixes (-ais, -ais, -ait, -ions, -iez,
same meaning than en dehors de (apart from), would -aient), which makes the detection harder, espe-
have the same effect on the negation scope. How- cially for supervised learning techniques.
ever, these prepositions can also play the role of
3 RELATED WORK
negation by themselves.
Examples (4) show that cues can also be We present several corpora and methods that have
included in medical concepts such as non been proposed in the existing work to tackle the
hodgkinien (non-Hodgkin’s). In our work, we chose tasks of speculation and negation detection.
to label hodgkinien as part of the negation scope.
Finally, example (5) shows the context in which 3.1 DATA
the ambiguous word pas, meaning both no/not and In the recent years, several specialized corpora in
footstep, has the non-negation meaning. In this ex- English have been annotated with speculation and
ample, pas is part of the idiomatic expression faire negation, which has resulted in models for their
les cent pas (pacing, walking around). Another ambi- automatic detection. These corpora can be divided
guity is related to the adverb plus meaning either into two categories: (1) corpora annotated with
more or, in conjunction with ne, no more. cues and scopes, such as Bioscope (Vincze et al.,
2008) or *SEM-2012, and (2) corpora focusing on
2.2 SPECULATION concepts and named entities, such as I2B2 and Mi-
The expression of speculation can be even more pacq. We briefly describe these corpora. The Bio-
complex than negation. Indeed, speculation can scope corpus (Vincze et al., 2008) contains reports
be triggered by many specific sequences of words. of radiological examinations, scientific articles,
We describe several of them below. and abstracts from biomedical articles. Each sen-
tence and each negation and speculation cue/scope
1. En effet, l’arrêt du traitement antituberculeux pair receives unique identifier. The *SEM-2012
en soi pourrait permettre un rétablissement corpus (Morante and Blanco, 2012) consists of a
de la fonction normale des héptocytes en Sherlock Holmes novel and three other short sto-
éliminant la source de l’atteinte hépatique. ries written by Sir Arthur Conan Doyle. It con-
(Indeed, stopping TB treatment per se could restore tains 5,520 sentences, among which 1,227 sen-
normal hepatocyte function by eliminating the source tences are negated. Each occurrence of the nega-
of liver damage.) tion, the cue and its scope are annotated, as well
2. Le bénéfice de l’association lénalomide + R- as the focus of the negation if relevant. In this cor-
CHOP au rapport par rapport au R-CHOP pus, cues and scopes can be discontinuous. The
reste à démontrer. (The benefit of the combination I2B2/VA-2010 challenge (Uzuner et al., 2011) fea-
of lenalomide + R-CHOP compared with R-CHOP re- tured several tasks using US clinical records. One
mains to be demonstrated.) task aimed the detection of statements and of their
224
scope. Medical concepts had to be associated with French French
the corresponding statement: present, absent, pos- clin. trials clin. cases
sible, conditional, hypothetical or not associated Documents – 200
with the patient. Mipacq (Albright et al., 2013) Sentences 6,547 3,811
is another corpus with clinical data in English an- Tokens 150,084 87,487
notated with syntactic and semantic labels. Each Vocabulary (types) 7,880 10,500
detected UMLS entity has two attribute locations: Negative sentences 1,025 804
negation (true or false) and status (none, possible, Speculative sentences 630 226
HistoryOf or FamilyHistoryOf ).
Table 1: Statistics on the two French corpora
3.2 EXPERT SYSTEMS
Among the rule based systems dedicated to the
negation detection, NegEx (Chapman et al., 2001) as a fall-back by Packard et al. (2014) when the
pioneered the area. It uses regular expressions main MRS (minimal recursion semantics) Crawler
to detect the cues and to identify medical terms cannot parse the sentence. Qian et al. (2016)
in their scope. It was later adapted to various addresses the scope detection with an approach
languages including French (Deléger and Grouin, based on a convolutional neural network which
2012). ConText (Harkema et al., 2009), derived extracts features from various syntactic paths be-
from NegEx, covers more objectives: negation, tween the cues and the candidate tokens in con-
temporality, and the subject concerned by this in- stituency and dependency parsed trees. Fancellu
formation in the clinical texts. It has been adapted et al. (2016) uses neural networks to solve the
to French (Abdaoui et al., 2017). In another problem of negation scope detection. One ap-
work, medical concepts may receive additional la- proach uses Feed-forward neural network, while
bels (positive, negative or uncertain) Elkin et al. the other, which appears to be more efficient for
(2005a). Özgür and Radev (2009); Øvrelid et al. the task, uses a bidirectional Long Short-Term
(2010); Kilicoglu and Bergler (2010) exploit lexi- Memory (BiLSTM) neural network. Given the re-
cal, grammatical and syntactic information to de- sults from the latter approach, it inspired our work.
tect speculation and its scope. ScopeFinder (Apos-
tolova et al., 2011) detects the scope of nega-
4 FRENCH MEDICAL CORPORA
tion and speculation with rules built automatically
from BioScope (lexico-syntactic patterns extrac-
We manually annotated two corpora from the
tion). NegBio (Peng et al., 2018) detects both
biomedical field. Table 1 presents some statistics
negation and speculation in radiology reports with
on these corpora: the number of words, the va-
rules based on universal dependency graphs.
riety of the vocabulary, the number of sentences,
3.3 SUPERVISED LEARNING the number of negative sentences with one or
To our knowledge, Light et al. (2004) is the first more negations. The Inter Annotator Agreement
work to include supervised learning for specula- (IAA) on negation annotation is high (Cohen’s
tion detection. It relies on SVM to select specula- κ=0.8461).
tive sentences in MEDLINE abstracts. Tang et al.
(2010) proposes a cascade method based on CRF 4.1 ESSAI: FRENCH CORPUS with
and SVM classifiers to detect speculation cues and CLINICAL TRIALS
another CRF classifier to identify their scopes.
Velldal et al. (2012) proposes a SVM-based cue One corpus contains clinical trial protocols in
detection system, trained on simple n-grams fea- French. They were mainly obtained from the Na-
tures computed on the local lexical context (words tional Cancer Institute registry1 . The typical pro-
and lemmas). This system offers a hybrid detec- tocol consists of two parts: the summary of the
tion of the scope, which combines expert rules, trial, which indicates the purpose of the trial and
operating on syntactic dependency trees, with a the methods applied; and a detailed description of
ranking SVM that learns a discriminative ranking the trial with the inclusion and exclusion criteria.
function over nodes in constituent trees. It was
further improved by Read et al. (2012) and is used 1
https://www.e-cancer.fr
225
Form Lemma POS Cue scope
Pas pas ADV pas
de de PRP de
dyspnée dyspnée NOM dyspnée
. . SENT
226
often shared, rare words can also get reliable rep- We performed the gradient descent using the L-
resentations. BFGS (Limited-memory BFGS) method with 0.1
These two word embedding models are trained L1 penalty and 0.01 L2 penalty. We only experi-
using the Skip-Gram algorithm, 100 dimensions, a ment with CRFs for the cue detection task, in com-
window of 5 words before and after each word, parison with BiLSTM-CRF.
a minimum count of five occurrences for each
5.4 EVALUATING LABELING SYSTEMS
word and negative sampling. The training data
are composed of the French Wikipedia articles We use standard evaluation measures: precision
and biomedical data. The latter includes the ES- P , recall R, and F1 score. The scope detection is
SAI and CAS corpora, the French Medical Corpus evaluated in two ways: (1) on individual scope to-
from CRTT2 and the Corpus QUAERO Médical kens which is the standard evaluation, and (2) on
du français3 (Névéol et al., 2014). These models exact scopes to assess more strictly how efficient
are trained using the Gensim4 (Rehurek and Sojka, our models are. For the latter, we use the avail-
2010) python library. able evaluation script5 . Each corpus is randomly
segmented into the training set (80%, 20% for val-
5.2 RECURRENT NEURAL NETWORK idation), and the test set (20%).
Recurrent neural network takes into account the 6 CUE DETECTION
previously seen data in addition to the currently
seen data. This is implemented with loops in The speculation and negation cue detection is the
the architecture of the network, which allows the first step of the task. To tackle this problem,
information to persist in memory. Among the we experiment with two supervised learning ap-
RNNs, long short-term memory networks (LSTM) proaches. First, we train a CRF using several fea-
(Hochreiter and Schmidhuber, 1997) are the most tures (words, lemmas and POS-tags) with empir-
efficient for the learning of long-term dependen- ically defined window over features. Our second
cies and are therefore more suitable to solve the approach uses a BiLSTM with a CRF output layer,
problem of discontinuous scope, which is typical which is trained on the same features. We did not
for the negation. LSTM cells are also more effi- use any pre-trained embeddings for this task.
cient at retaining useful information during back- Table 3 presents the results obtained with our
propagation. approaches on the ESSAI and CAS corpora. We
We use a bidirectional LSTM, which operates can see that cue detection shows high evaluation
forward and backward on the sentence, to detect values: 93.92 to 97.21 F-measure for negation
cues and scopes. The backward pass is relevant cues, and 86.88 to 91.30 F-measure for specula-
for the scope detection because the scope may be tion cues. Although there is little room for im-
before or after the cue. Prediction is computed provement on negation cue detection, indeed 10k-
by either a softmax or a CRF (suitable for the fold cross-validation with our CRF reaches more
sequence labeling) output layer. We use embed- than 95 F-measure on both corpora, speculation
dings of dimension k = 100 and a dimension- cue detection would benefit from more training
ality of the output space of 400 units per layer examples. Indeed, the potential number of cues
(backward/forward) with 0.5 dropout. 50 epochs and the numerous contexts in which they appear
achieve the highest F1 score on the validation sets. and do or do not express speculation makes them
harder to detect and will require more annotated
5.3 CONDITIONAL RANDOM FIELDS examples. For both negation and speculation cue
detection, our CRF is slightly more efficient than
Conditional random fields (CRFs) (Lafferty et al.,
the BiLSTM-CRF when the CAS corpus is in-
2001) are statistical methods used to label word se-
volved, which indicates that the CAS corpus con-
quences. By training a model on appropriate fea-
tains less complex examples than ESSAI.
tures and labels to be predicted, the CRFs gener-
ally obtain good results with much lower training 7 SCOPE DETECTION
time than neural networks.
In all the scope detection experiments proposed,
2
https://bit.ly/2LOJfEW we only train the neural networks on nega-
3
https://quaerofrenchmed.limsi.fr/
4 5
https://radimrehurek.com/gensim/ https://github.com/ffancellu/NegNN
227
System Corpus window size P R F1
CRF (4) 96.05 91.89 93.92
ESSAI
BiLSTM-CRF None 95.10 94.58 94.84
Negation
CRF (4) 97.05 97.37 97.21
CAS
BiLSTM-CRF None 97.02 97.02 97.02
CRF (4) 91.43 82.76 86.88
ESSAI
BiLSTM-CRF None 91 83.84 87.27
Speculation
CRF (4) 93.93 88.82 91.30
ESSAI+CAS
BiLSTM-CRF None 91.22 87.92 89.54
Table 3: Precision, Recall and F1 -score for the cue detection task on the two corpora (bold: best scores).
Table 4: Precision, Recall and F1 -score for the scope detection task (bold: best scores).
Table 5: Cross-corpora Precision, Recall and F1 -score for the scope detection task (bold: best scores).
228
tive/speculative sentences. The base system takes unit without renal failure after eight days of manage-
as input an instance I(w, c, t), where each word ment and five hemodialysis sessions.)
is represented by: w vector (word-embedding), c
The second example illustrates the error that im-
vector (cue-embedding), indicating if the word is a
pacts precision. Here, the model wrongly predicts
cue or not, t vector (POS-tag-embedding). Prelim-
that all tokens in the sentence are within the scope.
inary tests showed that adding lemmas as features
In the reference data, the cue aucun (any, no) of-
only decreases the F-measure. For each system,
ten occurs at the beginning of sentences, and in
we use the same empirically defined hyperparam-
sentences with many instances of negation. The
eters given before. During training, embeddings
model, mostly trained on this kind of examples,
weights are updated.
may try to reproduce these structures which causes
Table 4 indicate the results obtained for the
bad prediction in some cases.
scope detection task. One can see that it is easier
to predict the scope of negation cues (up to 90.81 • GOLD: Les colorations spéciales (PAS, col-
F-measure) than of speculation cues (up to 86.73 oration de Ziehl-Neelsen, coloration de Gro-
F-measure). Results show that using pre-trained cott) ne [mettaient en évidence] aucun [agent
embeddings improves F-measures for exact scope pathogène].
detection by up to 6 points for both negation and • PRED: [Les colorations spéciales (PAS,
speculation. Moreover, the CRF output layer ei- coloration de Ziehl-Neelsen, coloration de
ther outperforms the softmax layer or reaches an Grocott]) ne [mettaient en évidence] aucun
equivalent F-measure for exact scope detection. In [agent pathogène].
another experiment, we trained the models on one
(Special stains (PAS, Ziehl-Neelsen stain, Grocott
corpus and tested them on the other (Table 5): the
stain) showed no pathogens.)
models trained on ESSAI are more efficient, and
the negation and speculation structures are more In the third example, the error impacts both pre-
stable in the CAS corpus. However, even though cision and recall. In this example, we have two
CAS (speculation) was only trained on 226 exam- instances of negation with the same cues: n...pas.
ples, the model still shows decent results in scope Usually, its scope follows, however, in the first in-
tokens detection. stance it precedes. As we do not have many ex-
amples of this kind to train on, the model fails to
7.1 ERROR ANALYSIS correctly label the sequence. In the second nega-
An analysis of the results makes it possible to iso- tion instance, the scope may be shorter than usual,
late frequent types of errors. In the following ex- which impacts precision.
amples, the speculation and negation cues are un-
• GOLD: [Le retrait du matériel
derlined, the scope is between brackets, while the
d’ostéosynthèse incriminé] n’[est] pas
segments in bold correspond to predictions errors.
[systématique], ce qui explique qu’il n’[ait]
7.1.1 NEGATION pas [été proposé] à notre patient asymp-
In the first example, the prediction fails at labeling tomatique.
rénale (renal). In the majority of cases in the ref- • PRED: Le retrait du matériel
erence data, the scopes associated to the cue sans d’ostéosynthèse incriminé n’[est] pas
often only include one token, which may be caus- [systématique], ce qui explique qu’il n’[ait]
ing this error that impacts recall: pas [été proposé à notre patient asymptoma-
• GOLD: Le patient sortira du service de tique].
réanimation guéri et sans [insuffisance (The removal of the implicated osteosynthesis material
rénale] après huit jours de prise en charge is not systematic, which explains why it has not been
et cinq séances d’hémodialyse. proposed to our asymptomatic patient.)
• PRED: Le patient sortira du service de 7.1.2 SPECULATION
réanimation guéri et sans [insuffisance]
In the first example, the scope of the speculation
rénale après huit jours de prise en charge et
has been predicted up to the end of the preposition
cinq séances d’hémodialyse.
while in the reference data, the scope covers the
(The patient will be discharged from the intensive care verbal group. This typically impacts precision.
229
• GOLD: Ce médicament n’a pas la toxicité • PRED: L’objectif de cet essai est
de la chimiothérapie, mais entraine une dis- d’évaluer la diminution des complications
parition des lymphocytes B normaux pendant postopératoires de l’oesophagectomie qui
plusieurs mois, ce qui pourrait [favoriser la devrait [être obtenue en réalisant une partie
survenue d’infections graves] car ces lym- de l’intervention sous coelioscopie], chez
phocytes participent à la défense immuni- des patients ayant un cancer de l’oesophage
taire. résécable.
• PRED: Ce médicament n’a pas la toxicité (The objective of this trial is to evaluate the decrease
de la chimiothérapie, mais entraine une dis- in postoperative complications of esophagectomy that
parition des lymphocytes B normaux pendant should be achieved by performing part of the proce-
plusieurs mois, ce qui pourrait [favoriser la dure under laparoscopy, for patients with resectable
survenue d’infections graves car ces lympho- esophageal cancer.)
cytes participent à la défense immunitaire] 8 CONCLUSION and FUTURE WORK
(This drug does not have the toxicity of chemotherapy, The interest for the automatic detection of specula-
but causes the disappearance of normal B lymphocytes tion and negation in English with supervised ma-
for several months, which could increase the occur- chine learning has increased in the recent years.
rence of serious infections because these lymphocytes Yet, the lack of data for other languages and for
participate in the immune defense.) specialized domains hampers the further develop-
However, most of our errors impact recall, like ment of such approaches. In our work, we pre-
in the following example. sented new bodies of biomedical data in French
annotated with negation and speculation (cues and
• GOLD: Elles possèdent les caractéristiques their scope). Prior to the dissemination to the re-
de la tumeur et pourraient [permettre à search community, the French clinical trial pro-
l’avenir de faire le diagnostic de tumeur sans tocols corpus will be finalized through the inte-
biopsie ainsi que de suivre l’évolution de la gration of new data and the computation of the
tumeur traitée], les CTC disparaissant quand inter-annotator agreement. The French CAS cor-
le traitement fonctionne. pus will be distributed as more clinical cases are
• PRED: Elles possèdent les caractéristiques manually annotated, as we need more speculative
de la tumeur et pourraient [permettre à sentences to train supervised learning models on
l’avenir de faire le diagnostic de tumeur] this dataset. Another contribution of our work is
sans biopsie ainsi que de suivre l’évolution the study of different types of word vector repre-
de la tumeur traitée, les CTC disparaissant sentations and recurrent neural networks for the
quand le traitement fonctionne. detection of negation and speculation. There has
(It has the characteristics of the tumor and could in the
not been much work of this type on French cor-
future be used to diagnose the tumor without biopsy
pora, especially for the biomedical domain which
and to follow-up the evolution of the treated tumor, the
contains specific negation and speculation phe-
CTC disappearing when the treatment is efficient.)
nomena. We showed that a CRF layer yields
better performance than softmax on exact scope
Several examples show errors where both preci- match. Finally, the models have been applied in
sion and recall are impacted. Usually, the reason a cross-corpus context. Besides, we plan to im-
is that they are rare cases in our corpora where the prove our neural network performance by provid-
scope is reversed. For instance, in the example be- ing richer feature. In particular, recent embedding
low, most of the scope of devrait precedes it, while techniques, such as BERT or ELMO (Devlin et al.,
in the reference data, its scope follows this cue. 2018; Peters et al., 2018) may provide more accu-
• GOLD: L’objectif de cet essai est rate representation of the sentences. Moreover, in
d’évaluer [la diminution des complica- order to provide more accurate features, we plan
tions postopératoires de l’oesophagectomie to move from TreeTagger, which makes a substan-
qui] devrait [être obtenue] en réalisant tial number of mistakes on our datasets, to a POS
une partie de l’intervention sous coelio- tagger/lemmatizer dedicated to French biomedical
scopie, chez des patients ayant un cancer de texts. Syntactic parsing of sentences may also pro-
l’oesophage résécable. vide useful features for the detection of scope.
230
ACKNOWLEDGMENTS PL Elkin, SH Brown, BA Bauer, CS Husser, W Car-
ruth, LR Bergstrom, and DL Wahner-Roedler.
This work was partly funded by the French gov- 2005b. A controlled trial of automated classification
ernment support granted to the CominLabs LabEx of negation from clinical notes. BMC Med Inform
managed by the ANR in Investing for the Future Decis Mak. 5(13).
program under reference ANR-10-LABX-07-01. Federico Fancellu, Adam Lopez, and Bonnie Webber.
2016. Neural networks for negation scope detection.
In Proceedings of the 54th Annual Meeting of the As-
References sociation for Computational Linguistics. volume 1.
Amine Abdaoui, Andon Tchechmedjiev, William Di-
gan, Sandra Bringay, and Clement Jonquet. 2017. S Gindl, K Kaiser, and S Miksch. 2008. Syntactical
French context: Détecter la négation, la temporalité negation detection in clinical practice guidelines. In
et le sujet dans les textes cliniques français. In Stud Health Technol Inform. pages 187–92.
4ème Symposium sur l’Ingénierie de l’Information
Médicale, SIIM’17. Henk Harkema, John N Dowling, Tyler Thornblade,
and Wendy W Chapman. 2009. Context: an al-
Daniel Albright, Arrick Lanfranchi, Anwen Fredrik- gorithm for determining negation, experiencer, and
sen, William F Styler IV, Colin Warner, Jena D temporal status from clinical reports. Journal of
Hwang, Jinho D Choi, Dmitriy Dligach, Rodney D biomedical informatics 42(5):839–851.
Nielsen, James Martin, et al. 2013. Towards com-
prehensive syntactic and semantic annotations of the Sepp Hochreiter and Jürgen Schmidhuber. 1997.
clinical narrative. Journal of the American Medical Long short-term memory. Neural Comput. 9(8).
Informatics Association 20(5):922–930. https://doi.org/10.1162/neco.1997.9.8.1735.
Emilia Apostolova, Noriko Tomuro, and Dina Demner- Halil Kilicoglu and Sabine Bergler. 2010. A
Fushman. 2011. Automatic extraction of lexico- high-precision approach to detecting hedges and
syntactic patterns for detection of negation and spec- their scopes. In Proceedings of the Fourteenth
ulation scopes. In Proceedings of the 49th An- Conference on Computational Natural Language
nual Meeting of the Association for Computational Learning—Shared Task. Association for Computa-
Linguistics: Human Language Technologies: short tional Linguistics, pages 70–77.
papers-Volume 2. Association for Computational
Linguistics, pages 283–287. John Lafferty, Andrew McCallum, Fernando Pereira,
et al. 2001. Conditional random fields: Probabilistic
Piotr Bojanowski, Edouard Grave, Armand Joulin, and
models for segmenting and labeling sequence data.
Tomas Mikolov. 2017. Enriching word vectors with
In Proceedings of the eighteenth international con-
subword information. Transactions of the Associa-
ference on machine learning, ICML. volume 1.
tion for Computational Linguistics 5:135–146.
W Chapman, W Bridewell, P Hanbury, GF Cooper, and Marc Light, Xin Ying Qiu, and Padmini Srinivasan.
BG Buchanan. 2001. A simple algorithm for iden- 2004. The language of bioscience: Facts, specula-
tifying negated findings and diseases in discharge tions, and statements in between. In HLT-NAACL
summaries. J Biomed Inform 34(5):301–10. 2004 Workshop: Linking Biological Literature, On-
tologies and Databases.
Louise Deléger and Cyril Grouin. 2012. Detecting
negation of medical problems in french clinical Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Cor-
notes. In Proceedings of the 2nd ACM SIGHIT In- rado, and Jeff Dean. 2013. Distributed representa-
ternational Health Informatics Symposium. tions of words and phrases and their compositional-
ity. In Advances in neural information processing
JC Denny and JF Peterson. 2007. Identifying qt pro- systems.
longation from ecg impressions using natural lan-
guage processing and negation detection. In Med- Roser Morante and Eduardo Blanco. 2012. * sem 2012
info. pages 1283–8. shared task: Resolving the scope and focus of nega-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and tion. In * SEM 2012: The First Joint Conference
Kristina Toutanova. 2018. Bert: Pre-training of deep on Lexical and Computational Semantics–Volume 1:
bidirectional transformers for language understand- Proceedings of the main conference and the shared
ing. arXiv preprint arXiv:1810.04805 . task, and Volume 2: Proceedings of the Sixth In-
ternational Workshop on Semantic Evaluation (Se-
Peter L. Elkin, Steven H. Brown, Brent A. Bauer, mEval 2012). volume 1, pages 265–274.
Casey S. Husser, William Carruth, Larry R.
Bergstrom, and Dietlind L. Wahner-Roedler. Aurélie Névéol, Cyril Grouin, Jeremy Leixa, Sophie
2005a. A controlled trial of automated classi- Rosset, and Pierre Zweigenbaum. 2014. The quaero
fication of negation from clinical notes. BMC french medical corpus: A ressource for medical en-
medical informatics and decision making 5. tity recognition and normalization. In In Proc Bio-
https://doi.org/10.1186/1472-6947-5-13. TextM, Reykjavik. Citeseer.
231
Lilja Øvrelid, Erik Velldal, and Stephan Oepen. 2010. Özlem Uzuner, Brett R South, Shuying Shen, and
Syntactic scope resolution in uncertainty analysis. Scott L DuVall. 2011. 2010 i2b2/va challenge on
In Proceedings of the 23rd international conference concepts, assertions, and relations in clinical text.
on computational linguistics. Association for Com- Journal of the American Medical Informatics Asso-
putational Linguistics, pages 1379–1387. ciation 18(5):552–556.
Arzucan Özgür and Dragomir R Radev. 2009. Detect- Erik Velldal, Lilja Øvrelid, Jonathon Read, and
ing speculations and their scopes in scientific text. Stephan Oepen. 2012. Speculation and negation:
In Proceedings of the 2009 Conference on Empirical Rules, rankers, and the role of syntax. Computa-
Methods in Natural Language Processing: Volume tional Linguistics 38(2).
3-Volume 3. Association for Computational Linguis-
tics, pages 1398–1407. Veronika Vincze, György Szarvas, Richárd Farkas,
György Móra, and János Csirik. 2008. The bioscope
Woodley Packard, Emily M Bender, Jonathon Read, corpus: biomedical texts annotated for uncertainty,
Stephan Oepen, and Rebecca Dridan. 2014. Sim- negation and their scopes. BMC Bioinformatics 9.
ple negation scope resolution through deep parsing:
A semantic solution to a semantic problem. In Pro-
ceedings of the 52nd Annual Meeting of the Associa-
tion for Computational Linguistics (Volume 1: Long
Papers). volume 1, pages 69–78.
Yifan Peng, Xiaosong Wang, Le Lu, Mohammadhadi
Bagheri, Ronald Summers, and Zhiyong Lu. 2018.
Negbio: a high-performance tool for negation and
uncertainty detection in radiology reports. AMIA
2018 Informatics Summit .
Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt
Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. 2018. Deep contextualized word rep-
resentations. arXiv preprint arXiv:1802.05365 .
Zhong Qian, Peifeng Li, Qiaoming Zhu, Guodong
Zhou, Zhunchen Luo, and Wei Luo. 2016. Specula-
tion and negation scope detection via convolutional
neural networks. In Proceedings of the 2016 Con-
ference on Empirical Methods in Natural Language
Processing. pages 815–825.
Jonathon Read, Erik Velldal, Lilja Øvrelid, and
Stephan Oepen. 2012. Uio 1: Constituent-based dis-
criminative ranking for negation resolution. In Pro-
ceedings of the First Joint Conference on Lexical
and Computational Semantics-Volume 1: Proceed-
ings of the main conference and the shared task, and
Volume 2: Proceedings of the Sixth International
Workshop on Semantic Evaluation. Association for
Computational Linguistics, pages 310–318.
Radim Rehurek and Petr Sojka. 2010. Software frame-
work for topic modelling with large corpora. In In
Proceedings of the LREC 2010 Workshop on New
Challenges for NLP Frameworks. Citeseer.
Helmut Schmid. 1994. Probabilistic part-ofispeech
tagging using decision trees. In Proceedings of
International Conference on New Methods in Lan-
guage Processing, Manchester, UK..
Buzhou Tang, Xiaolong Wang, Xuan Wang, Bo Yuan,
and Shixi Fan. 2010. A cascade method for de-
tecting hedges and their scope in natural language
text. In Proceedings of the Fourteenth Conference
on Computational Natural Language Learning—
Shared Task. Association for Computational Lin-
guistics, pages 13–17.
232
Porting Multilingual Morphological Resources to OntoLex-Lemon
A significant number of linguistic resources have 2 The Linguistic Linked Open Data
been published on the Linguistic Linked Open Cloud
Data cloud (LLOD),1 and projects like ELEXIS2
and Prêt-à-LLOD3 are contributing to its further The LLOD initiative had its inception in 2012 at
extension. But we notice that only very few, if any, a workshop co-located with the 34th Annual Con-
specific morphological resources are included in ference of the German Linguistic Society (DGfS).
the LLOD cloud. Available morphology informa- The workshop was organized by members of the
tion is mostly contained in lexical or dictionary en- Open Knowledge Foundation,6 and the contribu-
tries. Our aim is to make also specialized morpho- tions to this workshop are available in (Chiarcos
logical resources available in this cloud. With this et al., 2012). The workshop has been a point of
step we want to support the interlinking of such focal activity for several research and infrastruc-
resources with other types of linguistic data, in a ture projects, as well as for the “Ontology Lex-
multilingual fashion, extending work described in ica” W3C Community Group.7 Those develop-
(Gromann and Declerck, 2019), which is linking ments are described in (McCrae et al., 2016). A
synsets of the Princeton WordNet (PWN) (Fell- major result of those activities is the development
baum, 1998) that are associated with plural forms of the OntoLex-Lemon model, which is described
to full lexical descriptions. in more details in Section 3.
A first step of our current work consisted We adopted OntoLex-Lemon for the representa-
in mapping the MMorph4 set of multilingual tion of morphological resources, as this model was
shown to be able to represent both classical lexico-
1
See http://www.linguistic-lod.org/ graphic description (McCrae et al., 2017) and lex-
2
https://elex.is/. See also (Krek et al., 2018) for
5
a general overview of ELEXIS and (Declerck et al., 2018) for See (Cimiano et al., 2016) and https://www.w3.
a focused description of the role of LLOD in the context of org/2016/05/ontolex/.
the project. 6
See https://okfn.org/.
3 7
https://www.pret-a-llod.eu. See for more details https://www.w3.org/
4
See (Petitpierre and Russell, 1995). community/ontolex/.
233
Proceedings of Recent Advances in Natural Language Processing, pages 233–238,
Varna, Bulgaria, Sep 2–4 2019.
ical semantics networks, like WordNet (McCrae
et al., 2014), to which we want to link full mor-
phological descriptions.
3 OntoLex-Lemon
The OntoLex-Lemon model was originally devel-
oped with the aim to provide a rich linguistic
grounding for ontologies, meaning that the natu-
ral language expressions used in the description
of ontology elements are equipped with an exten-
sive linguistic description.8 This rich linguistic Figure 1: The core module of OntoLex-Lemon: Ontol-
grounding includes the representation of morpho- ogy Lexicon Interface. Graphic taken from https:
logical and syntactic properties of lexical entries //www.w3.org/2016/05/ontolex/.
as well as the syntax-semantics interface, i.e. the
meaning of these lexical entries with respect to an
ontology or to specialized vocabularies. work can also be seen as preparing the field for a
The main organizing unit for those linguistic de- detailed representation of morphological compo-
scriptions is the lexical entry, which enables the nents of lexical data by first porting morphological
representation of morphological patterns for each resources to the core module of OntoLex-Lemon,
entry (a MWE, a word or an affix). The connection displayed in 1, before applying the representation
of a lexical entry to an ontological entity is marked guidelines of the morphology extension module,
mainly by the denotes property or is mediated which is not yet in a stable and final state.
by the LexicalSense or the LexicalConcept
4 The Morphological Resources
properties, as this is represented in Figure 1, which
displays the core module of the model. We considered two types of morphological data
OntoLex-Lemon builds on and extends the sets. One is an updated version of the multilingual
lemon model (McCrae et al., 2012b). A ma- MMorph resource (Petitpierre and Russell, 1995),
jor difference is that OntoLex-Lemon includes covering 5 languages. And we also mapped two
an explicit way to encode conceptual hierar- monolingual data sets, one for German and one for
chies, using the SKOS standard.9 As can be Italian. We will use those additional data sets for
seen in Figure 1, lexical entries can be linked, the comparison, cross-checking and merging of
via the ontolex:evokes property, to such monolingual morphological resources, using the
SKOS concepts, which can represent WordNet uniform representation of the data in OntoLex-
synsets. This structure is paralleling the relation Lemon.
between lexical entries and ontological resources,
which is implemented either directly by the 4.1 MMorph
ontolex:reference property or mediated by MMorph was originally developed by ISSCO
the instances of the ontolex:LexicalSense at University of Geneva in the past MULTEXT
class. project.11 For our purposes, we used the ex-
More recent developments of the model have tended MMorph version developed at DFKI LT
been described in (McCrae et al., 2017). Cur- Lab (MMorph3). This version includes huge mor-
rently two extension modules are being discussed: phological resources for English, French, German,
a lexicographic and a morphology module.10 Our Italian and Spanish.
8
See (McCrae et al., 2012a), (Cimiano et al., 2016) and
We choose this resource as it provides already
also https://www.w3.org/community/ontolex/ in its original format a largely unified representa-
wiki/Final_Model_Specification.
9
tion of the morphological data in the different lan-
SKOS stands for “Simple Knowledge Organization Sys-
tem”. SKOS provides “a model for expressing the basic struc- community/ontolex/wiki/Lexicography and
ture and content of concept schemes such as thesauri, clas- https://www.w3.org/community/ontolex/
sification schemes, subject heading lists, taxonomies, folk- wiki/Morphology.
sonomies, and other similar types of controlled vocabulary” 11
See https://www.issco.unige.ch/en/
(https://www.w3.org/TR/skos-primer/) research/projects/MULTEXT.html for more
10
See respectively https://www.w3.org/ details on the resulting MMorph 2.3.4 version.
234
guages, with only few differences across the dis- tion of RDF-graphs in rdf:xml, turtle sys-
tinct sources. ntax and other relevant formats.
Very generally, the MMorph tool relates a word In Listing 2 we show the resulting OntoLex-
to a morphosyntactic description (MSD) con- Lemon representation of the German noun “Aach-
taining free-definable attribute and values. The ener”.
MMorph lexicon used to realize such MSD con-
Listing 2: The OntoLex-Lemon entry for Aachener
sists of a set of lexical entries and structural rules.
For example, the following rule creates in English : lex_aachener a ontolex : LexicalEntry ;
l e x i n f o : gender l e x i n f o : masculine ;
a noun plural concatenating the singular form and l e x i n f o : p a r t O f S p e e c h l e x i n f o : noun ;
the noun suffix “s” (Petitpierre and Russell, 1995): ontolex : canonicalForm : form_aachener ;
ontolex : otherForm
" s " n o u n _ s u f f i x [ number= p l u r ] : form_aachener_dat_plural ,
: form_aachener_gen_singular ,
noun [ number= p l u r g e n d e r = $gen ] : f o r m _ a a c h e n e r _ n o m −gen−a c c _ p l u r a l .
<− noun [ number= s i n g g e n d e r = $gen ]
n o u n _ s u f f i x [ number= p l u r ] : f o r m _ a a c h e n e r a o n t o l e x : Form ;
lexinfo : case lexinfo : accusative ,
Note how the rule ensures that the gender does lexinfo : dative ,
not change in the plural form. Further adjustment lexinfo : nominative ;
l e x i n f o : number l e x i n f o : s i n g u l a r ;
rules are defined to catch the orthographic features o n t o l e x : w r i t t e n R e p " A a c h e n e r "@de .
of a specific language (e.g. box+s = boxes in En-
: form_aachener_dat_plural a
glish). o n t o l e x : Form ;
The MMorph lexica can be dumped to full form lexinfo : case lexinfo : dative ;
lists for the usage in further programs: l e x i n f o : number l e x i n f o : p l u r a l ;
o n t o l e x : w r i t t e n R e p " A a c h e n e r n "@de .
Listing 1: The MMorph entry for the German noun : form_aachener_gen_singular a
“Aachener” (inhabitant of Aachen) o n t o l e x : Form ;
" a a c h e n e r " = " a a c h e n e r " Noun [ g e n d e r =masc lexinfo : case lexinfo : gen iti ve ;
number= s i n g u l a r c a s e =nom | d a t | a c c ] l e x i n f o : number l e x i n f o : s i n g u l a r ;
" a a c h e n e r " = " a a c h e n e r " Noun [ g e n d e r =masc o n t o l e x : w r i t t e n R e p " A a c h e n e r s "@de .
number= p l u r a l c a s e =nom | gen | a c c ]
" a a c h e n e r n " = " a a c h e n e r " Noun [ g e n d e r =masc : f o r m _ a a c h e n e r _ n o m −gen−a c c _ p l u r a l a
number= p l u r a l c a s e = d a t ] o n t o l e x : Form ;
" a a c h e n e r s " = " a a c h e n e r " Noun [ g e n d e r =masc lexinfo : case lexinfo : accusative ,
number= s i n g u l a r c a s e = gen ] lexinfo : genitive ,
lexinfo : nominative ;
l e x i n f o : number l e x i n f o : p l u r a l ;
As the reader can observe in Listing 1, the nomi- o n t o l e x : w r i t t e n R e p " A a c h e n e r "@de .
nal entries are completed by appropriate features
describing case, gender, and number. Multiple The reader can observe how the relations be-
values of a feature are expressed by “|”. The tween a lemma (an instance of the class
user can freely define language- and word class- LexicalEntry) and its different morpholog-
specific features (e.g. clitics for verbal entries or ical forms (instances of the class Form) is made
rection of prepositions). As the example above explicit by the use of named properties. Another
demonstrates, the dumped lexica are ideally suited feature of our work is the re-use of established vo-
for the mapping into the OntoLex-Lemon format, cabularies, for example the LexInfo vocabulary13
as they present their data in a well structured fash- to represent the morpho-syntactic features.
ion. In Listing 3, we show examples of the resulting
Our German version of MMorph contains over data for the lemma “cura” in Spanish.
2.630.000 full-forms. Compared to the original
Listing 3: The OntoLex-Lemon entry for cura
version, it has specifically improved the coverage
: lex_cura_1 a ontolex : LexicalEntry ;
of compounds. lexinfo : gender l e x i n f o : feminine ;
To transform the MMorph data into OntoLex- lexinfo : p a r t O f S p e e c h l e x i n f o : noun ;
Lemon we used a Python script including the ontolex : canonicalForm : form_cura ;
ontolex : otherForm : f o r m _ c u r a _ p l u r a l .
rdflib module12 , which supports the genera-
12 13
See https://github.com/RDFLib/rdflib for See https://lexinfo.net/ and (Cimiano et al.,
more details. 2011) for more details
235
: lex_cura_2 a ontolex : LexicalEntry ; 4.2 Two Monolingual Resources
lexinfo : gender l e x i n f o : masculine ;
lexinfo : p a r t O f S p e e c h l e x i n f o : noun ; Other data sets we are considering are “Morph-
ontolex : canonicalForm : form_cura ; it!”18 for Italian and the “DE_morph_dict data”19 .
ontolex : otherForm : f o r m _ c u r a _ p l u r a l .
An entry in Morph-it! has the form displayed in
: f o r m _ c u r a a o n t o l e x : Form ; Listing 4:
l e x i n f o : number l e x i n f o : s i n g u l a r ;
o n t o l e x : w r i t t e n R e p " c u r a " @es . Listing 4: The Morph-it! entry for “abbassamento”
: f o r m _ c u r a _ p l u r a l a o n t o l e x : Form ; (lowering or reduction)
l e x i n f o : number l e x i n f o : p l u r a l ; abbassamento abbassamento NOUN−M: s
o n t o l e x : w r i t t e n R e p " c u r a s " @es . abbassamenti abbassamento NOUN−M: p
The reader can also see the harmonized repre- Listing 5: The OntoLex-Lemon representation for “ab-
sentation of morphological resources across lan- bassamento” (lowering or reduction)
guages (here German and Spanish). This is an : lex_abbassamento a
ontolex : LexicalEntry ;
important feature that will allow to link various l e x i n f o : gender l e x i n f o : masculine ;
lemmas (or senses) from different languages to l e x i n f o : p a r t O f S p e e c h l e x i n f o : noun ;
a unique reference point in external information ontolex : canonicalForm
: form_abbassamento ;
sources, like WordNet(s)15 or knowledge Graphs, ontolex : otherForm
like DBpedia16 or Wikidata17 . : form_abbassamento_m_p .
236
Listing 6: The DE_morph_dict entry for “Abgang” (de- Group and publishing the results in an accessible
parture or leaving etc) subset of the Linked Data cloud.
Abgang
Abgang NN, masc , acc , sing 5 Conclusion
Abgang NN, masc , nom , sing
Abgang NN, masc , d a t , sing We described our current work consisting in port-
Abgange
Abgang NN, masc , d a t , ing , old ing a number of (multilingual) morphological re-
Abganges sources to OntoLex-Lemon, in order to harmo-
Abgang NN, masc , gen , sing
Abgangs
nize those and to support their interlinking, cross-
Abgang NN, masc , gen , sing checking, but also their linking with other data
Abgä nge source in the Linguistic Linked Open Data, as for
Abgang NN, masc , nom , plu
Abgang NN, masc , acc , plu examples WordNets, or with data sets included in
Abgang NN, masc , gen , plu knowledge graphs, like DBpedia or Wikidata.
Abgä ngen As a final goal of our work, we see the possibil-
Abgang NN, masc , d a t , plu
ity to interlink or merge those morphological re-
The mapping of this entry to OntoLex-Lemon re- sources in the Linguistic Linked Open Data cloud.
sults in a representation that is by now familiar,
and which is given in Listing 7. Acknowledgments
Listing 7: The OntoLex-Lemon representation for “Ab- Work presented in this paper has been supported
gang” (departure or leaving etc) in part by the H2020 project “Prêt-à-LLOD” with
: lex_abgang a ontolex : LexicalEntry ; Grant Agreement number 825182. Contributions
l e x i n f o : gender l e x i n f o : masculine ;
l e x i n f o : p a r t O f S p e e c h l e x i n f o : noun ; by Thierry Declerck are also supported in part by
o n t o l e x : canonicalForm : form_abgang ; the H2020 project “ELEXIS” with Grant Agree-
ontolex : otherForm ment number 731015.
: form_abgang_dat_plu ,
: form_abgang_dat_sing ,
: form_abgang_gen_sing ,
: form_abgang_nom−gen−a c c _ p l u . References
: f o r m _ a b g a n g a o n t o l e x : Form ; Duygu Altinok. 2018. Demorphy, german language
lexinfo : case lexinfo : accusative , morphological analyzer. CoRR, abs/1803.00902.
lexinfo : dative ,
lexinfo : nominative ; Francis Bond and Kyonghee Paik. 2012. A survey of
l e x i n f o : number l e x i n f o : s i n g u l a r ; wordnets and their licenses. In Proceedings of the
o n t o l e x : w r i t t e n R e p " Abgang "@de . 6th Global WordNet Conference (GWC 2012), pages
64–71, Matsue.
: f o r m _ a b g a n g _ d a t _ p l u a o n t o l e x : Form ;
lexinfo : case lexinfo : dative ; Christian Chiarcos, Sebastian Nordhoff, and Sebastian
l e x i n f o : number l e x i n f o : p l u r a l ; Hellmann, editors. 2012. Linked Data in Linguistics
o n t o l e x : w r i t t e n R e p " Abgä ngen "@de . - Representing and Connecting Language Data and
Language Metadata. Springer.
( etc . )
With those resources represented in OntoLex- Philipp Cimiano, Paul Buitelaar, John McCrae, and
Michael Sintek. 2011. Lexinfo: A declarative model
Lemon, which are duplicating the German and for the lexicon-ontology interface. Journal of Web
Italian resources we have already from MMorph, Semantics: Science, Services and Agents on the
we aim at discovering possible inconsistencies World Wide Web, 9(1):29–51.
or similarities within resources for one language,
Philipp Cimiano, John P. McCrae, and Paul Buitelaar.
which could lead to both a improvement and a 2016. Lexicon Model for Ontologies: Community
merging of the original resources. Report.
We are in a sense extending a former experi-
Thierry Declerck, John McCrae, Roberto Navigli, Kse-
ment on automatically merging Italian morpholog- nia Zaytseva, and Tanja Wissik. 2018. Elexis - euro-
ical resources in the context of a finite automata pean lexicographic infrastructure: Contributions to
environment, and which is described in (Declerck and from the linguistic linked open data. In Pro-
et al., 2012). The new work is not only a multilin- ceedings of the 2nd GLOBALEX Workshop. GLOB-
ALEX (GLOBALEX-2018), Lexicography & Word-
gual extension, but is aiming at a broad interoper- Net, located at 11th Language Resources and Eval-
ability of morphological resources by using a de- uation Conference (LREC 2018), May 8, Miyazaki,
facto standard developed by a W3C Community Japan, pages 17–22. ELRA.
237
Thierry Declerck, Stefania Racioppa, and Karlheinz Multext deliverable 2.3.1, ISSCO, University of
Mörth. 2012. Automatized merging of italian lexical Geneva.
resources. In Proceeding of the LREC 2012 Work-
shop on Language Resource Merging, Paris. ELRA, Eros Zanchetta and Marco Baroni. 2005. Morph-it!
ELRA. a free corpus-based morphological resource for the
italian language. In Proceedings of Corpus Linguis-
Christiane Fellbaum, editor. 1998. WordNet: An Elec- tics 2005. University of Birmingham.
tronic Lexical Database. The MIT Press, Cam-
bridge, MA ; London.
238
Dependency-Based Self-Attention for Transformer NMT
239
Proceedings of Recent Advances in Natural Language Processing, pages 239–246,
Varna, Bulgaria, Sep 2–4 2019.
Self-attention Because the Transformer model does not in-
+ Encoder- clude recurrent or convolutional structures, it en-
Jd
Decoder attention codes word positional information as sinusoidal
Self-attention + FF positional encodings:
Je ..
+ FF .
.. P(pos,2i) = sin(pos/100002i/d ), (1)
. ..
. cos(pos/100002i/d ), (2)
.. P(pos,2i+1) =
. ..
.
..
. Self-attention where pos is the position, i is the dimension, and
Self-attention + Encoder- d is the dimension of the intermediate representa-
1 1
+ FF Decoder attention tion. At the first layers of the encoder and decoder,
+ FF the positional encodings calculated by Equations
Encoder Layers Decoder Layers (1) and (2) are added to the input embeddings.
(j)
The j-th encoder layer’s output Senc is gener-
Figure 1: Transformer model. ated by a self-attention layer Self Attn() and a
position-wise fully connected feed forward net-
work layer F F N () as follows:
word sequence rather than a word sequence to ad-
dress the translation of out-of-vocabulary words (j)
Henc = LN (Senc
(j−1) (j−1)
+ Self Attn(Senc )), (3)
(Sennrich et al., 2016). Therefore, we extend (j) (j) (j)
Senc = LN (Henc + F F N (Henc )), (4)
dependency-based self-attention to operate at sub-
word units created by byte pair encoding (BPE) (0) (j)
where Senc is the input of the encoder, Henc
rather than word-units.
is the output of the j-th encoder’s self-attention,
Our experiments demonstrate that the proposed
and LN () is layer normalization (Lei Ba et al.,
Transformer NMT model performs 1.0 BLEU (j)
points higher than the baseline Transformer NMT 2016). The j-th decoder layer’s output Sdec is
model, which does not incorporate dependency generated by an encoder-decoder attention layer
structures, on the WAT’18 Asian Scientific Pa- EncDecAttn() in addition to the two sublayers
per Excerpt Corpus (ASPEC) Japanese-to-English of the encoder (i.e., Self Attn() and F F N ()) as
translation task. The experiments also demon- follows:
strate the effectiveness of each of our propos- (j) (j−1) (j−1)
Hdec = LN (Sdec + Self Attn(Sdec )), (5)
als, namely, encoder-side dependency-based self-
(j) (j) (j)
attention, decoder-side dependency-based self- Gdec = LN (Hdec + EncDecAttn(Hdec )), (6)
attention, and extension for BPE. (j)
Sdec =
(j) (j)
LN (Gdec + F F N (Hdec )), (7)
2 Transformer NMT (0) (j)
where Sdec is the input of the decoder, Hdec is
We provide here an overview of the Transformer the output of the j-th decoder’s self-attention, and
(j)
NMT model (Vaswani et al., 2017), which is the Gdec is the output of the j-th decoder’s encoder-
basis of our proposed model. The outline of the decoder attention.
(J )
Transformer NMT model is shown in Fig. 1. The last decoder layer’s output Sdecd is linearly
The Transformer NMT model is an encoder- mapped to a V -dimensional matrix, where V is
decoder model that has a self-attention mech- the output vocabulary size. Then, the output se-
anism. The encoder maps an input sequence quence Y is generated based on P (Y | X), which
of symbol representations (i.e., a source sen- is calculated by applying the softmax function to
tence) X = (x1 , x2 , . . . , xnenc )T to an inter- the V -dimensional matrix.
mediate vector. Then, the decoder generates an Self-attention computes the strength of the re-
output sequence (i.e., a target sentence) Y = lationship between two words in the same sen-
(y1 , y2 , . . . , yndec )T , given the intermediate vec- tence (i.e., between two source words or be-
tor. The encoder and the decoder are composed tween two target words), and encoder-decoder
of a stack of Je encoder layers and of Jd decoder attention computes the strength of the relation-
layers, respectively. ship between a source word and a target word.
240
Both the self-attention and encoder-decoder at- Self-attention
tention are implemented with multi-head atten- + Encoder-
Jd
tion, which projects the embedding space into Decoder attention
nhead subspaces of the dhead = d/nhead dimen- + FF
sion and calculates attention in each subspace. In ..
.
the j-th layer’s self-attention, the previous layer’s Self-attention Dependency-
J
output S (j−1) ∈ Rn×d is linearly mapped to e + FF based
three dhead -dimensional subspaces, Qh , Kh , ...
(j) (j)
self-attention pd
(j) (j)
and Vh , using parameter matrices WhQ ∈ Dependency- + Encoder-
R head , Wh
d×d K (j)
∈ R head , and Wh
d×d V (j)
∈ pe based Decoder attention
R d×d head , where n is the length of the input se- self-attention + FF
+ FF ..
quence and 1 ≤ h ≤ nhead 1 . In the j-th de- .
.
coder layer’s encoder-decoder attention, the pre- .. Self-attention
(j−1) (j)
vious layer’s output Sdec is mapped to Qh , and 1 Self-attention + Encoder-
1
(J )
the last encoder layer’s output Sence is mapped to + FF Decoder attention
(j) (j) + FF
Kh and Vh .
Then, an attention weight matrix, where each Encoder Layers Decoder Layers
value represents the strength of the relationship
between two words, is calculated on each sub-
Figure 2: Proposed model.
space as follows:
(j) (j) (j)T
Ah = sof tmax(d−0.5
head Qh Kh ). (8) lations into self-attention on both source and tar-
(j) (j) get sides, dependency-based self-attention. In
By multiplying Ah and Vh , a weighted repre-
(j) particular, it parses the dependency structures of
sentation matrix Mh is obtained: source sentences and target sentences by one at-
(j) (j) (j) tention head of the pe -th encoder layer’s multi-
Mh = Ah Vh . (9)
head self-attention and one of the pd -th decoder
(j)
Mh in self-attention includes the strengths of layer’s multi-head self-attention, respectively, and
the relationships with all words in the same sen- translates a sentence based on the source-side
(j)
tence for each source or target word, and Mh in and target-side dependency structures. We use
encoder-decoder attention includes the strengths the deep bi-affine parser (Dozat and Manning,
of the relationships with all source words for each 2016) as a model for dependency parsing in
target word. the dependency-based self-attention according to
(j) LISA. There are two inherent differences between
Finally, the concatenation of all Mh (i.e.,
(j) LISA and our dependency-based self-attention: (i)
M1,2,...,nhead ) is mapped to a d-dimensional ma- our decoder-side dependency-based self-attention
trix M (j) as follows: masks future information on words, and (ii) our
(j) (j) dependency-based self-attention operates at sub-
M (j) = W M [M1 ; . . . ; Mn(j) ], (10)
head
word units created by byte pair encoding rather
(j)
where W M ∈ Rd×d is a parameter matrix. than word-units.
Note that, in training, the decoder’s self-
attention masks future words so as to ensure that 3.1 Dependency-Based Self-Attention
the attentions of a target word do not rely on un-
The dependency-based self-attention parses de-
predicted words in inference.
pendency structures by extending the multi-head
3 Proposed Method self-attention of the p-th layer of the encoder or
decoder2 . First, the p-th self-attention layer maps
Figure 2 shows the outline of the proposed model. the previous layer’s output S (p−1) of d-dimension
The proposed model incorporates dependency re- to dhead -dimensional subspaces of multi-head at-
1 (j) (j) (j)
S indicates Senc for the encoder and Sdec for the de-
2
coder. p indicates pe for the encoder and pj for the decoder.
241
tention as follows:
242
sentence pairs encoder and decoder and set nhead = 8 and
train 1,198,149 d = 512. For the proposed model, we incor-
dev 1,790 porated dependency-based self-attention into the
test 1,812 fourth layer in both the encoder and the decoder
(i.e., pe = pd = 4).
Table 1: Statistics of the ASPEC data.
We evaluated translation performance on
Model BLEU the WAT’18 ASPEC (Nakazawa et al., 2016)
Trans. 27.29 Japanese-to-English translation task. We to-
Trans. + DBSA(Enc) 28.05 kenized each Japanese sentence with KyT ea
Trans. + DBSA(Dec) 27.86 (Neubig et al., 2011) and preprocessed according
Trans. + DBSA(Enc) + DBSA(Dec) 28.29 to the recommendations from WAT’184 . We used
the vocabulary of 100K subword tokens based
Table 2: Translation performance. on BPE for both languages and used the first
1.5 million translation pairs as the training data.
element. As shown, future information on each In the training, long sentences with over 250
word is masked. For example, the dependency re- subword-tokens were filtered out. Table 1 shows
lation from “are” to “explained” is masked. the statistics of our experiment data.
We used Japanese dependency structures gen-
3.3 Subword Dependency-Based erated by EDA5 and English dependency struc-
Self-Attention tures generated by Stanford Dependencies6 in the
Recent NMT models have improved the trans- training of the source-side dependency-based self-
lation performance by treating a sentence as a attention and the target-side dependency-based
subword sequence rather than a word sequence. self-attention, respectively. Note that Stanford De-
Therefore, we extend dependency-based self- pendencies and EDA are not used in the testing.
attention to work for subword sequences. In our
subword dependency-based self-attention, a sen- 4.2 Training Details
tence is divided into a subword sequence by BPE We trained each model using Adam
(Sennrich et al., 2016). When a word is divided (Kingma and Ba, 2014), where the learning
into multiple subwords, the modifiee (i.e., the rate and hyperparameter settings are set following
head) of the rightmost subword is set to the modi- Vaswani et al. (2017). For the objective function,
fiee of the original word and the modifiee of each we set ϵls (Szegedy et al., 2016) in label smooth-
subword other than the rightmost one is set to the ing to 0.1 and both the hyperparameters λenc
right adjacent subword. and λdec to 1.0. We set the mini-batch size to
Figure 4 shows an example of subword-level 224 and the number of epochs to 20. We chose
dependency relations, where “@@” is a sub- the model that achieved the best BLEU score on
word segmentation symbol. “Fingerprint” is the development set and evaluated the sentences
divided into the three subwords: “Fing@@”, generated from the test set using beam search
“er@@“, and “print”. When the head of the with a beam size of 4 and length penalty α = 0.6
word “Fingerprint” is “input” in the original (Wu et al., 2016).
word-level sentence, the heads of the three sub-
words are determined as follows: “er@@” = 4.3 Experiment Results
head(“Fing@@”), “print” = head(“er@@”), and Table 2 lists the experiment results.
“input” = head(“print”). Translation performance is measured by
BLEU (Papineni et al., 2002). In Table 2,
4 Experiments DBSA denotes our dependency-based self-
4.1 Experiment Settings attention. As shown, our proposed model
In our experiments, we compared the proposed 4
http://lotus.kuee.kyoto-u.ac.jp/WAT/
model with a conventional Transformer NMT WAT2018/baseline/dataPreparationJE.html
5
model, which does not incorporate dependency http://www.ar.media.kyoto-u.ac.jp/
tool/EDA
structures, to confirm the effectiveness of the pro- 6
https://nlp.stanford.edu/software/
posed model. We stacked six layers for each stanford-dependencies.html
243
Fing@@ er@@ print is input as an image .
244
dependency-based self-attention. Our decoder- Diederik P Kingma and Jimmy Ba. 2014. Adam: A
side dependency-based self-attention masks future method for stochastic optimization. arXiv preprint
arXiv:1412.6980 .
information to avoid conflicts between training
and inference. In addition, our dependency-based Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin-
self-attention is extended to work well for sub- ton. 2016. Layer normalization. arXiv preprint
word sequences. Experimental results showed that arXiv:1607.06450 .
the proposed model achieved a 1.0 point gain in Dekang Lin. 2004. A path-based transfer model for
BLEU over the baseline Transformer model on machine translation. In Proceedings of the 20th In-
the WAT’18 ASPEC Japanese-English translation ternational Conference on Computational Linguis-
tics. pages 625–630.
task. In future work, we will explore the effective-
ness of our proposed method for language pairs Toshiaki Nakazawa, Manabu Yaguchi, Kiyotaka Uchi-
other than Japanese-to-English. moto, Masao Utiyama, Eiichiro Sumita, Sadao
Kurohashi, and Hitoshi Isahara. 2016. Aspec: Asian
scientific paper excerpt corpus. In Proc. of LREC
Acknowledgement 2016.
The research results have been achieved by “Re- Graham Neubig, Yosuke Nakata, and Shinsuke Mori.
search and Development of Deep Learning Tech- 2011. Pointwise prediction for robust, adaptable
nology for Advanced Multilingual Speech Trans- Japanese morphological analysis. In Proceedings
of the 49th Annual Meeting of the Association for
lation”, the Commissioned Research of National Computational Linguistics: Human Language Tech-
Institute of Information and Communications nologies. Association for Computational Linguis-
Technology (NICT) , JAPAN. This work was par- tics, pages 529–533.
tially supported by JSPS KAKENHI Grant Num-
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-
ber JP18K18110. Jing Zhu. 2002. Bleu: A method for automatic eval-
uation of machine translation. In Proceedings of the
40th Annual Meeting on Association for Computa-
References tional Linguistics. ACL ’02, pages 311–318.
Kehai Chen, Rui Wang, Masao Utiyama, Lemao Liu, Rico Sennrich, Barry Haddow, and Alexandra Birch.
Akihiro Tamura, Eiichiro Sumita, and Tiejun Zhao. 2016. Neural machine translation of rare words with
2017. Neural machine translation with source de- subword units. In Proceedings of the 54th Annual
pendency representation. In Proceedings of the Meeting of the Association for Computational Lin-
2017 Conference on Empirical Methods in Natural guistics (Volume 1: Long Papers). Association for
Language Processing. pages 2846–2852. Computational Linguistics, Berlin, Germany, pages
1715–1725.
Timothy Dozat and Christopher D Manning. 2016.
Deep biaffine attention for neural dependency pars- Emma Strubell, Patrick Verga, Daniel Andor,
ing. arXiv preprint arXiv:1611.01734 . David Weiss, and Andrew McCallum. 2018.
Linguistically-informed self-attention for semantic
Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, role labeling. In Proc. of EMNLP 2018.
and Noah A. Smith. 2016. Recurrent neural net- Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
work grammars. In Proceedings of the 2016 Con- Sequence to Sequence Learning with Neural Net-
ference of the North American Chapter of the Asso- works. In Z Ghahramani, M Welling, C Cortes, N D
ciation for Computational Linguistics: Human Lan- Lawrence, and K Q Weinberger, editors, Advances
guage Technologies. Association for Computational in Neural Information Processing Systems 27, Cur-
Linguistics, San Diego, California, pages 199–209. ran Associates, Inc., pages 3104–3112.
Akiko Eriguchi, Yoshimasa Tsuruoka, and Kyunghyun Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe,
Cho. 2017. Learning to parse and translate improves Jon Shlens, and Zbigniew Wojna. 2016. Rethinking
neural machine translation. In Proceedings of the the inception architecture for computer vision. In
55th Annual Meeting of the Association for Compu- The IEEE Conference on CVPR.
tational Linguistics (Volume 2: Short Papers). pages
72–78. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Jonas Gehring, Michael Auli, David Grangier, De- Kaiser, and Illia Polosukhin. 2017. Attention is all
nis Yarats, and Yann N Dauphin. 2017. Convolu- you need. In I Guyon, U V Luxburg, S Bengio,
tional sequence to sequence learning. In Proceed- H Wallach, R Fergus, S Vishwanathan, and R Gar-
ings of the 34th International Conference on Ma- nett, editors, Advances in Neural Information Pro-
chine Learning-Volume 70. JMLR. org, pages 1243– cessing Systems 30, Curran Associates, Inc., pages
1252. 5998–6008.
245
S. Wu, D. Zhang, Z. Zhang, N. Yang, M. Li,
and M. Zhou. 2018. Dependency-to-dependency
neural machine translation. IEEE/ACM Transac-
tions on Audio, Speech, and Language Processing
26(11):2132–2141.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V
Le, Mohammad Norouzi, Wolfgang Macherey,
Maxim Krikun, Yuan Cao, Qin Gao, Klaus
Macherey, et al. 2016. Google’s neural ma-
chine translation system: Bridging the gap between
human and machine translation. arXiv preprint
arXiv:1609.08144 .
246
Detecting Toxicity in News Articles: Application to Bulgarian
247
Proceedings of Recent Advances in Natural Language Processing, pages 247–258,
Varna, Bulgaria, Sep 2–4 2019.
While most previous research has focused almost The interested reader can learn more about “fake
exclusively on English, here we target Bulgar- news” from the overview by Shu et al. (2017),
ian. In particular, we built a dataset for our ex- which adopted a data mining perspective and fo-
periments based on the knowledge base of Me- cused on social media. Another recent survey
dia Scan, which has catalogued and character- (Thorne and Vlachos, 2018) took a fact-checking
ized many of the Bulgarian online media in the perspective on “fake news” and related problems.
past five years. If a medium published a toxic Yet another survey was performed by Li et al.
news, this was recorded and the article, as well (2016), and it covered truth discovery in gen-
as the medium, got labelled accordingly. The ana- eral. Moreover, there were two recent articles
lyzed media vary from digital newspapers, to me- in Science: Lazer et al. (2018) offered a general
dia groups and blogs. For some articles there is de- overview and discussion on the science of “fake
tailed explanation with examples about why they news”, while Vosoughi et al. (2018) focused on
were labelled like that. In some cases, the Me- the proliferation of true and false news online.
dia Scan website describes attempts to contact the The veracity of information has been stud-
authors of an article asking for clarification about ied at different levels: (i) claim (e.g., fact-
some questionable facts that are being reported. checking), (ii) article (e.g., “fake news” detection),
Here we use this information by performing (iii) user (e.g., hunting for trolls), and (iv) medium
multi-class classification over the toxicity labels: (e.g., source reliability estimation). Our primary
fake news, sensations, hate speech, conspiracies, interest here is at the article-level.
anti-democratic, pro-authoritarian, defamation,
delusion. Note that we allow multiple of these la- 2.1 Fact-Checking
bels simultaneously. We further add a non-toxic
label for articles that represent good news. At the claim-level, fact-checking and rumor de-
tection have been primarily addressed using infor-
2 Related Work mation extracted from social media, i.e., based on
how users comment on the target claim (Canini
The proliferation of false information has attracted et al., 2011; Castillo et al., 2011; Ma et al., 2016,
a lot of research interest recently. This includes 2017; Dungs et al., 2018; Kochkina et al., 2018).
challenging the truthiness of news (Brill, 2001; The Web has also been used as a source of in-
Hardalov et al., 2016; Potthast et al., 2018), of formation (Mukherjee and Weikum, 2015; Popat
news sources (Baly et al., 2018a, 2019; Dinkov et al., 2016, 2017; Karadzhov et al., 2017b; Mi-
et al., 2019), and of social media posts (Canini haylova et al., 2018; Baly et al., 2018b; Zlatkova
et al., 2011; Castillo et al., 2011), as well as et al., 2019).
studying credibility, influence, bias, and propa- In both cases, the most important information
ganda (Ba et al., 2016; Chen et al., 2013; Mihaylov sources are stance (does a tweet or a news article
et al., 2015a; Kulkarni et al., 2018; Baly et al., agree or disagree with the claim?), and source re-
2018a; Mihaylov et al., 2018; Barrón-Cedeño liability (do we trust the user who posted the tweet
et al., 2019; Da San Martino et al., 2019; Zhang or the medium that published the news article?).
et al., 2019). Research was facilitated by shared Other important sources are linguistic expression,
tasks such as the SemEval 2017 and 2019 tasks on meta information, and temporal dynamics.
Rumor Detection (Derczynski et al., 2017; Gor-
rell et al., 2019), the CLEF 2018 and 2019 Check-
2.2 Stance Detection
That! labs (Nakov et al., 2018; Elsayed et al.,
2019b,a), which featured tasks on automatic iden- Stance detection has been addressed as a task in
tification (Atanasova et al., 2018, 2019) and ver- its own right, where models have been devel-
ification (Barrón-Cedeño et al., 2018; Hasanain oped based on data from the Fake News Chal-
et al., 2019) of claims in political debates, the lenge (Riedel et al., 2017; Thorne et al., 2017;
FEVER 2018 and 2019 task on Fact Extraction Mohtarami et al., 2018; Hanselowski et al., 2018),
and VERification (Thorne et al., 2018), and the or from SemEval-2017 Task 8 (Derczynski et al.,
SemEval 2019 task on Fact-Checking in Com- 2017; Dungs et al., 2018). It has also been studied
munity Question Answering Forums (Mihaylova for other languages such as Arabic (Darwish et al.,
et al., 2019), among others. 2017b; Baly et al., 2018b; Mohtarami et al., 2019).
248
2.3 Source Reliability Estimation They found that “fake news” pack a lot of infor-
Unlike stance detection, the problem of source mation in the title (as the focus is on users who
reliability remains largely under-explored. In do not read beyond the title), and use shorter, sim-
the case of social media, it concerns modeling pler, and repetitive content in the body (as writing
the user1 who posted a particular message/tweet, fake information takes a lot of effort). Thus, they
while in the case of the Web, it is about the trust- argued that the title and the body should be ana-
worthiness of the source (the URL domain, the lyzed separately.
medium). In follow-up work, Horne et al. (2018b) created
The source reliability of news media has often a large-scale dataset covering 136K articles from
been estimated automatically based on the general 92 sources from opensources.co, which they
stance of the target medium with respect to known characterize based on 130 features from seven cat-
manually fact-checked claims, without access to egories: structural, sentiment, engagement, topic-
gold labels about the overall medium-level factu- dependent, complexity, bias, and morality. We use
ality of reporting (Mukherjee and Weikum, 2015; this set of features when analyzing news articles.
Popat et al., 2016, 2017, 2018). The assumption is In yet another follow-up work, Horne et al.
that reliable media agree with true claims and dis- (2018a) trained a classifier to predict whether a
agree with false ones, while for unreliable media given news article is coming from a reliable or
it is mostly the other way around. The trustwor- from an unreliable (“fake news” or conspiracy)2
thiness of Web sources has also been studied from source. Note that they assumed that all news from
a Data Analytics perspective. For instance, Dong a given website would share the same reliability
et al. (2015) proposed that a trustworthy source is class. Such an assumption is fine for training (dis-
one that contains very few false facts. tant supervision), but it is problematic for testing,
Note that estimating the reliability of a source where manual documents-level labels are needed.
is important not only when fact-checking a claim Potthast et al. (2018) used 1,627 articles from
(Popat et al., 2017; Nguyen et al., 2018), but nine sources, whose factuality has been manu-
such reliability scores can be used as an impor- ally verified by professional journalists from Buz-
tant prior when addressing article-level factuality zFeed. They applied stylometric analysis, which
tasks such as “fake news” and click-bait detec- was originally designed for authorship verifica-
tion (Brill, 2001; Hardalov et al., 2016; Karadzhov tion, to predict factuality (fake vs. real).
et al., 2017a; De Sarkar et al., 2018; Pérez-Rosas Rashkin et al. (2017) focused on the language
et al., 2018). used by “fake news” and compared the prevalence
of several features in articles coming from trusted
2.4 “Fake News” Detection sources vs. hoaxes vs. satire vs. propaganda.
Most work on “fake news” detection has relied on However, their linguistic analysis and their auto-
medium-level labels, which were then assumed to matic classification were at the article level and
hold for all articles from that source. they only covered eight news media sources.
Horne and Adali (2017) analyzed three small
datasets ranging from a couple of hundred to a few 2.5 Work for Bulgarian
thousand articles from a couple of dozen sources, We are aware of only one piece of previous work
comparing (i) real news vs. (ii) “fake news” vs. for Bulgarian that targets toxicity. In particular,
(iii) satire, and found that the latter two have a lot (Karadzhov et al., 2017a) built a fake news and
in common across a number of dimensions. They click-bait detector for Bulgarian based on data
designed a rich set of features that analyze the text from a hackaton.
of a news article, modeling its complexity, style,
While most of the above research has focused
and psychological characteristics.
on isolated and specific task (such as trustworthi-
1
User modeling in social media and news community fo- ness, fake news, fact-checking), here we try to cre-
rums has focused on finding malicious users such as opinion ate a holistic approach by exploring several toxic
manipulation trolls, paid (Mihaylov et al., 2015b) or just per- and non-toxic labels simultaneously.
ceived (Mihaylov et al., 2015a; Mihaylov and Nakov, 2016;
Mihaylov et al., 2018; Mihaylova et al., 2018), sockpuppets
2
(Maity et al., 2017), Internet water army (Chen et al., 2013), We show in parentheses, the labels from
and seminar users (Darwish et al., 2017a). opensources.co that are used to define a category.
249
Characteristic Value In addition to this dataset of only toxic articles,
we added some “non-toxic” articles, fetched from
Toxic articles 221
media without toxicity examples in Media Scan:
Non-toxic articles 96
we added a total of 96 articles from 25 media.
Media 164
Table 1 shows statistics about the dataset, and
Average title length (chars) 70.47
Figure 1 shows the distribution of the labels.
Average title length (words) 11.08
Average text length (chars) 3,613.70 4 Method
Average text length (words) 556.83
Average text length (sentences) 31.64 We used a feature-rich classifier based on logistic
regression and a neural network.
Table 1: Statistics about the dataset. For each article, we extracted its title and its
body. We further extracted some meta information
100 about the corresponding news medium. As some
80 NLP resources are only available or are better for
English, we translated the articles to English, by
Articles count
60
using Google Translate API, so that we can ex-
40 tract features from them as explained in subsec-
20 tions 4.2, 4.6, 4.7, and 4.8.
0 4.1 LSA
non-toxic
conspiracy
defamation
anti-democratic
pro-authoritarian
fake news
hate speech
delusion
sensation
250
4.3 Stylometric Features 4.4 Media Features
For the title and the body of each article, we cal- We further extracted binary and numerical features
culate the following features: characterizing the medium the article came from:
• editor: its value is 1 if the target medium has
• avg word length title: average length of the
a designated chief editor, and it is 0 other-
word in the article title;
wise;
• avg word length text: average length of the • responsible person: its value is 1 if the target
words in the article body medium has a responsible person, and it is 0
otherwise;
• word count title: number of words in the ar-
ticle title; • bg server: its value is 1 if the target
medium’s location is in Bulgaria, and it is 0
• word count text: number of words in the ar- otherwise;
ticle body;
• popularity: reciprocal value of the target
• char count title: number of characters in the medium’s rank the Web traffic analysis plat-
article title; form Alexa5 ;
• domain person: its value is 1 if the target
• char count text: number of characters in the
medium has a designated owner, and it is 0
article body;
otherwise;
• spec char count title: number of specific • days existing: number of days between when
(non-alpha-numeric) characters in the article the medium was created and 01.01.2019. As
title; this value is quite large, we take the logarithm
thereof. For example, a medium created on
• spec char count text: number of specific Januarty 1, 2005 would have 5,113 days of
(non-alpha-numeric) characters in the article existence, which would correspond to 3.70 as
body; this feature’s value.
• upper char count title: number of uppercase 4.5 XLM
characters in the article title; We further used cross-lingual representations from
the Facebook’s XLM model (Lample and Con-
• upper char count text: number of uppercase
neau, 2019), which creates cross-lingual repre-
characters in the article body;
sentations based on the Transformer, similarly to
BERT. This model is pretrained for 15 languages
• upper word count title: number of words
including Bulgarian and English. We use their pre-
starting with an uppercase character in the ar-
trained models, which were fine-tuned for Cross-
ticle title;
lingual Natural Language Interence (XNLI) tasks.
• upper word count text: number of words This yielded a 1024-demnsional representation for
starting with an uppercase character in the ar- the title, and another one for the article body.
ticle body; 4.6 Universal Sentence Encoder
• sentence count text: number of sentences in We also extracted representation using Google’s
the article; Universal Sentence Encoder, or USE, (Cer et al.,
2018). We used the pretrained model from TF
• avg sentence length char text: average Hub6 . As the model is only available for English,
length of the sentences in the article body, in we used the translations of the news articles. We
terms of characters; passed the model the first 300 tokens for each title
or body to generate 512-dimensional vectors.
• avg sentence length word text: average 5
http://www.alexa.com/
length of the sentences in the article body, in 6
http://tfhub.dev/google/
terms of words; universal-sentence-encoder/2
251
4.7 ElMo 5 Experiments and Evaluation
Next, we use deep contextualized word represen-
tations from ElMo, which uses generative bidirec- 5.1 Experimental Setup
tional language model pre-training (Peters et al., We used logistic regression as our main classifica-
2018). The model yields 1024-dimensional rep- tion method. As we have a small dataset, we per-
resentation, which we generate separately for the formed 5-fold cross-validation. For evaluation, we
article title and for its body. used accuracy and macro-average F1 score. The
4.8 NELA Features results are presented in Table 3.
Our baseline approach is based on selecting the
Finally, we use features from the NELA toolkit
most frequent class, i.e., non-toxic, which covers
(Horne et al., 2018a), which were previously
30.30% of the dataset (see Table 1).
shown useful for detecting fake news, political
bias, etc. The toolkit implements 129 features,
which we extract separately for the article title and 5.2 Individual Models
for its body:
We evaluated a total of 14 setups for feature com-
• Structure: POS tags, linguistic features bination. Four of them represent features gener-
(function words, pronouns, etc.), and fea- ated from the original article’s title and body as
tures for clickbait title classification from well as a combination thereof. The next four se-
(Chakraborty et al., 2016); tups present feature sets generated from the En-
glish translation as well as a combination thereof.
• Sentiment: sentiment scores using lexicons The final section of Table 3 shows three setups that
(Recasens et al., 2013; Mitchell et al., 2013) are somewhat language-independent: meta media,
and full systems (Hutto and Gilbert, 2014); all features combined together as well as a meta
classifier. We tuned the logistic regression for
• Topic: lexicon features to differentiate be-
each individual experimental setup, using an addi-
tween science topics and personal concerns;
tional internal cross-validation for the training part
• Complexity: type-token ratio, readability, of each experiment in the 5-fold cross-validation.
number of cognitive process words (identify- In total, 15,000 additional experiments have been
ing discrepancy, insight, certainty, etc.); conducted to complete the fine-tuning.
We can see in Table 3 that the BERT features
• Bias: features modeling bias (Recasens et al.,
(setups 2, 9) perform well both for English and for
2013; Mukherjee and Weikum, 2015) and
Bulgarian. The feature combinations (setups 6, 11,
subjectivity (Horne et al., 2017);
13) do not yield good results as this increases the
• Morality: features based on the Moral Foun- number of features, while the number of training
dation Theory (Graham et al., 2009) and lex- examples remains limited. Using only the right 6
icons (Lin et al., 2018) meta features about the target medium yields 12%
improvement over the baseline. Interestingly, LSA
A summary of all features is shown in Table 2. turns out to be the best text representation model.
252
Language N Feature Set Dimension Accuracy F1-macro
- 1 Baseline - 30.30 05.17
BG 2 BERT(title), BERT(text) 1,536 47.69 32.58
3 XLM(title), XLM(text) 2,048 38.50 24.58
4 Styl(title), Styl(text) 15 31.89 08.51
5 LSA(title), LSA(text) 215 55.59 42.11
6 Bulgarian combined 3,824 39.43 24.38
EN 7 USE(title), USE(text) 1,024 53.70 40.68
8 NELA(title), NELA(text) 258 36.36 23.04
9 BERT(title), BERT(text) 1,536 52.05 39.78
10 ElMO(title), ElMO(text) 2,048 54.60 40.95
11 English combined 4,878 45.45 31.42
- 12 Media meta 6 42.04 15.64
13 All combined 8,694 38.16 26.04
14 Meta classifier 153 59.06 39.70
anti-democratic 0 0 3 0 2 1 2 0 7
delusion 0 6 1 3 4 0 0 0 6
s
defamation 0 0 20 0 6 0 2 0 3 o
conspiracy 0 1 2 34 3 0 0 0 9 f
True labels
t
non-toxic 0 0 1 2 87 0 1 0 5 .
..........
m
..........
#classes
. 64
32 a
pro-authoritarian 0 0 3 4 2 0 0 0 4
..........
#features . 128
64 x
hate speech 0 0 1 2 4 0 16 0 3 .
. Output layer
sensation 0 0 1 0 9 0 2 1 4 .
.
fake news 0 2 1 14 7 0 3 0 23
hidden layers
de tic
fam n
ns n
pro no iracy
tho oxic
te an
sen ech
fak ion
s
ew
de sio
co atio
ra
ha ritari
sat
Input layer
e
en
-au n-t
u
oc
sp
p
l
em
ti-d
an
253
6 Discussion 7 Conclusion and Future Work
Below we compare the performance of the mod- We have presented experiments in detecting the
els when working with Bulgarian vs. English re- toxicity of news articles. While previous research
sources, and we further discuss issues related to was mostly limited to English, here we focused on
the small size of our dataset. Bulgarian. We created a new dataset by crawling
a website that has been collecting Bulgarian news
articles and manually categorized them in eight
6.1 Language-Related Issues
toxicity groups. We then trained a multi-class clas-
The first part of the feature comparison in Ta- sifier with nine categories: eight toxic and one
ble 3 is between English and Bulgarian, and it is non-toxic. We experimented with a variety of rep-
interesting to compare them. The first compari- resentations based on ElMo, BERT, xand XLM.
son is between the BERT features. Even though We further used a variety of domain-specific fea-
we used a pre-trained BERT model, with same tures, which we eventually combined in a meta
pooling techniques there is close to 4.5% absolute classifier. The evaluation results show an accuracy
improvement when using the English translation. of 59.0% and a macro-F1 score of 39.7%, which
This is probably due to the English BERT being represent sizable improvements over the majority-
trained on more data as the English Wikipedia is class baseline (Acc=30.3%, macro-F1=5.2%).
much bigger for English: 5.7M English articles vs. In future work, we plan to extend and also to
just 250K Bulgarian articles. balance the dataset. This can be achieved by ei-
Another notable comparison is between the ther exploring another source for articles using the
types of models. Two of the Bulgarian feature methodology of Media Scan, or by processing the
sets are created via local models (experiments 4 unstructured PDF article, which we ignored in the
and 5), while all of the English experiments are present study. We also plan to explore new infor-
from transfer-learning. We can see that LSA (ex- mation sources. From a technical point of view,
periment 5) is the best feature set, and one can ar- we would like to improve the neural network ar-
gue that this is to be expected. On such a small chitecture as well as the oversampling techniques
dataset in a non-English language, it is hard to rep- (with possible combination with undersampling).
resent the text with pre-trained models. Neverthe-
Data and Code We are releasing all of the code
less, we can see that a combination between only
for our experiments in a public repository that can
pre-trained models (experiment 11) performs bet-
be found in GitHub7 with explanations about how
ter compared to fusion between local and transfer-
to reproduce our environment. In that repository,
learning models (experiment 6).
we further release the full dataset together with
the generated features, all the textual data, all the
6.2 Data Size Issues translations and all the meta data about the articles.
There are several aspects of the above experiments
Acknowledgements
where we can observe the negative effect of hav-
ing insufficient data. First, in experiments 6, 11, This research is part of the Tanbih project,8 which
12 in Table 3, we can see that the combination of aims to limit the effect of “fake news”, propa-
features performs worse in each language group ganda and media bias by making users aware of
compared to single feature types. what they are reading. The project is developed
Another place where we felt we had insufficient in collaboration between the Qatar Computing
data was in the neural network experiments, where Research Institute (QCRI), HBKU and the MIT
we had many more parameters to train than in the Computer Science and Artificial Intelligence Lab-
simple logistic regression. oratory (CSAIL).
A related problem is that of class imbalance: we This research is also partially supported
have seen in Section 5.4 above that the smallest by Project UNITe BG05M2OP001-1.001-0004
classes were effectively ignored even by our best funded by the OP “Science and Education for
classifier. We can see in Figure 1 that those three Smart Growth” and the EU via the ESI Funds.
classes have less than 60 articles combined, while 7
github.com/yoandinkov/ranlp-2019
the “non-toxic” only had 96 articles. 8
http://tanbih.qcri.org/
254
References Ann M Brill. 2001. Online journalists embrace new
marketing function. Newspaper Research Journal
Pepa Atanasova, Lluı́s Màrquez, Alberto Barrón- 22(2):28–40.
Cedeño, Tamer Elsayed, Reem Suwaileh, Wajdi Za-
ghouani, Spas Kyuchukov, Giovanni Da San Mar- Kevin R. Canini, Bongwon Suh, and Peter L. Pirolli.
tino, and Preslav Nakov. 2018. Overview of the 2011. Finding credible information sources in so-
CLEF-2018 CheckThat! lab on automatic identifi- cial networks based on content and social structure.
cation and verification of political claims, Task 1: In Proceedings of the IEEE International Confer-
Check-worthiness. In CLEF 2018 Working Notes. ence on Privacy, Security, Risk, and Trust, and the
CEUR-WS.org, Avignon, France. IEEE International Conference on Social Comput-
ing. Boston, MA, USA, SocialCom/PASSAT ’11,
Pepa Atanasova, Preslav Nakov, Georgi Karadzhov, pages 1–8.
Mitra Mohtarami, and Giovanni Da San Martino.
2019. Overview of the CLEF-2019 CheckThat! Carlos Castillo, Marcelo Mendoza, and Barbara
Lab on Automatic Identification and Verification of Poblete. 2011. Information credibility on Twitter.
Claims. Task 1: Check-Worthiness. In CLEF 2019 In Proceedings of the 20th International Conference
Working Notes. CEUR-WS.org, Lugano, Switzer- on World Wide Web. Hyderabad, India, WWW ’11,
land. pages 675–684.
Mouhamadou Lamine Ba, Laure Berti-Equille, Kushal Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua,
Shah, and Hossam M. Hammady. 2016. VERA: Nicole Limtiaco, Rhomni St John, Noah Constant,
A platform for veracity estimation over web data. Mario Guajardo-Cespedes, Steve Yuan, Chris Tar,
In Proceedings of the 25th International Conference et al. 2018. Universal sentence encoder. arXiv
Companion on World Wide Web. Montréal, Québec, preprint arXiv:1803.11175 .
Canada, WWW ’16, pages 159–162.
Abhijnan Chakraborty, Bhargavi Paranjape, Kakarla
Ramy Baly, Georgi Karadzhov, Dimitar Alexandrov, Kakarla, and Niloy Ganguly. 2016. Stop clickbait:
James Glass, and Preslav Nakov. 2018a. Predict- Detecting and preventing clickbaits in online news
ing factuality of reporting and bias of news media media. In Proceedings of the 2016 IEEE/ACM In-
sources. In Proceedings of the Conference on Em- ternational Conference on Advances in Social Net-
pirical Methods in Natural Language Processing. works Analysis and Mining. San Francisco, CA,
Brussels, Belgium, EMNLP ’18, pages 3528–3539. USA, ASONAM ’16, pages 9–16.
Ramy Baly, Georgi Karadzhov, Abdelrhman Saleh, Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall,
James Glass, and Preslav Nakov. 2019. Multi-task and W Philip Kegelmeyer. 2002. SMOTE: synthetic
ordinal regression for jointly predicting the trustwor- minority over-sampling technique. Journal of artifi-
thiness and the leading political ideology of news cial intelligence research 16:321–357.
media. In Proceedings of the 17th Annual Confer- Cheng Chen, Kui Wu, Venkatesh Srinivasan, and
ence of the North American Chapter of the Asso- Xudong Zhang. 2013. Battling the Internet Water
ciation for Computational Linguistics. Minneapolis, Army: detection of hidden paid posters. In Pro-
MN, USA, NAACL-HLT ’19, pages 2109–2116. ceedings of the 2013 IEEE/ACM International Con-
ference on Advances in Social Networks Analysis
Ramy Baly, Mitra Mohtarami, James Glass, Lluı́s and Mining. Niagara, Canada, ASONAM ’13, pages
Màrquez, Alessandro Moschitti, and Preslav Nakov. 116–120.
2018b. Integrating stance detection and fact check-
ing in a unified corpus. In Proceedings of the Con- Giovanni Da San Martino, Seunghak Yu, Alberto
ference of the North American Chapter of the Asso- Barron-Cedeno, Rostislav Petrov, and Preslav
ciation for Computational Linguistics. New Orleans, Nakov. 2019. Fine-grained analysis of propaganda
LA, USA, NAACL-HLT ’18, pages 21–27. in news articles. In Proceedings of the 2019 Con-
ference on Empirical Methods in Natural Language
Alberto Barrón-Cedeño, Giovanni Da San Martino, Processing. Hong Kong, China, EMNLP ’19.
Israa Jaradat, and Preslav Nakov. 2019. Proppy:
Organizing the news based on their propagandis- Kareem Darwish, Dimitar Alexandrov, Preslav Nakov,
tic content. Information Processing & Management and Yelena Mejova. 2017a. Seminar users in the
56(5):1849 – 1864. Arabic Twitter sphere. In Proceedings of the 9th
International Conference on Social Informatics. Ox-
Alberto Barrón-Cedeño, Tamer Elsayed, Reem ford, UK, SocInfo ’17, pages 91–108.
Suwaileh, Lluı́s Màrquez, Pepa Atanasova, Wajdi
Zaghouani, Spas Kyuchukov, Giovanni Da San Mar- Kareem Darwish, Walid Magdy, and Tahar Zanouda.
tino, and Preslav Nakov. 2018. Overview of the 2017b. Improved stance prediction in a user similar-
CLEF-2018 CheckThat! lab on automatic identi- ity feature space. In Proceedings of the Conference
fication and verification of political claims, Task on Advances in Social Networks Analysis and Min-
2: Factuality. In CLEF 2018 Working Notes. ing. Sydney, Australia, ASONAM ’17, pages 145–
CEUR-WS.org, Avignon, France. 148.
255
Sohan De Sarkar, Fan Yang, and Arjun Mukherjee. Jesse Graham, Jonathan Haidt, and Brian A Nosek.
2018. Attending sentences to detect satirical fake 2009. Liberals and conservatives rely on different
news. In Proceedings of the 27th International sets of moral foundations. Journal of personality
Conference on Computational Linguistics. Santa Fe, and social psychology 96(5):1029.
NM, USA, COLING ’18, pages 3371–3380.
Andreas Hanselowski, Avinesh PVS, Benjamin
Leon Derczynski, Kalina Bontcheva, Maria Liakata, Schiller, Felix Caspelherr, Debanjan Chaudhuri,
Rob Procter, Geraldine Wong Sak Hoi, and Arkaitz Christian M. Meyer, and Iryna Gurevych. 2018. A
Zubiaga. 2017. SemEval-2017 Task 8: Ru- retrospective analysis of the fake news challenge
mourEval: Determining rumour veracity and sup- stance-detection task. In Proceedings of the 27th
port for rumours. In Proceedings of the 11th In- International Conference on Computational Lin-
ternational Workshop on Semantic Evaluation. Van- guistics. Santa Fe, NM, USA, COLING ’18, pages
couver, Canada, SemEval ’17, pages 60–67. 1859–1874.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Momchil Hardalov, Ivan Koychev, and Preslav Nakov.
Kristina Toutanova. 2019. BERT: Pre-training of 2016. In search of credible news. In Proceedings
deep bidirectional transformers for language under- of the 17th International Conference on Artificial In-
standing. In Proceedings of the 2019 Conference of telligence: Methodology, Systems, and Applications.
the North American Chapter of the Association for Varna, Bulgaria, AIMSA ’16, pages 172–180.
Computational Linguistics. Minneapolis, MN, USA,
NAACL-HLT ’2019, pages 4171–4186. Maram Hasanain, Reem Suwaileh, Tamer Elsayed,
Alberto Barrón-Cedeño, and Preslav Nakov. 2019.
Yoan Dinkov, Ahmed Ali, Ivan Koychev, and Preslav Overview of the CLEF-2019 CheckThat! Lab on
Nakov. 2019. Predicting the leading political ide- Automatic Identification and Verification of Claims.
ology of Youtube channels using acoustic, textual Task 2: Evidence and Factuality. In CLEF 2019
and metadata information. In Proceedings of the Working Notes. CEUR-WS.org, Lugano, Switzer-
20th Annual Conference of the International Speech land.
Communication Association. Graz, Austria, INTER-
SPEECH ’19. Benjamin Horne and Sibel Adali. 2017. This just in:
Xin Luna Dong, Evgeniy Gabrilovich, Kevin Murphy, Fake news packs a lot in title, uses simpler, repetitive
Van Dang, Wilko Horn, Camillo Lugaresi, Shao- content in text body, more similar to satire than real
hua Sun, and Wei Zhang. 2015. Knowledge-based news. CoRR abs/1703.09398.
trust: Estimating the trustworthiness of web sources. Benjamin Horne, Sibel Adali, and Sujoy Sikdar. 2017.
Proc. VLDB Endow. 8(9):938–949. Identifying the social signals that drive online dis-
Sebastian Dungs, Ahmet Aker, Norbert Fuhr, and cussions: A case study of Reddit communities. In
Kalina Bontcheva. 2018. Can rumour stance alone Proceedings of the 26th IEEE International Confer-
predict veracity? In Proceedings of the 27th Inter- ence on Computer Communication and Networks.
national Conference on Computational Linguistics. Vancouver, Canada, ICCCN ’17, pages 1–9.
Santa Fe, NM, USA, COLING ’18, pages 3360–
3370. Benjamin D. Horne, William Dron, Sara Khedr, and
Sibel Adali. 2018a. Assessing the news landscape:
Tamer Elsayed, Preslav Nakov, Alberto Barrón- A multi-module toolkit for evaluating the credibility
Cedeño, Maram Hasanain, Reem Suwaileh, Pepa of news. In Proceedings of the The Web Conference.
Atanasova, and Giovanni Da San Martino. 2019a. Lyon, France, WWW ’18, pages 235–238.
CheckThat! at CLEF 2019: Automatic identifica-
tion and verification of claims. In Proceedings of the Benjamin D. Horne, Sara Khedr, and Sibel Adali.
41st European Conference on Information Retrieval. 2018b. Sampling the news producers: A large news
Cologne, Germany, ECIR ’19, pages 309–315. and feature data set for the study of the complex
media landscape. In Proceedings of the Twelfth In-
Tamer Elsayed, Preslav Nakov, Alberto Barrón- ternational Conference on Web and Social Media.
Cedeño, Maram Hasanain, Reem Suwaileh, Gio- Stanford, CA, USA, ICWSM ’18, pages 518–527.
vanni Da San Martino, and Pepa Atanasova. 2019b.
Overview of the CLEF-2019 CheckThat!: Auto- Clayton Hutto and Eric Gilbert. 2014. VADER: A par-
matic identification and verification of claims. In simonious rule-based model for sentiment analysis
Experimental IR Meets Multilinguality, Multimodal- of social media text. In Proceedings of the 8th Inter-
ity, and Interaction. Springer, Lugano, Switzerland. national Conference on Weblogs and Social Media.
Ann Arbor, MI, USA, ICWSM ’14.
Genevieve Gorrell, Elena Kochkina, Maria Liakata,
Ahmet Aker, Arkaitz Zubiaga, Kalina Bontcheva, Georgi Karadzhov, Pepa Gencheva, Preslav Nakov, and
and Leon Derczynski. 2019. SemEval-2019 task 7: Ivan Koychev. 2017a. We built a fake news & click-
RumourEval, determining rumour veracity and sup- bait filter: What happened next will blow your mind!
port for rumours. In Proceedings of the 13th In- In Proceedings of the International Conference on
ternational Workshop on Semantic Evaluation. Min- Recent Advances in Natural Language Processing.
neapolis, MN, USA, SemEval ’19, pages 845–854. Varna, Bulgaria, RANLP ’17, pages 334–343.
256
Georgi Karadzhov, Preslav Nakov, Lluı́s Màrquez, Al- the ACM Conference on Computer Supported Coop-
berto Barrón-Cedeño, and Ivan Koychev. 2017b. erative Work and Social Computing. Portland, OR,
Fully automated fact checking using external USA, CSCW ’17, pages 243–246.
sources. In Proceedings of the International Confer-
ence on Recent Advances in Natural Language Pro- Todor Mihaylov, Georgi Georgiev, and Preslav Nakov.
cessing. Varna, Bulgaria, RANLP ’17, pages 344– 2015a. Finding opinion manipulation trolls in news
353. community forums. In Proceedings of the Nine-
teenth Conference on Computational Natural Lan-
Elena Kochkina, Maria Liakata, and Arkaitz Zubiaga. guage Learning. Beijing, China, CoNLL ’15, pages
2018. All-in-one: Multi-task learning for rumour 310–314.
verification. In Proceedings of the International
Conference on Computational Linguistics. Santa Fe, Todor Mihaylov, Ivan Koychev, Georgi Georgiev, and
NM, USA, COLING ’18, pages 3402–3413. Preslav Nakov. 2015b. Exposing paid opinion ma-
nipulation trolls. In Proceedings of the International
Vivek Kulkarni, Junting Ye, Steve Skiena, and Conference Recent Advances in Natural Language
William Yang Wang. 2018. Multi-view models for Processing. Hissar, Bulgaria, RANLP ’15, pages
political ideology detection of news articles. In Pro- 443–450.
ceedings of the 2018 Conference on Empirical Meth-
ods in Natural Language Processing. Brussels, Bel- Todor Mihaylov, Tsvetomila Mihaylova, Preslav
gium, EMNLP ’18, pages 3518–3527. Nakov, Lluı́s Màrquez, Georgi Georgiev, and Ivan
Koychev. 2018. The dark side of news community
Guillaume Lample and Alexis Conneau. 2019. Cross- forums: Opinion manipulation trolls. Internet Re-
lingual language model pretraining. CoRR search 28(5):1292–1312.
abs/1901.07291.
Todor Mihaylov and Preslav Nakov. 2016. Hunting for
David M.J. Lazer, Matthew A. Baum, Yochai Ben- troll comments in news community forums. In Pro-
kler, Adam J. Berinsky, Kelly M. Greenhill, Filippo ceedings of the 54th Annual Meeting of the Asso-
Menczer, Miriam J. Metzger, Brendan Nyhan, Gor- ciation for Computational Linguistics. Berlin, Ger-
don Pennycook, David Rothschild, Michael Schud- many, ACL ’16, pages 399–405.
son, Steven A. Sloman, Cass R. Sunstein, Emily A.
Thorson, Duncan J. Watts, and Jonathan L. Zit- Tsvetomila Mihaylova, Georgi Karadzhov, Pepa
train. 2018. The science of fake news. Science Atanasova, Ramy Baly, Mitra Mohtarami, and
359(6380):1094–1096. Preslav Nakov. 2019. SemEval-2019 task 8: Fact
checking in community question answering forums.
Yaliang Li, Jing Gao, Chuishi Meng, Qi Li, Lu Su, In Proceedings of the 13th International Workshop
Bo Zhao, Wei Fan, and Jiawei Han. 2016. A sur- on Semantic Evaluation. Minneapolis, MN, USA,
vey on truth discovery. SIGKDD Explor. Newsl. SemEval ’19, pages 860–869.
17(2):1–16.
Tsvetomila Mihaylova, Preslav Nakov, Lluı́s Màrquez,
Y. Lin, J. Hoover, G. Portillo-Wightman, C. Park, Alberto Barrón-Cedeño, Mitra Mohtarami, Georgi
M. Dehghani, and H. Ji. 2018. Acquiring back- Karadjov, and James Glass. 2018. Fact checking in
ground knowledge to improve moral value predic- community forums. In Proceedings of the Thirty-
tion. In Proceedings of the 2018 IEEE/ACM In- Second AAAI Conference on Artificial Intelligence.
ternational Conference on Advances in Social Net- New Orleans, LA, USA, AAAI ’18, pages 879–886.
works Analysis and Mining. Los Alamitos, CA,
USA, ASONAM ’18, pages 552–559. Lewis Mitchell, Morgan R Frank, Kameron Decker
Harris, Peter Sheridan Dodds, and Christopher M
Jing Ma, Wei Gao, Prasenjit Mitra, Sejeong Kwon, Danforth. 2013. The geography of happiness:
Bernard J. Jansen, Kam-Fai Wong, and Meeyoung Connecting Twitter sentiment and expression, de-
Cha. 2016. Detecting rumors from microblogs with mographics, and objective characteristics of place.
recurrent neural networks. In Proceedings of the PloS one 8(5):e64417.
25th International Joint Conference on Artificial In-
telligence. New York, NY, USA, IJCAI ’16, pages Mitra Mohtarami, Ramy Baly, James Glass, Preslav
3818–3824. Nakov, Lluı́s Màrquez, and Alessandro Moschitti.
2018. Automatic stance detection using end-to-end
Jing Ma, Wei Gao, and Kam-Fai Wong. 2017. De- memory networks. In Proceedings of the Annual
tect rumors in microblog posts using propagation Conference of the North American Chapter of the
structure via kernel learning. In Proceedings of the Association for Computational Linguistics. New Or-
55th Annual Meeting of the Association for Compu- leans, LA, USA, NAACL-HLT ’18, pages 767–776.
tational Linguistics. Vancouver, Canada, ACL ’17,
pages 708–717. Mitra Mohtarami, James Glass, and Preslav Nakov.
2019. Contrastive language adaptation for cross-
Suman Kalyan Maity, Aishik Chakraborty, Pawan lingual stance detection. In Proceedings of the 2019
Goyal, and Animesh Mukherjee. 2017. Detection Conference on Empirical Methods in Natural Lan-
of sockpuppets in social media. In Proceedings of guage Processing. Hong Kong, China, EMNLP ’19.
257
Subhabrata Mukherjee and Gerhard Weikum. 2015. Hannah Rashkin, Eunsol Choi, Jin Yea Jang, Svitlana
Leveraging joint interactions for credibility analysis Volkova, and Yejin Choi. 2017. Truth of varying
in news communities. In Proceedings of the 24th shades: Analyzing language in fake news and polit-
ACM International on Conference on Information ical fact-checking. In Proceedings of the 2017 Con-
and Knowledge Management. Melbourne, Australia, ference on Empirical Methods in Natural Language
CIKM ’15, pages 353–362. Processing. Copenhagen, Denmark, EMNLP ’17,
pages 2931–2937.
Preslav Nakov, Alberto Barrón-Cedeño, Tamer El-
sayed, Reem Suwaileh, Lluı́s Màrquez, Wajdi Za- Marta Recasens, Cristian Danescu-Niculescu-Mizil,
ghouani, Pepa Atanasova, Spas Kyuchukov, and and Dan Jurafsky. 2013. Linguistic models for ana-
Giovanni Da San Martino. 2018. Overview of the lyzing and detecting biased language. In Proceed-
CLEF-2018 CheckThat! lab on automatic identifica- ings of the 51st Annual Meeting of the Associa-
tion and verification of political claims. In Proceed- tion for Computational Linguistics. Sofia, Bulgaria,
ings of CLEF. Avignon, France, pages 372–387. ACL ’13, pages 1650–1659.
An T. Nguyen, Aditya Kharosekar, Matthew Lease, Benjamin Riedel, Isabelle Augenstein, Georgios P Sp-
and Byron C. Wallace. 2018. An interpretable joint ithourakis, and Sebastian Riedel. 2017. A simple but
graphical model for fact-checking from crowds. In tough-to-beat baseline for the Fake News Challenge
Proceedings of the Thirty-Second AAAI Conference stance detection task. ArXiv:1707.03264 .
on Artificial Intelligence. New Orleans, LA, USA,
AAAI ’18, pages 1511–1518. Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, and
Huan Liu. 2017. Fake news detection on social me-
Verónica Pérez-Rosas, Bennett Kleinberg, Alexandra dia: A data mining perspective. SIGKDD Explor.
Lefevre, and Rada Mihalcea. 2018. Automatic de- Newsl. 19(1):22–36.
tection of fake news. In Proceedings of the 27th In-
James Thorne, Mingjie Chen, Giorgos Myrianthous,
ternational Conference on Computational Linguis-
Jiashu Pu, Xiaoxuan Wang, and Andreas Vlachos.
tics. Santa Fe, NM, USA, COLING ’18, pages
2017. Fake news stance detection using stacked en-
3391–3401.
semble of classifiers. In Proceedings of the EMNLP
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Workshop on Natural Language Processing meets
Gardner, Christopher Clark, Kenton Lee, and Luke Journalism. Copenhagen, Denmark, pages 80–83.
Zettlemoyer. 2018. Deep contextualized word rep- James Thorne and Andreas Vlachos. 2018. Automated
resentations. In Proceedings of the Conference of fact checking: Task formulations, methods and fu-
the North American Chapter of the Association for ture directions. In Proceedings of the International
Computational Linguistics. New Orleans, LA, USA, Conference on Computational Linguistics. Santa Fe,
NAACL-HLT ’18, pages 2227–2237. NM, USA, COLING ’18, pages 3346–3359.
Kashyap Popat, Subhabrata Mukherjee, Jannik James Thorne, Andreas Vlachos, Christos
Strötgen, and Gerhard Weikum. 2016. Credibil- Christodoulopoulos, and Arpit Mittal. 2018.
ity assessment of textual claims on the web. In FEVER: a large-scale dataset for fact extraction and
Proceedings of the 25th ACM International on VERification. In Proceedings of the Conference
Conference on Information and Knowledge Man- of the North American Chapter of the Association
agement. Indianapolis, IN, USA, CIKM ’16, pages for Computational Linguistics. New Orleans, LA,
2173–2178. USA, NAACL-HLT ’18, pages 809–819.
Kashyap Popat, Subhabrata Mukherjee, Jannik Soroush Vosoughi, Deb Roy, and Sinan Aral. 2018.
Strötgen, and Gerhard Weikum. 2017. Where the The spread of true and false news online. Science
truth lies: Explaining the credibility of emerging 359(6380):1146–1151.
claims on the Web and social media. In Proceedings
of the 26th International Conference on World Wide Yifan Zhang, Giovanni Da San Martino, Alberto
Web Companion. Perth, Australia, WWW ’17, Barrón-Cedeño, Salvatore Romeo, Jisun An, Hae-
pages 1003–1012. woon Kwak, Todor Staykovski, Israa Jaradat,
Georgi Karadzhov, Ramy Baly, Kareem Darwish,
Kashyap Popat, Subhabrata Mukherjee, Jannik and Preslav Nakov James Glass. 2019. Tanbih:
Strötgen, and Gerhard Weikum. 2018. CredEye: A Get to know what you are reading. In Proceedings
credibility lens for analyzing and explaining misin- of the 2019 Conference on Empirical Methods in
formation. In Proceedings of The Web Conference Natural Language Processing. Hong Kong, China,
2018. Lyon, France, WWW ’18, pages 155–158. EMNLP ’19.
Martin Potthast, Johannes Kiesel, Kevin Reinartz, Dimitrina Zlatkova, Preslav Nakov, and Ivan Koychev.
Janek Bevendorff, and Benno Stein. 2018. A stylo- 2019. Fact-checking meets fauxtography: Verifying
metric inquiry into hyperpartisan and fake news. In claims about images. In Proceedings of the 2019
Proceedings of the 56th Annual Meeting of the Asso- Conference on Empirical Methods in Natural Lan-
ciation for Computational Linguistics. Melbourne, guage Processing. Hong Kong, China, EMNLP ’19.
Australia, ACL ’18, pages 231–240.
258
De-Identification of Emails:
Pseudonymizing Privacy-Sensitive Data in a German Email Corpus
Elisabeth Eder Ulrike Krieg-Holz
Institut für Germanistik
Alpen-Adria-Universität Klagenfurt, Klagenfurt, Austria
{elisabeth.eder | ulrike.krieg-holz}@aau.at
Udo Hahn
Jena University Language & Information Engineering (JULIE) Lab
Friedrich-Schiller-Universität Jena, Jena, Germany
udo.hahn@uni-jena.de
259
Proceedings of Recent Advances in Natural Language Processing, pages 259–269,
Varna, Bulgaria, Sep 2–4 2019.
Surprisingly, despite its high relevance for NLP on German language data. Since surrogate gener-
operating on UGC, the topic of data privacy has ation constitutes a highly constrained case of lan-
long been neglected by the mainstream of NLP re- guage generation, in Section 6 we describe the re-
search. While it has always been of utmost im- sults of an evaluation study to assess the natural-
portance for medical, i.e., clinical, NLP (Meystre, ness of these replacements with native speakers of
2015), it has received almost no attention in NLP’s German, as well as the performance of a recog-
non-medical camp for a long time (for two early nizer for privacy-relevant text stretches on original
exceptions, cf. Rock (2001); Medlock (2006)). and already pseudonymized data.
This naı̈ve perspective is beginning to change
these days with the ever-growing importance of
2 Related Work
social media documents for text analytics. How- The main thrust of work on de-identification has
ever, there are currently no systematic actions been performed for clinical NLP.3 Most influential
taken to hide personally sensitive information for progress in this field have been two challenge
from down-stream applications when dealing with competitions within the context of the I 2 B 2 (In-
chat, blog, SMS or email raw data. Since this formatics for Integrating Biology & the Bedside)
attitude also faces legal implications, a quest for initiative4 which focused on 18 different types of
the protection of individual data privacy has been Protected Health Information (PHI) categories as
raised and, in the meantime, finds active response required by US legislation (HIPAA).5 The first of
in the most recent work of the NLP community (Li these challenge tasks was launched in 2006 for
et al., 2018; Coavoux et al., 2018). 889 hospital discharge summaries, with a total of
We distinguish two basic approaches to elimi- 19,498 PHI instances of person-identifying verbal
nate privacy-bearing data from raw text data. The expressions (Uzuner et al., 2007). The second was
first one, anonymization, identifies instances of run in 2014 and addressed an even broader set of
relevant privacy categories (e.g., person names or PHI categories (Stubbs et al., 2015a). In sum-
dates) and replaces sensitive strings by some ar- mary, the best system performances peaked in the
tificial code (e.g., ‘xxx’). This blinding approach high 90s (F1 score) using classical machine learn-
might be appropriate to eliminate privacy-bearing ing methods, Conditional Random Fields (CRFs)
data in the medical world, but it is inappropriate in particular, and hand-written rules, or a mixture
for most NLP applications since crucial discrim- of both. As a successor to I 2 B 2, the C EGS -NG RID
inative information and contextual clues will be Shared Tasks and Workshop on Challenges in NLP
erased by such a scrubbing procedure. for Clinical Data created a corpus of 1,000 man-
The second approach, pseudonymization, pre- ually de-identified psychiatric evaluation records
serves such valuable information by replacing (Stubbs et al., 2017). Interestingly, for the au-
privacy-bearing text strings with randomly chosen tomatic de-identification task performance values
alternative synthetic instances from the same pri- dropped significantly down to 79.85 F1 for the
vacy type (e.g., the person name ‘Suzanne Walker’ best-performing system indicating an only mod-
is mapped to ‘Caroline Snyder’). As a common est potential for domain and text genre portability
denominator, the term de-identification subsumes (moving from discharge summaries to psychiatric
both, anonymization and pseudonymization. evaluation records).
Recently, the deep learning wave has also hit
The focus of this paper will be on pseudo-
the (clinical) de-identification community. For this
nymization and more precisely on the methods
task, bidirectional Long-Short Term Memory Net-
needed to produce realistic synthetic replace-
works (Bi-LSTMs) became quite popular as ev-
ments, a process often also referred to as surrogate
idenced by the work of Dernoncourt et al. (2017)
generation. We start with a discussion of related
work in Section 2 and then introduce the seman- 3
Note that we have to distinguish between data protec-
tic types we consider as relevant carriers of per- tion in structured data contained in (clinical) information sys-
tems, (for which k-anonymity (Sweeney, 2002) is a well-
sonal information in emails in Section 3. Next, known model to minimize a person’s re-identification risk)
we provide an overview of the email corpus our and pseudonym-based textual variant generation for unstruc-
tured verbal data we here focus on.
experiments are based on in Section 4, including 4
https://www.i2b2.org/
manual annotation efforts. In Section 5, we turn 5
https://www.hhs.gov/hipaa/
to the process of surrogate generation, with focus for-professionals/privacy/index.html
260
who achieve an F1 score of 97.85 on the I 2 B 2 2014 codes of the encountered semantic type (e.g., each
dataset, or Liu et al. (2017) who report perfor- specific email address was replaced by the type
mance figures ranging from 95.11% over 96.98% symbol EMAIL), not by an alternative semantic
up to 98.28% micro F1 score under increasingly token, i.e., a pseudonym. The same strategy was
sloppier matching criteria on the same dataset. also chosen by Chen and Kan (2013) for their SMS
Note that these challenges were focusing on the corpus that contains more than 71,000 messages,
recognition of privacy-relevant text stretches tex- focusing on English and Mandarin. However, nei-
tual but did not incorporate pseudonymization, a ther are the methods of automatic anonymization
more complex task (Stubbs et al., 2015b). Carrell described in detail, nor are performance figures
et al. (2013) deal with the latter challenge within of this process reported in both papers (Chen and
the context of the ‘Hiding In Plain Sight’ approach Kan (2013) only mention the use of regular ex-
where the detected privacy-bearing identifiers are pressions for the anonymization process).
replaced with realistic synthetic surrogates in or- In conclusion, pseudonymization has to the best
der to collectively render the few ’leaked’ iden- of our knowledge only been seriously applied
tifiers difficult to distinguish from the synthetic to medical documents, up until now. Hence,
surrogates—a major advantage for pseudonymiza- our investigation opens this study field for the
tion over anonymization. Targeting English medi- first time ever to non-medical applications of
cal texts S CRUB (Sweeney, 1996) is one of the first pseudonymization. Such de-identified corpora can
surrogate generation systems followed by work then easily be distributed via public sites and so
from Uzuner et al. (2007), Yeniterzi et al. (2010), might stimulate further NLP research.
Deléger et al. (2014), Stubbs et al. (2015b) and
Stubbs and Uzuner (2015). Similar procedures 3 Named Entities for De-Identification
have been proposed for Swedish (Alfalahi et al., Perhaps the most relevant source and starting point
2012) and Danish (Pantazos et al., 2011) clinical for determining types of personally identifying in-
corpora, yet not for German ones. formation pieces in written documents is a cata-
Work outside the clinical domain is rare. While logue of Personal Health Information (PHI) items
we found no work dealing with the anonymization that has been derived from the Health Informa-
or even pseudonymization of emails and Twitter- tion Privacy Act (HIPAA) which is binding law
style social media data,6 anonymizing SMSes is in the US. PHI enumerates altogether 18 privacy-
a topic of active research. Patel et al. (2013) in- sensitive items organized into eight main cate-
troduce a system capable of anonymizing SMS gories (Stubbs and Uzuner, 2015):
(Short Message Service) communication. Their
• Name includes the names of patients, doctors
study builds on 90,000 authentic French text mes-
and user names,
sages and uses dictionaries as well as decision
trees as machine learning technique. Their eval- • Profession of persons mentioned,
uation task is, however, very coarse-grained— • Location includes rooms, clinical depart-
select those SMSes from a test corpus of 23,055 ments, hospital names, names of organiza-
messages that either have or have not to be tions, street names, city names, state names,
anonymized. There is no breakdown to PHI-like names of countries, ZIPs, etc.,
categories known from the medical domain. • Age of persons,
Treurniet et al. (2012) were taking care of • Date expressions,
privacy-relevant data for a Dutch SMS corpus • Communication data, e.g., phone or fax num-
(52,913 messages, in total) in much more detail. bers, email addresses, URLs, IP addresses,
They automatically anonymized all occurrrences • all sorts of IDs such as Social Security num-
of dates, times, decimal amounts, and numbers ber, medical record number, health plan num-
with more than one digit (telephone numbers, ber, account number, license number, vehicle
bank accounts, etc.), e-mail addresses, URLs, and ID, device ID, biometric ID, etc.,
IP addresses. All sensitive information was re- • any Other form of personally sensitive data.
placed with corresponding semantic placeholder
While some types of categories from above are
6
Lüngen et al. (2017) report on manual anonymization ef- generally useful also for non-medical anonymiza-
forts for German chat data. tion procedures, others are quite domain-specific,
261
because they are intrinsically attached to the clin- ZipCode (ZIP), and CityName (CITY) which
ical domain (such as the names of patients, doc- stands for villages, towns, cities, larger metropoli-
tors or nurses, the names of hospitals and their de- tan areas (e.g., ‘Larger Manchester’) and regions
partments, or various forms of IDs, e.g., health in- smaller than a state (e.g., ‘Bay Area’); it also in-
surance numbers). Hence, we adapted this list for cludes derivations of these names (e g., ‘Roman’).
email documents while, at the same time, we tried Finally, Address (ADD) encompasses EmailAd-
to avoid over-fitting to this text genre. dress (EMAIL), PhoneNumber (PHONE), includ-
We, finally, came up with the category set de- ing fax numbers, and URL (URL), as well as other
picted in Figure 1 which we stipulate to univer- forms of domain names.
sally account for all types of emails, irrespective of Unlike some approaches from the field of clin-
any particular natural language and email (header) ical NLP (Stubbs et al., 2015b), we did not take
encoding. The categories are organized in a con- ages or professions into account, because our use
cise hierarchy whose top level categories are So- case is not that sensible and ages or professions
cialActor (ACTOR), Date (DATE), FormalIdenti- probably are mentioned far more often in clini-
fier (FID), Location (LOC), and Address (ADD). cal reports than in emails. Furthermore, unspe-
We anticipate that this hierarchy can be further re- cific dates like ‘Christmas’ or ‘next week’ and geo-
fined and accomodated to other text genres as well. graphical information such as landmarks, rivers or
lakes were not tagged for de-identification since
their contribution to possible re-identification is
fairly limited due to their generality.
262
Table 1 shows Cohen’s Kappa (Cohen, 1960) as cording to the findings of the error analysis. The
a measure for IAA for the pairs of annotators cal- outcome of this overhaul constituted the final gold
culated on the entities represented by the BIO an- standard annotations of C ODE A LLTAGS+d for the
notation scheme.9 Hence, not only the token label de-identification task.
itself but also matching starting and ending points
of an entity are taken into account, as well. The 5 Surrogate Generation
agreement is quite high, especially between anno-
Once pi-relevant named entities have been recog-
tator 1 and 3.
nized they undergo a replacement process where
A1 - A2 A2 - A3 A3 - A1 original identifiers are substituted by synthetic,
0.925 0.933 0.958 though natural, surrogates. For this step, we dis-
tinguish language-independent criteria from those
Table 1: Cohen’s Kappa for BIO tags on C ODE which are intrinsically language-specific. Only the
A LLTAGS+d latter have to be re-specified for languages other
than German.
Based on these 50 emails the annotators also ex-
amined and discussed differences of their annota- 5.1 Language-Independent Criteria
tions and decided on the gold standard by major- Personal information belonging to the categories
ity vote, which we applied for further evaluation in ADD (EMAIL, PHONE or fax number, URL),
order to measure precision (Prec), recall (Rec) and FID (PASSword, UFID), USER name and ZIP
F1 score (F1 ). Table 2 shows the outcomes regard- code are relatively simple to replace. Each digit
ing BIO tags per annotator and the overall result of the string is substituted with a randomly gen-
calculated over the joint annotations of the anno- erated alternative digit, each alphabetic character
tators. We took the averages from the outcomes of is replaced with a randomly generated alternative
the single categories weighted by the number of letter of the same case and alphabet. Other char-
true instances for each label. acters like ‘@’ or punctuation marks are left as is.
Prec Rec F1 For URLs, we also keep the subdomain ‘www’ and
A1 98.06 95.63 96.72 commonly used URL schemes like ‘http’, ‘https’,
A2 87.67 77.92 80.36 ‘ftp’, ‘file’ and ‘mailto’. In contrast to Stubbs et al.
A3 94.14 84.79 86.82 (2015b) we did not implement any other restric-
A1+A2+A3 93.37 86.11 88.66 tions for the selection of characters. As a conse-
quence, the resulting surrogates may have an un-
Table 2: Weighted average of precision, recall and realistic appearance.
F1 score with respect to the gold standard for BIO To maintain temporal ordering in the document
tags of C ODE A LLTAGS+d we generate a time shift separately for each text to
make re-identification difficult, if not impossible.
An error analysis revealed that, besides mostly We shifted the dates by a random interval between
accidental errors, a higher disagreement on tag- 365 days forward or backward. As we try to keep
ging ORGs (due to overlap or confusion with the original language-dependent formats, the user
city or product names and rather generic orga- has to determine the formats to be generated.
nizations)10 and an uncertainty regarding DATEs In order to maintain coreferences between pi
could be observed. The latter problem was solved entities, we replace multiple occurrences of an en-
by the decision to treat all dates as pi regardless tity by the same surrogate. To account for different
of their specificity. As a consequence, one anno- spellings of names regarding lower and upper case
tator worked through the entire corpus to re-tag we treat different possibilities of combinations,
each DATE and, if necessary, also re-tag ORGs ac- the original spelling, lower case, upper case and
9
a normalized format which is language-specific.
‘B’ preceding a token’s tag stands for the Beginning of
an entity, ‘I’ for its continuation (Inside), and ‘O’ for any We decided not to consider misspellings, because
stretch of text not belonging to an entity (Outside). checking for slightly different names, e.g., em-
10
Stubbs and Uzuner (2015) also report confusions of or- ploying the Levenshtein distance, could also lead
ganizations with other subcategories from their location class
(department, hospital) and Stubbs et al. (2017) witness an un- to coreference breaks since quite a few names dif-
certainty for tagging quasi-generic organizations. fer only in one letter (like ‘Lena’ and ‘Lina’).
263
For resolving initials and abbreviations of stituted with an alternative month name trying to
GIVEN, FAMILY and CITY names, we adopt the keep differences between standard varieties (e.g.,
approach by Stubbs et al. (2015b) and use letter- ‘Januar’ (standard German) vs. ‘Jänner’ (Aus-
to-letter mappings generated for each document. trian German) for ‘January’) and also preserve ab-
This means that each name of the respective cat- breviations (e.g., ‘Jan.’).
egory starting with a certain character is replaced
Street names. For names of streets, we rec-
with a name with its first letter corresponding to
ognize the abbreviations ‘S|-str(.)’ (for ‘Straße’
the mapping of this character or in case of initials
(street)) and ‘P|-pl(.)’ (for ‘Platz’ (place)) for
and abbreviations with the mapping solely. For
look-up and coreference. We do not handle inflec-
example ‘Gandalf’ and ‘Gimli’ would be substi-
tions of street names like ‘Ligusterweg(e)s’ (gen-
tuted by names starting with the same character
itive) because, in emails, street names most often
like ‘Bilbo’ and ‘Boromir’. Hence, an initial ‘G.’,
are part of an address and thus lack inflection.
will also be replaced with ‘B.’; it does not have to
The list of surrogates is built from large repos-
be disambiguated and assigned to any of the previ-
itories of German street names11 jointly with
ously occurring names (a task left to a coreference
Austrian ones from different provinces12 that do
resolution module we currently do not provide).
not contain special characters or are composed
We also try to account for frequency distri-
of more than two terms (e.g., ‘Albert-Einstein-
butions of GIVEN, FAMILY, CITY, ORG and
Straße’). Furthermore, we restricted them to
STREET names by constraining these random
contain only names with standard street suffixes
letter-to-letter-mappings to map first letters to let-
(‘-straße’, ‘-weg’, ‘-platz’, etc.), because we had
ters with a similar frequency. In this way, map-
to get rid of village names that do not have any
pings of very common first letters, such as ‘A’ in
named streets (such as with ‘Wegscheid 15’).
case of German first names, to rare ones, like ‘X’,
can be avoided. This approach still allows rare Given Names and Family Names. Common
substitutes for common names. However, we cir- German proper nouns are singular and do not
cumvent the problem of adding ambiguity to the change number. (Duden, 2009, p. 191) Hence, our
text, if we only have few names starting with ‘X’ system does not process any plural forms of fore-
(Stubbs et al. (2015b) also mentioned but did not names or surnames (which rarely may occur), yet
implement this idea). As we map distributions in handles the genitive singular case. Therefore, our
quite a rough way, we do not think that this could system is also capable of resolving coreferences
cause a leak of information of the original text, but between uninflected and inflected genitive forms.
distributional mappings are optional and the user To acquire suitable look-up dictionaries we ex-
may choose the granularity and distribution in his tracted female and male names with their asso-
or her own module or language extension. ciated nicknames from a list of first names.13
These lists are restricted to more or less com-
5.2 Language-Dependent Criteria for German mon forenames in German-speaking countries ex-
Since the categories DATE, STREET, CITY, cept for names with rare first letters, where we in-
GIVEN and FAMILY name, and ORG are affected cluded less frequent names, too. An alternative list
by language-specific influences we implemented of German surnames14 is frequency-independent,
a German language module for these named en- thus includes also lots of uncommon names.
tity types. In contrast to other surrogate genera-
City names. For the CITY category, the geni-
tion systems we know of, our solution takes inflec-
tive singular case is handled too. While person
tion into account (relying on the NLP tool SPACY
names and city, town or village names are mostly
(Honnibal and Montani, 2017)).
11
http://www.datendieter.de/item/Liste_
Dates. The formats we handle include typical von_deutschen_Strassennamen_.csv
12
combinations of day, month and year according http://www.statistik.at/strasse/
suchmaske.jsp
to German formatting style (e.g., days precede 13
ftp://ftp.heise.de/pub/ct/listings/
months), yet also account for different spellings 0717-182.zip
14
(e.g., ‘01.06.2019’ or ‘1.6.19’). Days, months, http://www.namenforschung.
net/fileadmin/user_upload/dfa/
and years occurring in isolation are processed as Inhaltsverzeichnisse_etc/Index_Band_
well. If a month is given in letter format, it is sub- I-V_Gesamt_Stand_September_2016.pdf
264
used without determiner, a few German names Organizations. Similarly, we consider the gen-
for regions (e.g., ‘die Steiermark’, ‘das Drautal’) itive case for organizations. As the same com-
always require a determiner (Duden, 2009, pp. pany or institution in a document might be denoted
299ff.). These names can be of every gender and by different name forms, such as its full name
also pluralia tantum, whereas city, town or vil- (‘Stadtwerke GmbH’), with or without the corpo-
lage names are neuter and singularia tantum (Du- rate form (‘Stadtwerke’) or an acronym (‘STW’),
den, 2009, pp. 160f.). Unfortunately, we cur- with respect to coreference chains a more sophisti-
rently do not dispose of repositories for gender- cated solution is required. For now, we only check
and number-specific CITY names large enough, so for names without a list of corporate forms. The
we rather tolerate possible mistakes than a poten- substitution is performed with a list of German
tial information leakage. company names,19 restricted to names not contain-
In contrast to person names, we consider deriva- ing any GIVEN or FAMILY name. Due to gender
tions of locations and implemented rules for signi- variability and the lack of a list of institutions, we
fying inhabitants (‘die Klagenfurterin’) and adjec- here added fictional acronyms and randomly gen-
tivized toponyms ending on ‘-(i)sch/-erisch/-er’ erated letter combinations.
(‘kärntnerisch’ (carinthian), ‘Wiener Dialekt’ (Vi-
ennese dialect)). We check for coreferences using 5.3 System Architecture
Levenshtein distance on previously seen CITYs. Our system for surrogate generation (see Figure
To catch lemmata occurring later we form appro- 2) accepts any type of text, not only emails, but
priate candidates in a rule-based manner together requires BRAT annotations of pi-relevant entities
with a lexicon look-up. For naming inhabitants, as described in Section 3. It allows for an easy
we only treat the standard forms ending on ‘-er’ adaptation to languages other than German, since
such as ‘Wiener’ (Viennese) and do not care about the base module of the surrogate generation sys-
non-standard names like ‘Hesse’ (Hessian). tem can implement a language module for alterna-
The generation of derivations from the sub- tive languages, too. While language-independent
stitute lemma is restricted to the most common categories (Section 5.1) do not need any further
cases. We produce derivatives by concatenating consideration, this language module has to provide
the lemma and ‘-er’ or ‘-r’, if the lemma ends with allowed date formats and lists with language- and
‘-e’. For adjectivized forms ending with ‘-isch/- category-appropriate substitutes for the language-
sch/-erisch’ that often allow a variation of these dependent classes (Section 5.2). Furthermore, fre-
suffixes (Duden, 2016, p. 685) we decided on gen- quency mappings of first letters may be specified
erating derivatives with ‘-erisch’ (e.g., ‘Wiener- in order to take distributions of names with respect
isch’ (Viennese)). Our system produces the inflec-
tional forms for derivations by copying the origi-
nal inflectional suffix to the generated form.
To maintain local national information we
employed separate lists of location names for
the three major German-speaking countries15 on
which we perform a dictionary look-up to deter-
mine which country the location name is from.
Admittedly, this approach fails, if the place is ei-
ther not mentioned or occurs in multiple countries.
For substitution, we provide cleaner lists contain-
ing only villages, towns and cities for Germany,16
Austria,17 and Switzerland.18
15
http://download.geonames.org/export/ Figure 2: Schematic system architecture and sur-
dump rogate generation workflow; dashed parts optional
16
http://www.fa-technik.adfc.de/code/
opengeodb/PLZ.tab plz/PLZO_CSV_LV03.zip
17 19
http://www.statistik.at/strasse/ https://www.datendieter.de/item/
suchmaske.jsp Liste_von_deutschen_Firmennamen_.txt
18
http://data.geo.admin.ch/ch. extracted from O PEN S TREET M AP (http://www.
swisstopo-vd.ortschaftenverzeichnis_ openstreetmap.org/)
265
to their first letters into account. If a use case 6.2 Frequency Analysis
requires a special treatment of a category, exten-
As Deléger et al. (2014) remark, large fre-
sional functions have to be defined, e.g., for in-
quency imbalances between corpora may influ-
flection generation. Otherwise, the base system re-
ence the performance of machine learning sys-
places the entities with default entries in the substi-
tems. Therefore, we also assessed frequency im-
tute lists, i.e., generally (non-inflected) lemmata.
balances between different corpora: The origi-
6 Evaluation of Pseudonymization nal non-pseudonymized C ODE A LLTAGS+d cor-
pus with hand-annotated pi entities (referred to
6.1 Grammaticality and Acceptability Tests as ORIG), the pseudonymized20 form of the orig-
In a first round of evaluations, we wanted to test inal corpus (PSEUD) and pseudonymized20 C ODE
whether the surrogates we had generated were A LLTAGS+d with automatically recognized pi en-
well-formed in terms of grammaticality and se- tities (PSEUD PIR). For the latter, we retrieved the
mantically ‘natural’ in terms of acceptability. For pi entities by training a model on 9/10 of ORIG
privacy reasons, we refrained from explicitly eval- and applying it to the unseen part on ten different
uating whether coreference relations were pre- folds. We used a system for recognizing privacy-
served, because it is difficult to keep track of them bearing information (PIR) based on N EURO NER
without the original wording. (Dernoncourt et al., 2017; Lee et al., 2016), with
First, the pseudonymized emails were scored slight modifications of its neural network architec-
on both evaluation dimensions on a scale from 1 ture from our side.
(worst) to 5 (best) in packages of about 30 emails Table 3 shows that the number of tokens de-
by different annotators. Each email was annotated clines for both pseudonymized corpora compared
only once. While grammaticality refers to the to the original corpus. Taking a closer look, the
agreement in number, gender and case between the category discrepancy almost entirely results from
generated surrogate and the sentence constituents the category ORG and, to a much lesser extent
the surrogate is embedded in, a surrogate is ac- though, also from STREET and CITY names. We
ceptable if it semantically fits into the surrounding conclude that our substitution dictionaries obvi-
context so that a reasonable semantic interpreta- ously contain shorter names, i.e., entities consist-
tion can be made. For example, ‘We bought the car ing of fewer tokens. Contrasting ORIG and the
at Amici Pizza Express’ is considered as seman- automatically annotated PSEUD PIR, we witness
tically inacceptable because cars normally cannot an increase of pi entities. For one thing, this
be bought at a pizzeria. In contrast common names can be explained by taking the tokens of one en-
replaced with rare ones (e.g., ‘Paris’ with a small tity separately and thus splitting one single entity
village name) are regarded as acceptable. into multiple ones, which is also reflected in the
For the evaluation, the manually tagged 1,390 smaller difference between the number of tokens.
German-language emails of C ODE A LLTAGS+d As this phenomenon especially occurs with URLs
underwent the surrogate generation process twice and PHONE or fax numbers (these categories are
to be sure not to reveal any original pi items to the substituted randomly) it has no effect on surrogate
annotators. The automatically generated output generation. Again, ORGanizations play a major
was checked manually, also substituting phrases role here, because the PIR system achieves the
not covered by our categories, such as course worst results for this category.
names (the corpus contains lots of university re-
lated emails donated by our students). After that it ORIG PSEUD PSEUD PIR
was fed again to the surrogate generation system. # token 151,229 150,166 150,425
With 4.90 for grammaticality and 4.73 for se- # types21 21,159 22,320 22,455
mantic acceptability the results are pretty sound. # pi entities 8,866 8,866 9,427
The lower outcome for semantic acceptability is # pi tokens 12,649 11,586 11,865
mostly due to the surrogates for the ORG category.
Table 3: Quantitative breakdown of the corpora
Occasionally, rather odd combination like ‘Serial
used for evaluation; “#” stands for frequency count
Knitters, IT solutions’ (for ‘Institute for XX, YY-
University’) and substitutes with a different, inac- 20
Processed by the surrogate generation system without
curate function were found. any further reworking.
266
6.3 Automatic Recognition Performance Train|Test Prec Rec F1 p
Further, we followed Yeniterzi et al. (2010) and ORIG | PSEUD 77.97 73.12 74.93
Deléger et al. (2014) and tested the performance PSEUD | ORIG 70.25 61.32 64.57 0.0
ORIG | PSEUD PIR 85.23 75.21 78.39
of the PIR system on the three corpora. Like Yen-
PSEUD PIR| ORIG 67.31 63.21 63.88 0.0
iterzi et al. (2010), we also found better results
for training and testing on the PSEUD corpus than Table 5: Results of the PIR system trained on
on ORIG,22 while the performance difference de- Train and tested on Test (10-fold cross validation
creases for PSEUD PIR and ORIG (see Table 4). without overlap); significance difference (p-value)
Among others, this may be a consequence of the over reverse setting (paired t-test)
frequency imbalance, because the pseudonymized
data contain fewer tokens especially those related diverse. Further, they may contain rarer names.
to the hard to recognize ORGanizations. Both phenomena potentially lead to a decrease of
performance on these data sets when trained on
Corpus Prec Rec F1 p-value
the ORIG corpus. Regarding the drop of the re-
ORIG 83.02 82.26 82.52 verse experiment the fewer occurrences of ‘I’-tags
PSEUD 86.15 86.24 86.07 0.007 (from the BIO format) have an impact, too.
PSEUD PIR 82.26 85.79 83.86 0.063 In contrast to Yeniterzi et al. (2010) and our re-
Table 4: Results of the PIR system on differ- sults, Deléger et al. (2014) report a smaller per-
ent corpora (10-fold cross validation); significance formance difference between training on original
difference (p-value) with respect to ORIG (paired and testing on pseudonymized data and vice versa.
t-test) This is probably due to the fact that they replaced
pi entities with different entities of the same cat-
For training and testing on different corpora, we egory taken from the same original corpus, thus
eliminated the emails found in the test set from almost retaining the original personal information.
the train set in a 10-fold cross-validation man- This approach bears the potential of causing a leak
ner, even if they were pseudonymized in one of of personal information due to categories with lim-
the datasets to avoid any overlap. Again, similar ited occurrences.
to Yeniterzi et al. (2010), training on ORIG and
testing on PSEUD yields significantly better results 7 Conclusion
than the other way round (see Table 5). Also the In this paper, we moved the de-identification prob-
F1 score plunges deeply when training on PSEUD - lem out of the medical domain (cf. also our work
PIR and testing on ORIG compared to ORIG and in this field described in Kolditz et al. (2019)) into
testing on PSEUD PIR. the realm of user-generated content, emails in our
If we subsume the language-independent cate- use case. In particular, we focused on the surro-
gories, because they are randomized and treated gate generation step of that task, i.e., substituting
similarly in surrogate generation, we get compa- named entities bearing privacy-relevant informa-
rable outcomes for all experiments with the PIR tion by synthetic, yet mostly natural surrogates.
system with an F1 score around 90.00. In con- Our main contributions are the specification of
trast, language-dependent categories, with the ex- a language-independent type hierarchy composed
ception of DATEs, which accomplish equivalent of named entities that carry privacy-relevant in-
results for nearly all experiments, too, consistently formation, and the realization of the first German
perform worse compared to training and testing on non-medical surrogate generation pipeline. It is
the same corpus, especially regarding ORGaniza- composed of a language-dependent part for Ger-
tions, CITYs and STREETs. As the results in Ta- man input and a language-independent one which
ble 3 reveal, the amount of word types is higher in can readily be reused for languages other than Ger-
the pseudonymized corpora; hence, they are more man, without any changes.
21
Types of tokens excluding punctuation and stop words. We also ran a series of experiments on emails
22
When the k-fold splits are performed on sentences rather from the German-language C ODE A LLTAGS+d
than on emails the results get notably better achieving 86.24 corpus. In this evaluation of our surrogate gener-
F1 score for ORIG. But for reasons of comparability between
non-pseudonymized and pseudonymized corpora we had to ation system, we found high scores for the gram-
split on emails. maticality and acceptability of the automatically
267
generated surrogates. A frequency analysis of dif- SMS corpus. Language Resources and Evaluation
ferent variants of C ODE A LLTAGS+d revealed a 47(2):299–335.
quantitative imbalance between the original cor- Maximin Coavoux, Shashi Narayan, and Shay B. Co-
pus, the pseudonymized one, and a de-identified hen. 2018. Privacy-preserving neural representa-
variant that was built using an automatic recog- tions of text. In EMNLP 2018 — Proceedings of
nizer for privacy-relevant named entities. Experi- the 2018 Conference on Empirical Methods in Nat-
ural Language Processing. Brussels, Belgium, Oc-
ments on these three corpora further exposed dif- tober 31 - November 4, 2018. pages 1–10.
ferences in recognition performance already dis-
cussed in the literature. Jacob Cohen. 1960. A coefficient of agreement for
nominal scales. Educational and Psychological
Our future work will focus on a more ade- Measurement 20(1):37–46.
quate treatment of German derivational morphol-
ogy and coreferences rooted in varying spellings. Louise Deléger, Todd Lingren, Yizhao Ni, Megan
The main methodological desideratum concerns Kaiser, Laura Stoutenborough, Keith Marsolo,
Michal Kouril, Katalin Molnar, and Imre Solti.
the investigation of ways to deal with organiza- 2014. Preparing an annotated gold standard cor-
tions with different functions in order to improve pus to share with extramural investigators for de-
semantic acceptability. Last but not least, we identification research. Journal of Biomedical In-
will have to demonstrate that the results from the formatics 50:173–183.
small-scale corpus we currently dealt with (C ODE Franck Dernoncourt, Ji Young Lee, Özlem Uzuner,
A LLTAGS+d ) will scale up to much larger docu- and Peter Szolovits. 2017. De-identification of pa-
ment collections (e.g., C ODE A LLTAGXL as de- tient notes with recurrent neural networks. Jour-
nal of the American Medical Informatics Associa-
scribed in Krieg-Holz et al. (2016)).
tion 24(3):596–606.
Acknowledgments Duden. 2009. Die Grammatik: Unentbehrlich für
richtiges Deutsch, volume 4 of Der Duden in 12
We want to thank our RANLP reviewers for point- Bänden. Dudenverlag, Mannheim etc., 8th edition.
ing out improvements of the original submission.
These hints were gratefully acknowledged and in- Duden. 2016. Das Wörterbuch der sprachlichen
Zweifelsfälle: Richtiges und gutes Deutsch, vol-
corporated in the final version of the paper. ume 9 of Der Duden in 12 Bänden. Dudenverlag,
This work was funded by the Austrian Berlin, 8th edition.
HRSM project Kompetenznetzwerk Digitale Edi-
tion (KONDE). Matthew Honnibal and Ines Montani. 2017. SPAC Y 2:
Natural language understanding with Bloom embed-
dings, convolutional neural networks and incremen-
tal parsing. https://spacy.io/.
References
Eytan Adar. 2007. User 4XXXXX9: anonymiz- Chris Jay Hoofnagle, Bart van der Sloot, and Fred-
ing query logs. In Proceedings of the Workshop erik Zuiderveen Borgesius. 2019. The European
on Query Log Analysis: Social and Technological Union general data protection regulation: what it is
Challenges @ WWW 2007. Banff, Alberta, Canada, and what it means. Information & Communications
May 8, 2007. Technology Law 28(1):65–98.
Alyaa Alfalahi, Sara Brissman, and Hercules Dalia- Bryan Klimt and Yiming Yang. 2004. The E NRON
nis. 2012. Pseudonymisation of personal names and corpus: a new dataset for email classification re-
other PHIs in an annotated clinical Swedish cor- search. In ECML 2004 – Proceedings of the 15th
pus. In BioTxtM 2012 — Proceedings of the 3rd European Conference on Machine Learning, Pisa,
Workshop on Building and Evaluating Resources for Italy, September 20-24, 2004. Springer, pages 217–
Biomedical Text Mining @ LREC 2012. Istanbul, 226.
Turkey, May 26, 2012. pages 49–54.
Tobias Kolditz, Christina Lohr, Johannes Hellrich,
David S. Carrell, Bradley A. Malin, John S. Aberdeen, Luise Modersohn, Boris Betz, Michael Kiehntopf,
Samuel Bayer, Cheryl Clark, Benjamin Wellner, and and Udo Hahn. 2019. Annotating German clinical
Lynette Hirschman. 2013. Hiding In Plain Sight: documents for de-identification. In MedInfo 2019 —
use of realistic surrogates to reduce exposure of pro- Proceedings of the 17th World Congress on Medical
tected health information in clinical text. Journal and Health Informatics. Lyon, France, 25-30 August
of the American Medical Informatics Association 2019. IOS Press.
20(2):342–348.
Ulrike Krieg-Holz, Christian Schuschnig, Franz
Tao Chen and Min-Yen Kan. 2013. Creating a live, Matthies, Benjamin Redling, and Udo Hahn. 2016.
public short message service corpus: the NUS C OD E A LLTAG: a German-language e-mail corpus.
268
In LREC 2016 — Proceedings of the 10th Interna- Pontus Stenetorp, Sampo Pyysalo, Goran Topić,
tional Conference on Language Resources and Eval- Tomoko Ohta, Sophia Ananiadou, and Jun’ichi Tsu-
uation. Portorož, Slovenia, 23-28 May 2016. pages jii. 2012. B RAT: a Web-based tool for NLP-assisted
2543–2550. text annotation. In EACL 2012 — Proc. of the 13th
Conf. of the European Chapter of the Association for
Ji Young Lee, Franck Dernoncourt, Özlem Uzuner, and Computational Linguistics: Demonstrations. Avi-
Peter Szolovits. 2016. Feature-augmented neural gnon, France, April 25-26, 2012. pages 102–107.
networks for patient note de-identification. In Clini-
calNLP 2016 — Proceedings of the Clinical Natural Amber Stubbs, Michele Filannino, and Özlem Uzuner.
Language Processing Workshop @ COLING 2016. 2017. De-identification of psychiatric intake
Osaka, Japan, December 11, 2016. pages 17–22. records: overview of 2016 CEGS NGRID Shared
Tasks Track 1. Journal of Biomedical Informatics
Yitong Li, Timothy Baldwin, and Trevor Cohn. 2018. 75(Supplement):S4–S18.
Towards robust and privacy-preserving text repre-
sentations. In ACL 2018 — Proceedings of the 56th Amber Stubbs, Christopher Kotfila, and Özlem Uzuner.
Annual Meeting of the Association for Computa- 2015a. Automated systems for the de-identification
tional Linguistics. Melbourne, Victoria, Australia, of longitudinal clinical narratives: overview of 2014
July 15-20, 2018. volume 2, pages 25–30. i2b2/UTHealth Shared Task Track 1. Journal of
Biomedical Informatics 58(Supplement):S11–S19.
Zengjian Liu, Buzhou Tang, Xiaolong Wang, and
Qingcai Chen. 2017. De-identification of clinical Amber Stubbs and Özlem Uzuner. 2015. Annotating
notes via recurrent neural network and conditional longitudinal clinical narratives for de-identification:
random field. Journal of Biomedical Informatics the 2014 i2b2/UTHealth corpus. Journal of Biomed-
75(Supplement):S34–S42. ical Informatics 58(Supplement):S20–S29.
Harald Lüngen, Michael Beißwenger, Laura Herzberg, Amber Stubbs, Özlem Uzuner, Christopher Kotfila, Ira
and Cathrin Pichler. 2017. Anonymisation of the Goldstein, and Peter Szolovits. 2015b. Challenges
Dortmund Chat Corpus 2.1. In cmccorpora17 — in synthesizing surrogate PHI in narrative EMRs.
Proceedings of the 5th Conference on CMC and So- In Aris Gkoulalas-Divanis and Grigorios Loukides,
cial Media Corpora for the Humanities. Bolzano, editors, Medical Data Privacy Handbook, Springer
Italy, October 3-4, 2017. pages 21–24. International Publishing, pages 717–735.
Ben Medlock. 2006. An introduction to NLP-based Latanya Sweeney. 1996. Replacing personally-
textual anonymisation. In LREC 2006 — Proceed- identifying information in medical records, the
ings of the 5th International Conference on Lan- S CRUB system. In AMIA ’96 — Proceedings of the
guage Resources and Evaluation. Genoa, Italy, 22- 1996 AMIA Annual Fall Symposium. Washington,
28 May, 2006. pages 1051–1056. D.C., USA, October 26-30, 1996. pages 333–337.
Stéphane M. Meystre. 2015. De-identification of un- Latanya Sweeney. 2002. k-anonymity: a model for
structured clinical data for patient privacy protec- protecting privacy. International Journal of Un-
tion. In Aris Gkoulalas-Divanis and Grigorios certainty, Fuzziness and Knowledge-based Systems
Loukides, editors, Medical Data Privacy Handbook, 10(5):557–570.
Springer International Publishing, pages 697–716.
Maaske Treurniet, Orphée De Clercq, Henk van den
Stephen P. Mulligan, Wilson C. Freeman, and Chris D. Heuvel, and Nelleke Oostdijk. 2012. Collecting a
Linebaugh. 2019. Data protection law: an overview. corpus of Dutch SMS. In LREC 2012 — Proceed-
Technical Report CRS Report R45631, Congres- ings of the 8th International Conference on Lan-
sional Research Service. guage Resources and Evaluation. Istanbul, Turkey,
May 21-27, 2012. pages 2268–2273.
Kostas Pantazos, Sören Lauesen, and Sören Lippert.
2011. De-identifying an EHR database: anonymity, Özlem Uzuner, Yuan Luo, and Peter Szolovits. 2007.
correctness and readability of the medical record. In Evaluating the state-of-the-art in automatic de-
MIE 2011 — Proc. of the 23rd Conference of the identification. Journal of the American Medical In-
European Federation of Medical Informatics. Oslo, formatics Association 14(5):550–563.
Norway, August 28-31, 2011. pages 862–866.
Shomir Wilson et al. 2016. The creation and analysis
Namrata Patel, Pierre Accorsi, Diana Z. Inkpen, Cédric of a website privacy policy corpus. In ACL 2016
Lopez, and Mathieu Roche. 2013. Approaches of — Proceedings of the 54th Annual Meeting of the
anonymisation of an SMS corpus. In CICLing 2013 Association for Computational Linguistics. Berlin,
— Proceedings of the 14th International Conference Germany, August 7-12, 2016. pages 1330–1340.
on Computational Linguistics and Intelligent Text
Processing. Karlovasi, Samos, Greece, March 24- Reyyan Yeniterzi, John S. Aberdeen, Samuel Bayer,
30, 2013. Springer, pages 77–88. Benjamin Wellner, Lynette Hirschman, and
Bradley A. Malin. 2010. Effects of personal iden-
Frances Rock. 2001. Policy and practice in the tifier resynthesis on clinical text de-identification.
anonymisation of linguistic data. International Journal of the American Medical Informatics
Journal of Corpus Linguistics 6(1):1–26. Association 17(2):159–168.
269
Lexical Quantile-Based Text Complexity Measure
Abstract
c w
ARI(d) = 4.71 × + 0.5 × − 21.43 (1)
This paper introduces a new approach to esti- w s
mating the text document complexity. Com-
where c refers to the total number of letters in the
mon readability indices are based on average
length of sentences and words. In contrast to
document d, w is the total number of words and s
these methods, we propose to count the num- denotes the total number of sentences in d.
ber of rare words occurring abnormally often Since readability indexes rely on a few basic
in the document. We use the reference corpus factors, precise assessment requires aggregation
of texts and the quantile approach in order to of many scores. Thus, Coh-Metrix-PORT tool
determine what words are rare, and what fre- (Aluisio et al., 2010) includes more than 50 dif-
quencies are abnormal. We construct a general ferent indexes for Portuguese language. The tool
text complexity model, which can be adjusted
is based on Coh-Metrix (Graesser et al., 2004)
for the specific task, and introduce two special
models. The experimental design is based on a principles to estimate complexity and cohesion not
set of thematically similar pairs of Wikipedia only for explicit text, but for the mental represen-
articles, labeled using crowdsourcing. The ex- tation of the document.
periments demonstrate the competitiveness of Readability indexes are interpretable and easy
the proposed approach. to implement. However, the great number of con-
stants tuned specifically for the English language
1 Introduction texts, lack of the semantics consideration and tai-
loring to the US grade level system restrains the
Automated text complexity measurement tools number of possible applications.
have been proposed in order to help teachers to As for the non-English languages, several lex-
select textbooks that correspond to the students’ ical and morphological features for Italian to
comprehension level and publishers to explore solve text simplification problem were presented
whether their articles are readable. Thus, plenty (Brunato et al., 2015), supervised approach in
of readability indexes were developed. Mea- readability estimations was introduced (vor der
sures like Automated Readability Index (Senter Brck et al., 2008) and the complexity estimations
and Smith, 1967), Flesch-Kincaid readability tests for legal documents in Russian were explored
(Flesh, 1951), SMOG index (McLaughlin, 1969), (Dzmitryieva, 2017).
Gunning fog (Gunning, 1952) and etc. use heuris- In this paper we introduce a new approach to
tics based on simple statistics such as total number gauge the complexity of the documents based on
of words, mean number of words per sentence, to- their lexical features. Our research is motivated
tal number of sentences or even number of sylla- by information retrieval applications such as ex-
bles to evaluate how complex given text is. By ploratory search for learning or editorial purposes
combining these statistics with different weight- (Marchionini, 2006; White and Roth, 2009; Palagi
ing factors, readability indexes assign the given et al., 2017). In the exploratory search, the user
document a complexity score, which is, in most needs a hint which of the found documents to read
cases, the approximate representation of the US first, gradually moving from simple to more com-
grade level needed to comprehend the text. For in- plex documents. Reading order optimization is an
stance, an Automated Readability Index (ARI) has alternative way to content consumption that de-
the following form for the document d: parts from the typical ranked lists of documents
270
Proceedings of Recent Advances in Natural Language Processing, pages 270–275,
Varna, Bulgaria, Sep 2–4 2019.
based on their relevance (Koutrika et al., 2015). tion.
The more specific terms document contains, and nd
X
the more rare they are, the more complex the doc- W (d) = w(ti )[c(ti ) > Cγ (ti )] (2)
ument is. To formalize this consideration, we esti- i=1
mate the complexity of each term in the document where [ ] refers to the Iverson notation (i.e.
and then aggregate them to get the complete doc- [true] = 1, [f alse] = 0).
ument complexity score. We use Wikipedia as a By defining weights w(ti ) and complexity
reference collection of moderately complex texts scores c(ti ) for all terms ti specialize the complex-
in order to determine what term frequencies are ity model.
abnormal. Some examples of interpretable weights w(ti )
In section 2 we describe quantile approach to are presented in Table 1.
estimate the single term complexity. We present
highly flexible general model in section 3 and w(ti ) Meaning of w(ti )
models in subsections 3.1 and 3.2. The way of 1 number of complex terms
evaluating the proposed methods is introduced in 1/nd ×100% complex terms percentage
section 4 and the experiments result are provided c(ti ) total complexity
in section 5. c(ti )/nd mean complexity
c(ti ) − Cγ (ti ) excessive complexity
2 Single Term Complexity Estimation
(c(ti ) − Cγ (ti ))/nd mean excessive complexity
Reference collection: Let D denote a reference
Table 1: Weights w(ti ) examples.
collection. Let document d ∈ D consist of terms
t1 , t2 , . . . tnd , where nd refers to the length of doc-
ument d. Each term can be either a single word or
3.1 Distance-Based Complexity Model
a key phrase.
Quantile approach: In general case each term The following model relies on the assumption,
can occur in different complexity states, which proposed in (Birkin, 2007). Consider an arbi-
may depend on a position in text or context sur- trary document d which is the sequence of terms
rounding the term. Each complexity state of the t1 , t2 , . . . tnd . Let r(ti ) be a distance in terms to
term ti standing in position i is described with a the previous occurrence of the same term ti in doc-
term complexity score c(ti ). Consider a complex- ument d. Formally,
ity scores empirical distribution for each term over
r(ti ) = min {i − j | ti = tj }. (3)
the reference collection. Assume that term ti is in 1≤j<i
complex state if its complexity c(ti ) in current text
If i is the first occurrence of term ti in document
position i is greater than γ-quantile Cγ (ti ) of the
d, it means that r(ti ) is undefined. In such cases
distribution over c(ti ), where γ is a hyperparam-
we take r(ti ) equal to nd . Hence, for terms with
eter, responsible for the complexity level. There-
the only occurrence in d complexity scores are the
fore, when estimating complexity score of the doc-
greatest.
ument, we count c(ti ) only for terms ti which are
If term t does not appear in the reference collec-
in the complex state, defined by the γ parameter.
tion, we set Cγ equal to −∞, therefore counting it
For instance, c(ti ) can be a constant, which
as a constantly complex term.
means all terms have identical complexity, or can
Assume that term t in the position i is more
be set equal to 0 if it occurs in the reference collec-
complex than the same term in the position j if
tion and 1 otherwise. In this case, we count new
r(ti ) > r(tj ). Consider there are no separators
terms (for the reference collection) as complex and
between documents in the reference collection, so
all other terms as simple.
it becomes a single document dall . Thus, it is pos-
sible to count distributions of r(t) of each unique
3 General Document Complexity Model
term t in dall and corresponding γ-quantiles Cγ (t)
Document d complexity W (d) can be calculated of these distributions.
by aggregating complexity scores of terms that For the document d, which complexity we try to
form d. In this paper we propose a weighed sum estimate, we calculate rd (ti ) values for all terms
over the complex terms to be the aggregate func- ti ∈ d.
271
We define mean distance rd,i (ti ) for term ti in 0.24
i-th position in the document d as
0.22
Pi
j=1 rd (ti )[ti = tj ]
r̄d,i (ti ) = Pi (4) 0.2
j=1 [ti = tj ]
0.18
which aggregates all occurrences of the term ti
from the document start.
Finally c(ti ) has the form: 5 10 15 20 25 30
Distance r(t)
c(ti ) = r̄(ti ) − r̄d,i (ti ) (5)
Figure 2: Distribution of distances r(t), calculated over
where r̄(ti ) is the mean distance of the reference the complete Wikipedia dataset for the word ‘nlp’.
collection scores r(ti ) for the term ti .
Intuitively, this means, that term is more com-
complexity score c(t) and the γ-quantile we count
plex if it occurs less in reference collection and
is now a constant Cγ .
occurs more in document d.
Hence, the model has the following form:
Figures 1 and 2 show distributions of distances
r(t) for the simple term ‘algebra’ and the com-
nd
plex term ‘nlp’, calculated over the reference col- X 1
W (d) = w(ti ) > Cγ (6)
lection containing 1.5M documents of the Rus- count(ti )
i=1
sian Wikipedia. For the ‘algebra’ term most
occurrences are relatively close to each other, where w(ti ) corresponds to the term weights in-
whether ‘nlp’ occurrences have fairly greater dis- troduced before.
tance scores. Assume the term t1 is more complex than the
term t2 if number of occurrences in the reference
0.1 collection of the term t1 is lesser than the number
0.08 of occurrences of the term t2 .
Let count(t) denote number of occurrences of
0.06 the term t in the reference collection. Thus, the
0.04 complexity score function can be defined as
0.02 1
c(t) = (7)
0 count(t)
5 10 15 20 25 30 so the assumption above is satisfied.
Distance r(t) For each term t we calculate counters count(t)
and complexity scores c(t) over the reference col-
Figure 1: Distribution of distances r(t), calculated over lection. Having the distribution of c(t), we obtain
the complete Wikipedia dataset for the word ‘algebra’.
γ-quantiles Cγ . The described distribution for the
Russian Wikipedia reference collection is shown
So, using the formula for c(ti ) as above and on Figure 3.
choosing weights w(ti ) we get the distance-based Thus, we have defined c(t) for all terms possible
complexity model. and the distribution necessary to count the Cγ . By
varying weights w(ti ) described in section 3, we
3.2 Counter-Based Complexity Model obtain the counter-based model for the complexity
The second model presented in this paper is based estimation.
on the assumption that each term has an inde-
4 Quality Metric
pendent fixed complexity in the whole language.
Thus, in this section we consider not the complex- To measure the quality of proposed algorithms,
ity distribution of a single term, but the general we asked assessors to label 10K pairs of Russian
complexity distribution over all terms in the lan- Wikipedia articles. Assessors were asked to care-
guage. Hence, each term t is assigned the only fully read both articles and to choose which was
272
0.2 5 Experiments
0.15 Two types of experiments were done. In first case
we used full Russian Wikipedia articles dataset
0.1 (1.5M documents) as a reference collection. In
0.05 second type we used only Wikipedia articles
from the math domain. To do that, we built a
0 topic model using ARTM (Additive Regulariza-
10 20 30 tion of Topic Models) technique (Vorontsov and
Number of occurences count(t) Potapenko, 2015), which clusters documents into
monothematic groups.
Figure 3: Distribution of count(t), calculated over
complete Wikipedia articles dataset. 5.1 Complete Wikipedia Dataset
Preprocessing: All Wikipedia articles were lem-
matized (i.e. reduced to normal form). In this
more difficult to comprehend. If person cannot de- experiment we assume term to be either a single
termine which document is more complex, then he word or a bigram (i.e. two words combination). To
was asked to choose ‘documents are equal’ option. extract them, RAKE algorithm (Rose et al., 2010)
If documents in the given pair are from different was used. Hence, each document in the collection
scientific domains, then we ask assessor to choose was turned into the sequence of such terms.
‘invalid pair’ option. Reference collection: Preprocessed Wikipedia
Documents were chosen from math, physics, articles were used as a reference collection. r(t)
chemistry and programming areas. Clustering for every term position and count(t) for every
was performed using the topic modeling technique unique term were counted.
(Hofmann, 1999). BigARTM open-source library Documents to estimate complexity on: We
was used to perform the clustering (Vorontsov used the labeled pairs described in Section 4 to
et al., 2015). Pairs were formed so that both doc- evaluate the models. Accuracy was used as a qual-
uments belong to a single topic and their lengths ity metric.
are almost identical. Examples of document pairs Models to evaluate: Models introduced in
to assess are introduced in Table 2. 3.1 and 3.2 with different w(ti ) parameters were
tested. We took ARI and Flesch-Kincaid readabil-
Document 1 Document 2 Result
ity test as benchmarks.
Matrix Tensor RIGHT
The results of the experiments are introduced in
Neural network Linear regression LEFT
Table 3. Also we tested how the bigrams extrac-
Electric charge Molecule EQUAL tion affects final quality with fixed weight function
Mac OS X Convex Hull INVALID w(t) = c(t)/nd . The results are given in Table 4.
Table 2: Examples of labeled document pairs.
Model w(t) Accuracy
ARI - 46%
Each pair was labeled twice in order to avoid Flesch-Kincaid - 57%
human factor mistakes. We assume that the pair Distance-based c(t) 68%
was labeled correctly if labels were not controver- Distance-based c(t)/nd 71%
sial, i.e. first assessor labeled the first document Counter-based c(t) 77%
as more complex, while second assessor chose the Counter-based c(t)/nd 81%
second document. If one or both grades were ‘doc-
uments are equal’ then we assume the pair to be Table 3: Results of experiment 1 with different weight
correctly labeled. function.
8K pairs out of 10K were labeled correctly and
were used to compare for the different versions of Results show that both distance- and counter-
algorithms. For each we calculated the accuracy based approaches work twice as well as readabil-
score, which is the rate of correctly chosen docu- ity indexes. Counter-based model with w(t) =
ment in the pair. c(t)/nd weights show the best results.
273
Model Terms Accuracy Model w(t) Accuracy
Distance-based Words 63% ARI - 41%
Distance-based Words+Bigrams 71% Flesch-Kincaid - 49%
Counter-based Words+Bigrams 74% Distance-based c(t) 55%
Counter-based Bigrams 81% Distance-based c(t)/nd 61%
Counter-based c(t) 79%
Table 4: Results of experiments 1 with terms differ- Counter-based c(t)/nd 84%
ently defined.
Table 5: Results of experiment 2 on math collection of
Wikipedia articles with different weights.
274
text simplification. Konstantin Vorontsov, Oleksandr Frei, Murat Apishev,
Petr Romov, and Marina Suvorova. 2015. Bi-
A.A. Birkin. 2007. Speech Codes. Hippocrat, Saint- gartm: Open source library for regularized mul-
Peterburg. timodal topic modeling of large collections. In
Dominique Brunato, Felice Dell’Orletta, Giulia Ven- AIST’2015, Analysis of Images, Social networks and
turi, and Simonetta Montemagni. 2015. Design and Texts, pages 370–384. Springer International Pub-
annotation of the first italian corpus for text simpli- lishing Switzerland, Communications in Computer
fication. pages 31–41. and Information Science (CCIS).
Tim vor der Brck, Sven Hartrumpf, and Hermann Hel- Ryen W. White and Resa A. Roth. 2009. Exploratory
big. 2008. A readability checker with supervised Search: Beyond the Query-Response Paradigm.
learning using deep syntactic and semantic indica- Synthesis Lectures on Information Concepts, Re-
tors. trieval, and Services. Morgan and Claypool Publish-
ers.
Aryna Dzmitryieva. 2017. The art of legal writing: A
quantitative analysis of russian constitutional court
rulings. Sravnitel’noe konstitucionnoe obozrenie,
3:125–133.
R. Flesh. 1951. How to test readability. New York,
Harper and Brothers.
Arthur Graesser, Danielle McNamara, Max Louwerse,
and Zhiqiang Cai. 2004. Coh-metrix: Analysis of
text on cohesion and language. Behavior research
methods, instruments, computers : a journal of the
Psychonomic Society, Inc, 36:193–202.
Robert Gunning. 1952. The technique of clear writing.
McGraw-Hill, New York.
Thomas Hofmann. 1999. Probabilistic latent semantic
indexing. In Proceedings of the 22Nd Annual Inter-
national ACM SIGIR Conference on Research and
Development in Information Retrieval, SIGIR ’99,
pages 50–57, New York, NY, USA. ACM.
Georgia Koutrika, Lei Liu, and Steven Simske. 2015.
Generating reading orders over document collec-
tions. In 2015 IEEE 31st International Conference
on Data Engineering, pages 507–518.
Gary Marchionini. 2006. Exploratory search: From
finding to understanding. Commun. ACM, 49(4):41–
46.
G. H. McLaughlin. 1969. Smog grading: A new read-
ability formula. Journal of Reading, 12(8):639–646.
Emilie Palagi, Fabien Gandon, Alain Giboin, and
Raphaël Troncy. 2017. A survey of definitions and
models of exploratory search. In ESIDA17 - ACM
Workshop on Exploratory Search and Interactive
Data Analytics, Mar 2017, Limassol, Cyprus, pages
3–8.
Stuart Rose, Dave Engel, Nick Cramer, and Wendy
Cowley. 2010. Automatic Keyword Extraction from
Individual Documents.
R.J. Senter and E.A. Smith. 1967. Automated readabil-
ity index. AMRL-TR, 66(22).
K. V. Vorontsov and A. A. Potapenko. 2015. Additive
regularization of topic models. Machine Learning,
Special Issue on Data Analysis and Intelligent Opti-
mization with Applications, 101(1):303–323.
275
Demo Application for LETO: Learning Engine Through Ontologies
Suilan Estevez-Velarde1 , Andrés Montoyo2 , Yudivián Almeida-Cruz1 ,
Yoan Gutiérrez3 , Alejandro Piad-Morffis1 , and Rafael Muñoz2
1
School of Math and Computer Science, University of Havana, Cuba
{sestevez,yudy,apiad}@matcom.uh.cu
2
Department of Languages and Computing Systems, University of Alicante, Spain
3
U.I. for Computer Research (IUII), University of Alicante, Spain
{montoyo,ygutierrez,rafael}@dlsi.ua.es
276
Proceedings of Recent Advances in Natural Language Processing, pages 276–284,
Varna, Bulgaria, Sep 2–4 2019.
Memory
learn connect with
most relevant your experience
text processing previous new
memorize knowledge knowledge
(i.e. read text)
image
recognition
merge
learning process
cus on extracting knowledge and exploiting the that can be crosschecked in the future.
semi-structured format of web resources , e.g, The rest of the paper is organized as follows
ARTEQUAKT (Alani et al., 2003), SOBA (Buite- to facilitate a detailed description of our proposal:
laar et al., 2006) and WEB->KB (Craven et al., Section 2 describes the proposed architecture of
2000). In order to extract relevant knowl- a general framework for knowledge discovery. In
edge from natural language text, NLP techniques Section 3 we present an application of LETO to
have been introduced in systems such as OP- a specific knowledge discovery problem combin-
TIMA (Kim et al., 2008) and ISODLE (Weber ing Twitter and IMDB. Finally, in Section 4 we
and Buitelaar, 2006). Natural language features present the main conclusions of the research and
can be used to build rule-based systems (e.g., outline possible future works.
OntoLT (Buitelaar and Sintek, 2004)) or systems
based on statistical or probabilistic models trained 2 Learning Engine Through
on NLP corpora, such as LEILA (Suchanek Ontologies (LETO)
et al., 2006) or Text2Onto (Cimiano and Völker, In this section we present LETO, a general archi-
2005). Some systems address the issue of infer- tecture for a framework designed to discover rel-
ring more abstract knowledge from the extracted evant knowledge from a variety of data sources,
facts, often using unsupervised techniques to dis- both structured and unstructured.
cover inherent structures. Relevant examples of The LETO framework is divided into 6 mod-
this approach are OntoGain (Drymonas et al., ules, which are interrelated. Each module has a
2010), ASIUM (Faure and Poibeau, 2000) and specific responsibility defining the inputs and out-
BOEMIE (Castano et al., 2007). puts that establish the intercommunication among
Most of the mentioned systems focus on one it- the rest of the modules within the framework. Fig-
eration of the extraction process. However, more ure 2 shows a general overview of the framework.
recent approaches, like NELL (Mitchell et al.,
2018), attempt to learn continuously from a stream As shown in Figure 2 the top layer (Data
of web data, and increase over time both the Sources) represents the sources of data that serve
amount and the quality of the knowledge discov- as input for the framework. The middle layer con-
ered. tains the Main Modules, which perform the pro-
One of the main characteristics of LETO, in cessing of the input data to extract and discover
contrast to similar proposals in the literature (such the relevant knowledge latent in this data. Figure 2
as NELL (Mitchell et al., 2018) or BOEMIE (Peta- also shows the subprocesses that occur inside each
sis et al., 2011)), is the explicit management of module. The main modules always communicate
separated pieces of knowledge. By isolating the with each other by sharing ontologies. The follow-
knowledge for different domains, it is possible to ing sections 2.1, 2.2 and 2.3 explain in detail the
apply different techniques and/or parameters as inner workings of the main modules. The bottom
appropriate. Besides, this allows the temporal ex- layer (Backend) contains modules used by the rest
istence of contradictions or unreliable information of framework:
277
Structured Unstructured
Data Sources
...
Structured Semantic Text Social Web
Databases
documents networks documents networks pages
Main Modules
Structured Data Knowledge Unstructured Data
Processing Discovery Processing
Mapping Tagging
Generation Evaluation
Sensors Structure Knowledge
Backend
Algorithms Long Term Organizational
Library Memory Ontology
Algorithms Library: Contains different algo- an schema of this module. This module performs
rithms and mathematical models for solv- two main tasks:
ing specific problems, along with associated
metadata. Mapping: Since there are many different struc-
tured formats, the first stage of this module
Long Term Memory: Contains all the knowl- is to convert any of these representations into
edge accumulated by the other modules, in a standard representation for internal use, in
the form of individual ontologies with meta- the form of an ontology, using a mapping pro-
data that describes their content. cess (Choi et al., 2006; Y. An and Mylopou-
los, 2006; Noy and Musen, 2003). The cur-
Organizational Ontology: An internal represen- rent implementation infers classes and rela-
tation of the framework’s components in an tions from CSV or TSV input files using a
ontological format which enables the auto- rule-based approach, and outputs and ontol-
matic construction of the user interface. ogy in OWL format.
2.1 Structured Data Processing Tagging: This step attaches several tags, such as
source, domain, topic and reliability to the
This module is responsible for processing struc-
mapped ontology. This tags can be either
tured data. Sources for structured data are avail-
inferred automatically (e.g., the domain and
able online in different formats. Among the dif-
reliability) or provided by the user (e.g., the
ferent types of structures for representing informa-
source). The current implementation requires
tion, such as relational databases, concept maps,
a manual input by a domain expert.
knowledge graphs, and others, LETO proposes the
use of ontologies for their semantic richness. On-
tologies were chosen because they are more ex- 2.2 Unstructured Data Processing
pressive than other DTO (Data Transfer Object) The sources for unstructured data are extremely
formats. varied in format and computational representation.
The general pipeline that this module performs Text is one of the most common forms for storing
can be thought of as a classic Extract, Transform and communicating human knowledge, but pic-
and Load process (ETL) (Vassiliadis, 2009; Her- tures, sound files, and videos are also interesting
mida et al., 2012). Afterwards, the normalized and and increasingly popular forms of communication.
tagged block of knowledge (stored as an ontology) Also, in contrast with structured sources, there is
is handled to the knowledge processing module, a lot of variety in the level of reliability and com-
for further refinement and storage. Figure 3 shows pleteness of unstructured sources.
278
Data Sources Structured Data Processing Structural Level: The data tokens extracted from
the original source are processed as a group
Converting to
Internal Format
Adding Tags
and Metadata to find an underlying structure. Techniques
Knowledge
Discovery
Structured documents
origin implemented in this stage include Latent Se-
read input
formats
domain
Semantic networks
write output
complexity
mantic Analysis (LSA) (Hofmann, 2017),
ontolgy
279
matching among entities, relations and in- is known to be of high quality. These steps are rep-
stances in two or more ontologies that are resented in the figure with the numbers 1a and 1b
deemed similar (Shvaiko and Euzenat, 2013). and performed in LETO using the Structured Data
Processing module (see Fig. 3).
Evaluation: After the new ontology is created, The next step involves the processing of a con-
this step provides quality evaluation metrics tinuous stream of Twitter messages (2a). These
that assert the reliability, completeness or are obtained through the standard Twitter query
soundness of the new knowledge. These met- API, filtering with the hashtag #Oscars, which
rics are based on comparing the new ontology returned 3375 messages that span a period of
with the existing knowledge. 2 weeks. Using standard NLP techniques,
each tweet is processed to obtain named enti-
Inputs Knowledge Discovery ties (Nadeau and Sekine, 2007) and an opinion la-
Generate new Evaluate bel (Pang et al., 2008; Liu, 2012) (2b). The entity
Unstructured
Data knowledge knowledge
Long Term
sensor was implemented using spaCy (Honnibal
Memory
Processing
merging of find
contradictions
ontologies
and Montani, 2017), which returned 524 unique
find
create new
ontologies ontologies inconsistencies
from others
PERSON instances, from a total of 1961 PER-
Structured
Data SON mentions. The document level emotion sen-
Processing
Generation
Evaluation
280
2a
#Oscar2018
1a
IMDb Database
Match named entities with
Film or Actor instances.
3b
2c Match equivalent entities Store ontologies
in multiple tweets Result:
Decide tools and data Actor and Film instances
3a
Algorithms
2d
Remove outliers and
reduce entities count 3d
2f
Domain: Cinema
Result: Entities mentioned together
2e
with aggregated opinion Confidence: 0.3
Source: twitter.com
Domain: Cinema
Date: 2018
Confidence: 0.3
3e
Source: twitter.com
repeat periodically
Date: 2018
Figure 6: Schematic representation of the process for mapping emotions in reaction to movie reviews
with IMDB.
Figure 7: Example execution of the Entity Sensor for one tweet. A similar interface allows the batch
execution for a collection of tweets.
281
Figure 8: Main UI of LETO, specifically the Task Management view.
282
ied sensors, and more complex mechanisms for Paul Buitelaar and Michael Sintek. 2004. Ontolt ver-
knowledge integration (e.g., ontology merging and sion 1.0: Middleware for ontology extraction from
text. In Proc. of the Demo Session at the Interna-
mapping processes). Another line for future re-
tional Semantic Web Conference.
search is related to context mismatch and recogni-
tion, specifically in the Unsupervised Processing Silvana Castano, Sofia Espinosa, Alfio Ferrara, Van-
gelis Karkaletsis, Atila Kaya, Sylvia Melzer, Ralf
Module. This process is necessary for accurately
Möller, Stefano Montanelli, and Georgios Peta-
matching portions of unstructured text to sections sis. 2007. Ontology dynamics with multimedia
of an already stored ontology. We will also focus information: The boemie evolution methodology.
on extending the automation processes currently In International Workshop on Ontology Dynamics
available in LETO. (IWOD-07). page 41.
Balakrishnan Chandrasekaran. 1986. Generic tasks
Acknowledgments in knowledge-based reasoning: High-level build-
ing blocks for expert system design. IEEE expert
This research has been supported by a Carolina 1(3):23–30.
Foundation grant in agreement with University Namyoun Choi, Il-Yeol Song, and Hyoil Han. 2006. A
of Alicante and University of Havana. More- survey on ontology mapping. ACM Sigmod Record
over, it has also been partially funded by both 35(3):34–41.
aforementioned universities, the Generalitat Va- Philipp Cimiano, Alexander Mädche, Steffen Staab,
lenciana and the Spanish Government through the and Johanna Völker. 2009. Ontology learning. In
projects SIIA (PROMETEU/2018/089), LIVING- Handbook on ontologies, Springer, pages 245–267.
LANG (RTI2018-094653-B-C22) and INTEGER Philipp Cimiano and Johanna Völker. 2005. text2onto.
(RTI2018-094649-B-I00). In International Conference on Application of Nat-
ural Language to Information Systems. Springer,
pages 227–238.
References Mark Craven, Dan DiPasquo, Dayne Freitag, Andrew
Harith Alani, Sanghee Kim, David E Millard, Mark J McCallum, Tom Mitchell, Kamal Nigam, and Seán
Weal, Wendy Hall, Paul H Lewis, and Nigel R Shad- Slattery. 2000. Learning to construct knowledge
bolt. 2003. Automatic ontology-based knowledge bases from the world wide web. Artificial intelli-
extraction from web documents. IEEE Intelligent gence 118(1-2):69–113.
Systems 18(1):14–21. James Davidson, Benjamin Liebald, Junning Liu,
Palash Nandy, Taylor Van Vleet, Ullas Gargi, Su-
Nathalie Aussenac-Gilles and Marie-Paule Jacques. joy Gupta, Yu He, Mike Lambert, Blake Livingston,
2006. Designing and evaluating patterns for ontol- et al. 2010. The youtube video recommendation sys-
ogy enrichment from texts. In International Con- tem. In Proceedings of the fourth ACM conference
ference on Knowledge Engineering and Knowledge on Recommender systems. ACM, pages 293–296.
Management. Springer, pages 158–165.
Euthymios Drymonas, Kalliopi Zervanou, and Euripi-
Ken Barker, Bhalchandra Agashe, Shaw Yi Chaw, des GM Petrakis. 2010. Unsupervised ontology ac-
James Fan, Noah Friedland, Michael Glass, Jerry quisition from plain texts: the OntoGain system. In
Hobbs, Eduard Hovy, David Israel, Doo Soon Kim, International Conference on Application of Natural
et al. 2007. Learning by reading: A prototype sys- Language to Information Systems. Springer, pages
tem, performance baseline and lessons learned. In 277–287.
AAAI. volume 7, pages 280–286.
S. Estevez-Velarde, Y. Gutierrez, A. Montoyo, A. Piad-
Eva Blomqvist. 2009. Ontocase-automatic ontology Morffis, R. Munoz, and Y. Almeida-Cruz. 2018.
enrichment based on ontology design patterns. In Gathering object interactions as semantic knowledge
International Semantic Web Conference. Springer, (accepted). In Proceedings of the 2017 International
pages 65–80. Conference on Artificial Intelligence (ICAI’17).
David Faure and Thierry Poibeau. 2000. First ex-
Sergey Brin and Lawrence Page. 1998. The anatomy of periments of using semantic knowledge learned by
a large-scale hypertextual web search engine. Com- asium for information extraction task using intex.
puter networks and ISDN systems 30(1-7):107–117. In Proceedings of the ECAI workshop on Ontology
Learning.
Paul Buitelaar, Philipp Cimiano, Stefania Racioppa,
and Melanie Siegel. 2006. Ontology-based informa- Javi Fernández, Yoan Gutiérrez, José M Gómez, and
tion extraction with soba. In Proceedings of the In- Patricio Martı́nez-Barco. 2015. Social rankings:
ternational Conference on Language Resources and análisis visual de sentimientos en redes sociales.
Evaluation (LREC). Procesamiento del Lenguaje Natural 55:199–202.
283
David Ferrucci, Anthony Levas, Sugato Bagchi, David Natalya F Noy and Mark A Musen. 2003. The
Gondek, and Erik T Mueller. 2013. Watson: beyond prompt suite: interactive tools for ontology merg-
jeopardy! Artificial Intelligence 199:93–105. ing and mapping. International Journal of Human-
Computer Studies 59(6):983–1024.
Abhishek Gattani, Digvijay S Lamba, Nikesh Garera,
Mitul Tiwari, Xiaoyong Chai, Sanjib Das, Sri Sub- Natalya Fridman Noy, Mark A Musen, et al. 2000.
ramaniam, Anand Rajaraman, Venky Harinarayan, Algorithm and tool for automated ontology merg-
and AnHai Doan. 2013. Entity extraction, link- ing and alignment. In Proceedings of the 17th Na-
ing, classification, and tagging for social media: tional Conference on Artificial Intelligence (AAAI-
a wikipedia-based approach. Proceedings of the 00). Available as SMI technical report SMI-2000-
VLDB Endowment 6(11):1126–1137. 0831.
Richard Gross. 2015. Psychology: The science of mind Bo Pang, Lillian Lee, et al. 2008. Opinion mining and
and behaviour 7th edition. Hodder Education. sentiment analysis. Foundations and Trends
R in In-
formation Retrieval 2(1–2):1–135.
Q Guo, W Wu, DL Massart, C Boucon, and S De Jong.
2002. Feature selection in principal component F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
analysis of analytical data. Chemometrics and In- B. Thirion, O. Grisel, M. Blondel, P. Pretten-
telligent Laboratory Systems 61(1-2):123–132. hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-
sos, D. Cournapeau, M. Brucher, M. Perrot, and
Jesús M Hermida, Santiago Meliá, Jose-Javier E. Duchesnay. 2011. Scikit-learn: Machine learning
Martı́nez, Andrés Montoyo, and Jaime Gómez. in Python. Journal of Machine Learning Research
2012. Developing semantic rich internet applica- 12:2825–2830.
tions with the s m 4ria extension for oide. In Inter-
national Conference on Web Engineering. Springer, Georgios Petasis, Vangelis Karkaletsis, Georgios
pages 20–25. Paliouras, Anastasia Krithara, and Elias Zavitsanos.
2011. Ontology population and enrichment: State
Thomas Hofmann. 2017. Probabilistic latent semantic of the art. In Knowledge-driven multimedia infor-
indexing. In ACM SIGIR Forum. ACM, volume 51, mation extraction and ontology evolution. Springer-
pages 211–218. Verlag, pages 134–166.
Matthew Honnibal and Ines Montani. 2017. spacy 2:
Pavel Shvaiko and Jérôme Euzenat. 2013. Ontology
Natural language understanding with bloom embed-
matching: state of the art and future challenges.
dings, convolutional neural networks and incremen-
IEEE Transactions on knowledge and data engi-
tal parsing. To appear .
neering 25(1):158–176.
Murphy Kevin. 2012. Machine learning: a probabilis-
Fabian M Suchanek, Georgiana Ifrim, and Gerhard
tic perspective.
Weikum. 2006. Leila: Learning to extract informa-
Sang-Soo Kim, Jeong-Woo Son, Seong-Bae Park, Se- tion by linguistic analysis. In Proceedings of the
Young Park, Changki Lee, Ji-Hyun Wang, Myung- 2nd Workshop on Ontology Learning and Popula-
Gil Jang, and Hyung-Geun Park. 2008. Optima: An tion: Bridging the Gap between Text and Knowl-
ontology population system. In 3rd Workshop on edge. pages 18–25.
Ontology Learning and Population (July 2008).
Joseph Turian, Lev Ratinov, and Yoshua Bengio. 2010.
Bing Liu. 2012. Sentiment analysis and opinion min- Word representations: a simple and general method
ing. Synthesis lectures on human language tech- for semi-supervised learning. In Proceedings of the
nologies 5(1):1–167. 48th annual meeting of the association for compu-
tational linguistics. Association for Computational
T. Mitchell, W. Cohen, E. Hruschka, P. Talukdar, Linguistics, pages 384–394.
B. Yang, J. Betteridge, A. Carlson, B. Dalvi,
M. Gardner, B. Kisiel, J. Krishnamurthy, N. Lao, Panos Vassiliadis. 2009. A survey of extract–
K. Mazaitis, T. Mohamed, N. Nakashole, E. Pla- transform–load technology. International Journal
tanios, A. Ritter, M. Samadi, B. Settles, R. Wang, of Data Warehousing and Mining (IJDWM) 5(3):1–
D. Wijaya, A. Gupta, X. Chen, A. Saparov, 27.
M. Greaves, and J. Welling. 2018. Never-
ending learning. Commun. ACM 61(5):103–115. Nicolas Weber and Paul Buitelaar. 2006. Web-based
https://doi.org/10.1145/3191513. ontology learning with isolde. In Proc. of the Work-
shop on Web Content Mining with Human Lan-
Andrés Montoyo, Patricio Martı́Nez-Barco, and guage at the International Semantic Web Confer-
Alexandra Balahur. 2012. Subjectivity and senti- ence, Athens GA, USA. volume 11.
ment analysis: An overview of the current state of
the area and envisaged developments. A. Borgida Y. An and J. Mylopoulos. 2006. Build-
ing semantic mappings from databases to ontolo-
David Nadeau and Satoshi Sekine. 2007. A sur- gies. volume 21st National Conference on Artificial
vey of named entity recognition and classification. Intelligence (AAAI 06).
Lingvisticae Investigationes 30(1):3–26.
284
Sentence Simplification for Semantic Role Labelling and Information
Extraction
Richard Evans Constantin Orăsan
Research Institute in Information Research Institute in Information
and Language Processing and Language Processing
University of Wolverhampton University of Wolverhampton
United Kingdom United Kingdom
r.j.evans@wlv.ac.uk c.orasan@wlv.ac.uk
285
Proceedings of Recent Advances in Natural Language Processing, pages 285–294,
Varna, Bulgaria, Sep 2–4 2019.
eye tracking (Klerke et al., 2015; Timm, 2018), text simplification system for this purpose. Ex-
and reading comprehension testing (Orăsan et al., trinsic evaluation offers the possibility of meet-
2018) to evaluate text simplification systems. ing this requirement. Text simplification has also
There are several challenges in these ap- been claimed to improve automatic text process-
proaches to evaluation. The development of gold ing (e.g. Vickrey and Koller, 2008; Evans, 2011;
standards in text simplification is problematic be- Hasler et al., 2017), though the evidence for this
cause they are difficult to produce and numerous has been fairly limited. In this paper, we explore
variant simplifications are acceptable. As a result, whether syntactic simplification can facilitate two
existing metrics may not accurately reflect the use- NLP tasks: semantic role labelling (SRL) and in-
fulness of the simplification system being evalu- formation extraction (IE).
ated. Even when there are detailed guidelines for In Section 2 of this paper, we present an
the simplification task, there is still likely to be a overview of previous related work. In Section 3,
variety of means by which a human might sim- we present an overview of Evans and Orăsan’s
plify a text to produce a reference simplification. (2019) method for sentence simplification, which
Further, due to the difficulty of the human simpli- is the simplification method used in our current pa-
fication task, it may be that evaluation measures per. In Section 4, we present each of the extrin-
such as BLEU and SARI are unable to exploit a sic evaluation experiments based on SRL (Section
sufficiently large set of reference simplifications. 4.1) and IE (Section 4.2). Each of these sections
Evaluation of text simplification methods us- describes the task, the test data used, the NLP sys-
ing automatic readability metrics is problematic tem whose output is used for extrinsic evaluation
because the extent to which all but a handful of of the sentence simplification system, our moti-
readability metrics correlate with human reading vation for considering that accuracy of the NLP
comprehension is uncertain. Evaluation via opin- system may be improved via a preprocessing step
ion surveys of readers is difficult because partici- in which sentence simplification is performed, the
pants may have varying expectations about the up- evaluation method, our results, and a discussion of
per and lower limits of sentence complexity, mak- the results. In Section 5, we draw conclusions and
ing responses to Likert items unreliable. Partici- consider directions for future work.
pants also vary in terms of linguistic ability and
personal background knowledge. These variables, 2 Related Work
which affect reading behaviour and may affect re- Chandrasekar and Srinivas (1997) hypothesised
sponses to opinion surveys, are difficult to control. that approaches to sentence simplification may
When using methods such as eye tracking to evoke improvements in subsequent text processing
evaluate text simplification, previous work has tasks. In previous work, researchers have sought
shown that differences in reading behaviour de- to determine whether or not a preprocessing step
pend on participants’ reading goals (Yeari et al., based on text simplification can facilitate subse-
2015). This variable is usually controlled by ask- quent natural language processing. In the current
ing participants to respond to text-related opinion paper, our concern is to investigate the impact of a
surveys or multiple choice reading comprehension system simplifying sentences which contain com-
questions. One adverse effect of this is that these pound clauses. Hogan (2007) and Collins (1999)
evaluations may be of limited validity when con- observed that, for dependency parsers, dependen-
sidering the usefulness of system output for other cies involving coordination are identified with by
purposes. While we may learn whether a sentence far the worst accuracy of any dependency type
simplification method improves participants’ per- (F1 -score ≈ 61%). This is one factor motivating
formance in answering short reading comprehen- our research in this direction.
sion questions, it is not clear whether similar ben- Sentence simplification has also been applied
efits would be obtained in terms of readers’ abili- as a preprocessing step in neural machine transla-
ties to be entertained by the text or to understand it tion and hierarchical machine translation (Hasler
well enough to be able to summarise it for friends. et al., 2017). In their approach, the approach to
Given that text simplification is usually made sentence simplification was sentence compression.
for a particular purpose, the evaluation method One contribution of our current paper is an investi-
should offer insights into the suitability of the gation of the use of an information preserving ap-
286
proach to sentence simplification as a preprocess- The algorithm (Algorithm 1) integrates a sen-
ing step in the NLP applications. tence transformation function which implements
Vickrey and Koller (2008) applied their sen- the transformation schemes listed in Table 1.
tence simplification method to improve perfor-
Input: Sentence s0 , containing at least one
mance on the CoNLL-2005 shared task on SRL.3
sign of syntactic complexity of class c,
For sentence simplification, their method exploits
where c ∈ {CEV, SSEV}.
full syntactic parsing with a set of 154 parse tree
Output: The set of sentences A derived from
transformations and a machine learning compo-
s0 , that have reduced propositional
nent to determine which transformation operations
density.
to apply to an input sentence. They find that a SRL
1 The empty stack W ;
system based on a syntactic analysis of automati-
2 O ← ∅;
cally simplified versions of input sentences outper-
3 push(s0 , W );
forms a strong baseline. In their evaluation, Vick-
4 while isEmpty(W ) is f alse do
rey and Koller (2008) focus on the overall perfor-
5 pop(si , W );
mance of their SRL system rather than on the par-
6 if si contains a sign of syntactic
ticular contribution made by the sentence simpli-
complexity of class c (specified in Input)
fication method. As noted earlier, in our current
then
paper, we isolate sentence simplification as a pre-
7 si1 , si2 ← transf ormc (si );
processing step and investigate its impact on sub-
8 push(si1 , W );
sequent NLP tasks.
9 push(si2 , W );
3 Sentence Simplification System 10 else
11 O ← O ∪ {si }
Evans and Orăsan (2019) presented an itera- 12 end
tive rule-based method for sentence simplifica- 13 end
tion based on a shallow syntactic analysis step. Algorithm 1: Sentence simplification algorithm
Their system transforms input sentences contain-
ing compound clauses and complex NPs into se- In its original implementation, the transform
quences of simpler sentences that do not contain function (line 7 of Algorithm 1) included 28 sen-
these types of syntactic complexity. tence simplification rules to implement one trans-
The first stage of sentence simplification is a formation scheme simplifying compound clauses
shallow syntactic analysis step which tags tex- and 125 rules implementing three transforma-
tual markers of syntactic complexity, referred to tion schemes simplifying sentences which con-
as signs, with information about the syntactic con- tain complex NPs. Evaluation of the method re-
stituents that they coordinate or of which they are vealed that simplification of sentences containing
boundaries. The signs of syntactic complexity are complex NPs was significantly less reliable than
a set of conjunctions, complementisers, wh-words, simplification of sentences containing compound
punctuation marks, and bigrams consisting of a clauses. For this reason, in the extrinsic evalua-
punctuation mark followed by a lexical sign. In the tions presented in this paper, we deactivated the
analysis step, syntactic constituents are not identi- rules simplifying sentences that contain complex
fied. It is only the signs which are tagged. The NPs. Each of the remaining implemented rules in-
automatic sign tagger was developed by Dornescu cludes a rule activation pattern which, when de-
et al. (2013). In their scheme, clause coordinators tected in the input sentence, triggers an associated
are tagged CEV4 while the left boundaries of sub- transformation operation. Table 1 presents the
ordinate clauses are tagged SSEV.5 transformation scheme used to simplify compound
After shallow syntactic analysis of the sen- clauses and an example of the sentence transfor-
tence, an iterative algorithm is applied to sentences mation that it makes. Input sentences are trans-
containing compound clauses and complex NPs. formed if they match any of the rule activation
3
patterns, which are expressed in terms of partic-
http://www.lsi.upc.edu/˜srlconll/
spec.html. Last accessed 14th May 2019.
ular words, parts of speech, and tagged signs of
4
Coordinator of Extended projections of a Verb. syntactic complexity. Each application of a rule
5
Start of Subordinate Extended projection of a Verb. transforms a single input sentence into two sim-
287
Scheme Input Sentence Output Sentence 1 Output Sentence 2
{They were formally found not
{They were formally found {They were formally found
guilty by the recorder Michael
not guilty by the recorder not guilty by the recorder
A [B CEV C] D. Gibbon QC after}A [a witness,
Michael Gibbon QC after}A Michael Gibbon QC after}A
→ A B D. A C who cannot be identified, withdrew
a witness, who cannot be prosecutor Susan Ferrier
D. from giving evidenceB andCEV
identified, withdrew from offered no further evidenceC
prosecutor Susan Ferrier offered
giving evidenceB {}D . {}D
no further evidenceC ]{}D .
Table 1: Sentence transformation scheme used to simplify sentences containing compound clauses
pler sentences which are added to the working set the higher-numbered arguments”6 (Palmer et al.,
(stack W in Algorithm 1). 2005). The scheme includes semantic roles for
The iterative nature of the algorithm enables it “general, adjunct-like arguments” providing infor-
to convert complex sentences containing multiple mation on the verb’s cause [AMCAU], direction
signs of syntactic complexity such as (1) into the [AMDIR], discourse relations [AMDIS], loca-
sequence of simple sentences (2). tion [AMLOC], manner [AMMNR], modal func-
(1) Kattab, of Eccles, Greater Manchester, was required
tion7 [AMMOD], negation [AMNEG], purpose
to use diluted chloroform water in the remedy, but the [AMPNC], and time [AMTMP], among others.
pharmacy only kept concentrated chloroform, which For extrinsic evaluation of the sentence simplifi-
is 20 times stronger.
cation method, we focused on verbal predicates8 ,
(2) a. Kattab, of Eccles, Greater Manchester, was re- their arguments, and the nine listed adjunct-like ar-
quired to use diluted chloroform water in the gument types.
remedy.
b. The pharmacy only kept concentrated chloro- Table 2 provides an example of SRL to analyse
form. sentence (3).
c. Concentrated chloroform is 20 times stronger.
(3) When Disney offered to pay Mr. Steinberg a pre-
4 Experimental Setup mium for his shares, the New York investor didn’t
demand the company also pay a premium to other
We evaluated the sentence simplification method shareholders.
extrinsically via two NLP applications. In each
The table contains a row of information about
case, the application was treated as a black box.
the semantic roles associated with each of the four
We compared performance of the system when
main verbs occurring in the sentence. For exam-
processing input in its original form and in an au-
ple, it encodes information about the agent (the
tomatically simplified form generated by the sim-
New York investor), patient or theme (the company
plification method. As noted in Section 3, our
also pay a premium to other shareholders), time
approach to sentence simplification is syntactic
(When Disney offered to pay Mr. Steinberg a pre-
rather than lexical. As they are based to some ex-
mium for his shares), and negation (n’t) of the verb
tent on exact string matching, the experiments de-
demand.
scribed in this paper would be unsuitable for eval-
Test Data. No suitable test data exist to evalu-
uation of lexical simplification systems.
ate a SRL system as a means of extrinsically eval-
4.1 Semantic Role Labelling uating the sentence simplification method. Al-
Semantic role labelling (SRL) is the task of au- though annotated data from the CONLL-2004/59
tomatically detecting the different arguments of shared tasks on SRL are available, this test data
predicates expressed in input sentences. We evalu- is available only for the original versions of input
ated a system performing SRL in accordance with sentences and not for simplified versions which
the Propbank formalism. In this scheme, an “in- may be generated using sentence simplification
dividual verb’s semantic arguments are numbered, systems. Given that it is difficult to map verbs,
beginning with zero. For a particular verb, [A0] 6
Such as [A2], etc.
7
is generally the argument exhibiting features of Applicable to verbs.
8
a Prototypical Agent (Dowty, 1991), while [A1] As opposed to prepositional, adjectival, or other types of
predicate.
is a Prototypical Patient or Theme. No consis- 9
http://www.lsi.upc.edu/˜srlconll/
tent generalizations can be made across verbs for home.html. Last accessed 23rd May 2019.
288
A0 V A1 A2 A3 AMDIS AMNEG AMTMP
to pay Mr.
Disney offered Steinberg a
premium for
his shares
Disney pay his shares Mr. Steinberg a premium
When Disney
the company
the New offered to pay
demand also pay a pre- n’t
York in- Mr. Steinberg
mium to other
vestor a premium for
shareholders
his shares
their arguments, and the semantic labels of these the test data. Table 3 contains an example of
arguments from sentences in their original form the semantic roles labelled in one of the test sen-
to groups of sentences in their automatically gen- tences that we used. In this table, arguments iden-
erated simplifications, we developed a new set of tified more accurately in simplified sentences are
test data for this purpose. We used a 7270-token underlined. For cases in which the SRL performed
collection of news articles from the METER cor- by Senna differed when processing the original
pus (Gaizauskas et al., 2001) to derive a new man- and automatically simplified versions of input sen-
ually annotated data set. The original version of tences, we manually inspected the two analyses,
this dataset contains 265 sentences while the auto- and recorded the number of cases for which SRL
matically simplified one contains 470 sentences. of the original sentence was superior to that of the
NLP System. We made our extrinsic evaluation simplified sentence, and vice versa. The inspec-
of the sentence simplification method using Senna tion was made by a single annotator. In future
(Collobert et al., 2011), a SRL system which tags work, we will seek to employ additional annota-
predicates and their arguments in accordance with tors for this task.
the formalism used in Propbank Results. Our manual evaluation of output from
Motivation. In our previous work (Evans and Senna revealed that 86.39% (1707) of the argu-
Orăsan, 2019), we used six metrics to assess the ments identified in the two versions of the texts
readability of the original and simplified versions were identical. Of the remaining arguments,
of texts which include those that we use as test 5.31% (105) of those correctly identified in the
data for the SRL task. We found that the automat- original versions of the texts were not identified
ically simplified news texts have a lower proposi- in the simplified versions, while 8.29% (164) of
tional density (0.483 vs. 0.505) and reading grade the arguments correctly identified in the simpli-
level (5.4 vs. 10.3) and greater syntactic simplic- fied versions of the texts were not identified in the
ity (89.07 vs. 46.81) and temporal consistency, original versions. Of the 269 arguments identified
assessed in terms of tense and aspect (30.15 vs. in only one of the versions of the texts, 60.97%
27.76) than the original news texts. We deter- were arguments identified more accurately in the
mined the scores for these readability metrics us- simplified version, while 39.03% were arguments
ing the CPIDR tool (Covington, 2012)10 and the identified more accurately in the original versions
Coh-Matrix Web Tool (McNamara et al., 2014). of the texts.
As a task dependent on accurate syntactic pars- Table 4 shows the number of semantic roles la-
ing, we would expect that automatic SRL would belled more accurately, by type, when Senna pro-
be more accurate when processing the simplified cesses the original (Orig) and the automatically
versions of the input texts. simplified (Simp) versions of news articles. To
Evaluation Method. We applied Senna to the illustrate, when processing the original versions
original and automatically simplified versions of of the news texts, Senna correctly identifies the
10
http://ai1.ai.uga.edu/caspr/CPIDR-3. agents (arguments with semantic role label A0)
2.zip. Last accessed 31st May 2019. of 14 verbs that it did not identify when process-
289
Original sentence: But Smith had already been arrested - her clothing had been found
near his home and DNA tests linked him to it.
A0 V A1 A2 AMDIS AMLOC AMTMP
arrested Smith But already
Table 3: Example of more accurate semantic role labelling in automatically simplified text.
290
and any qualifying information about the finding Motivation. An analysis of the readability of the
(QUALIFIER). In this article, we focus on auto- original and simplified versions of the clinical vi-
matic extraction of information pertaining to phys- gnettes did not provide a strong indication that the
ical examinations. The goal of the IE system is to automatic sentence simplification method would
identify the phrases used in the clinical vignette improve the accuracy of the IE system. The 286
that denote findings and related concepts and add original clinical vignettes in the test data have a
them to its database entry for the vignette. mean propositional density of 0.4826 ideas per
Test Data. Our test data comprises a set of 286 word and 5.499 ideas per sentence. The values of
clinical vignettes and completed IE templates, en- these metrics for the simplified versions of the vi-
coding information about TECHNIQUEs, LOCA - gnettes are 0.4803 ideas per word and 5.269 ideas
TION s, SYSTEM s, and QUALIFIER s, associated per sentence, respectively. Although they are of
with the 719 FINDINGs that they contain. This the correct polarity, these differences are not sta-
test data was developed in the context of an earlier tistically significant (p = 0.5327 and p = 0.1407,
project and is based on clinical vignettes owned by respectively). However, previous work in sentence
the National Board of Medical Examiners.11 simplification for IE (Jonnalagadda et al., 2009;
NLP System. For the experiments described in Evans, 2011; Peng et al., 2012; Niklaus et al.,
this paper, we used a simple IE system in which 2016) has demonstrated that automatic sentence
input texts are tokenised and part of speech tagged, simplification can improve the accuracy of IE sys-
domain-specific gazetteers are used to identify ref- tems. This motivated us to evaluate the impact
erences to medical concepts and a simple set of fi- of the automatic sentence simplification method in
nite state transducers (FSTs) is used to group adja- this task.
cent references to concepts into multiword terms. Evaluation Method. For the IE task, our evalua-
The gazetteers and FSTs were developed in previ- tion metric is based on F1 -score averaged over all
ous work presented by Evans (2011). slots in the IE templates and all templates in the
After tagging references to clinical concepts in test data. Identification of true positives is based
the vignettes, IE is performed using a small num- on exact matching of system-identified slot fillers
ber of simple rules. To summarize, vignettes are with those in the manually completed IE templates
processed by considering each sentence in turn. in our test data.
Every mention of a clinical FINDING or SYMP - Results. The accuracy scores obtained by each
TOM is taken as the basis for a new IE template. variant of the IE system are presented in Table 5.
The first tagged TECHNIQUE, SYSTEM, and LO - Inspection of this table reveals that FINDINGs and
CATION within the sentence containing the focal all related concepts are identified more accurately
SYMPTOM or FINDING is considered to be related in the simplified versions of the input texts.
to it.12 Q UALIFIERS (e.g. bilateral or peripheral) Sentence (4) and its automatically simplified
are extracted in the same way, except in sentences variant (5) provide an example of the difference
containing the word no. In these cases, the QUAL - in performance obtained by the two systems. In
IFIER related to the FINDING is identified as none. these examples, identified FINDINGs are italicised
The sentences in the test data were simplified and associated concepts are underlined. Multi-
using the method presented in Section 3. We then word terms appear in square brackets.
ran the IE system in two settings. In the first
(4) She has truncalLOC obesity and
(IEORIG ), it processed the original collection of
pigmentedQU AL abdominalLOC striae.
vignettes. In the second (IESIM P ), it processed
the automatically simplified vignettes which con- (5) a. She has truncalLOC [obesity striae].
tain a reduced number of compound clauses. b. She has pigmentedQU AL
abdominalLOC striae.
11
https://www.nbme.org/. Last accessed 31st May
2019. In (5-a), the FINDING obesity is not tagged cor-
12
Versions of the system in which the closest tagged con-
cept was extracted in each case, rather than the first, were rectly because the SYMPTOM striae is erroneously
significantly less accurate in both cases (overall accuracy of grouped with obesity to form a new FINDING,
0.6542 for IE from the original vignettes, and 0.6567 for IE obesity striae which does not match the FIND -
from vignettes automatically simplified using the system de-
scribed in Section 3). See Table 5 for results obtained using ING listed in the gold standard. By contrast, LO -
the superior IE system. CATIONS in (5) are identified with greater accu-
291
IEORIG IESIM P
Template Best
slot Acc 95% CI Acc 95% CI Performer
F INDING 0.8819 [0.847, 0.914] 0.8861 [0.853, 0.917] 0.5486
T ECHNIQUE 0.8514 [0.814, 0.886] 0.8903 [0.858, 0.922] 0.9344
S YSTEM 0.8097 [0.769, 0.850] 0.8431 [0.806, 0.881] 0.873
Q UALIFIER 0.7431 [0.697, 0.786] 0.7708 [0.728, 0.814] 0.794
L OCATION 0.8431 [0.806, 0.881] 0.8611 [0.825, 0.894] 0.735
All 0.8258 [0.808, 0.843] 0.8503 [0.834, 0.867] 0.976
racy than those in (4) because IEORIG erroneously information extraction task. The probability of
extracts the same LOCATION (truncal) for both agreement with our gold standard is greater for
FINDING s in (4). IESIM P than for IEORIG , although the probabil-
We applied a bootstrapping method to obtain ity of agreement is already large for IEORIG . This
confidence intervals for accuracy of extraction of evaluation indicates that the automatic sentence
each of the IE template slots. For this purpose, simplification method facilitates IE.
50% of the the output of each system was ran-
domly sampled in each of 100 000 evaluations. 5 Conclusions
The confidence intervals are presented in the 95% As a result of various difficulties identified in cur-
CI columns of Table 5. The figures in the Best Per- rent approaches to intrinsic evaluation of sentence
former column of this table indicate the proportion simplification methods, we performed an extrinsic
of evaluations for which the IESIM P system was evaluation of one information-preserving sentence
more accurate than the IEORIG system. Differ- simplification method via three NLP tasks. We
ences in the accuracy of IE were found to be sta- found that the sentence simplification step brings
tistically significant in all cases, using McNemar’s improvements to the performance of IE and SRL
test (p < 0.00078), with the exception of differ- systems. In a third experiment, not described here
ences when extracting FINDINGs (p = 0.6766). due to space restrictions, we evaluated the sen-
Discussion. Chinchor (1992) notes that assess- tence simplification method extrinsically with re-
ment of the statistical significance of differences spect to a multidocument summarisation task us-
in accuracy between different IE systems is chal- ing MEAD (Radev et al., 2006) to summarise clus-
lenging. In our evaluation experiment, Dos San- ters of documents developed for Task 2 of DUC-
tos et al. (2018) framed the comparison between 2004.13 We found that the simplification step had
IEORIG and IESIM P using a binomial regression no impact on this task. As a result, although
model. Given that such models apply only when the findings reported in our current paper seem
the variables being considered are independent, promising, it is difficult to know the extent to
dos Santos et al. (2018) included a latent variable which they are applicable to other NLP tasks or
in the analysis to represent the effect of the text on to tasks which differ only with respect to the test
the performance of the two systems (the two eval- data used. This is one issue that we are interested
uations are not independent because both systems in exploring in future work. Another is a test of
process the same text). They showed that the odds whether extrinsic evaluation methods sensitive to
ratio of agreement between IESIM P and the gold information about the types of changes made in
standard is 1.5 times greater than that between the simplification step would perform better than
IEORIG and the gold standard. For all slots in the black box methods used in the current paper.
the IE template, the probability of agreement be- 13
Information about the DUC conferences is accessible
tween IEORIG and the gold standard is 0.937. The from https://www-nlpir.nist.gov/projects/
probability of agreement between IESIM P and the duc/index.html (last accessed 22nd August 2018).
gold standard is 0.957. This difference is statis- Guidelines about the tasks presented in DUC-2004 are
available at https://www-nlpir.nist.gov/
tically significant. They conclude that IEORIG projects/duc/guidelines/2004.html (last
and IESIM P differ in their performance on the accessed 22nd August 2018).
292
References David Dowty. 1991. Thematic proto-roles and argu-
ment selection. Language 67:547–619.
Mandya Angrosh, Tadashi Nomoto, and Advaith Sid-
dharthan. 2014. Lexico-syntactic text simplification Richard Evans. 2011. Comparing methods for the
and compression with typed dependencies. In Pro- syntactic simplification of sentences in information
ceedings of COLING 2014, the 25th International extraction. Literary and Linguistic Computing 26
Conference on Computational Linguistics: Techni- (4):371–388.
cal Papers. Association for Computational Linguis-
tics, Dublin, Ireland, pages 1996–2006. Richard Evans and Constantin Orăsan. 2019. Identify-
ing signs of syntactic complexity for rule-based sen-
Catherine Blake, Julia Kampov, Andreas K. Or- tence simplification. Natural Language Engineer-
phanides, David West, and Cory Lown. 2007. Unc- ing 25 (1):69–119.
ch at duc 2007: Query expansion, lexical simpli-
fication and sentence selection strategies for multi- Dan Feblowitz and David Kauchak. 2013. Sentence
document summarization. In Proceedings of the simplification as tree transduction. In Proceedings
Document Understanding Conference (DUC-2007). of the 2nd Workshop on Predicting and Improv-
National Institute of Standards and Technology. ing Text Readability for Target Reader Populations
(PITR). Association for Computational Linguistics,
Yvonne Canning. 2002. Syntactic Simplification of Sofia, Bulgaria, pages 1–10.
Text. Ph.d. thesis, University of Sunderland.
Robert Gaizauskas, Jonathan Foster, Yorick Wilks,
David Caplan and Gloria S. Waters. 1999. Verbal John Arundel, Paul Clough, and Scott Piao. 2001.
working memory and sentence comprehension. Be- The meter corpus: A corpus for analysing journal-
havioural and Brain Sciences 22:77–126. istic text reuse. In Proceedings of Corpus Linguis-
tics 2001 Conference. Lancaster University Centre
Raman Chandrasekar and Bangalore Srinivas. 1997. for Computer Corpus Research on Language, pages
Automatic induction of rules for text simplification. 214–223.
Knowledge-Based Systems 10:183–190.
Goran Glavas and Sanja Stajner. 2013. Event-centered
Nancy Chinchor. 1992. The statistical significance of simplication of news stories. In Proceedings of
the muc-4 results. In Proceedings of the Fourth Mes- the Student Workshop held in conjunctuion with
sage Understanding Conference. McLean, Virginia, RANLP-2013. RANLP, Hissar, Bulgaria, pages 71–
pages 30–50. 78.
Michael Collins. 1999. Head-Driven Statistical Mod- Ralph Grishman. 2005. The Oxford Handbook
els for Natural Language Parsing. Ph.d thesis, Uni- of Computational Linguistics. Oxford University
versity of Pennsylvania. Press, chapter Information Extraction, pages 545–
559.
Ronan Collobert, Jason Weston, Léon Bottou,
Michael Karlen, Koray Kavukcuoglu, and Eva Hasler, Adrià de Gispert, Felix Stahlberg, and Au-
Pavel Kuksa. 2011. Natural language pro- relien Waite. 2017. Source sentence simplification
cessing (almost) from scratch. Journal of for statistical machine translation. Computer Speech
Machine Learning Research 12:2493–2537. & language 45:221–235.
http://dl.acm.org/citation.cfm?id=1953048.2078186.
Deirdre Hogan. 2007. Coordinate noun phrase disam-
Michael A. Covington. 2012. CPIDR
R
5.1 user man- biguation in a generative parsing model. In Proceed-
ual. Technical report, Institute for Artificial In- ings of the 45th Annual Meeting of the Association of
telligence, University of Georgia, Athens, Georgia, Computational Linguistics. Association for Compu-
U.S.A. tational Linguistics, Prague, Czech Republic, pages
680–687.
Iustin Dornescu, Richard Evans, and Constantin
Orasan. 2013. A Tagging Approach to Identify Timothy B. Jay. 2003. The psychology of language.
Complex Constituents for Text Simplification. In Pearson, Upper Saddle Rive, NJ.
Proceedings of Recent Advances in Natural Lan-
guage Processing. Hissar, Bulgaria, pages 221 – Tomás̆ Jelı́nek. 2014. Improvements to dependency
229. parsing using automatic simplification of data. In
Proceedings of LREC-2014 the 22nd International
Larissa Sayuri Futino Castro dos Santos, Mar- Conference on Computational Linguistics (Coling
cos Oliveira Prates, Gisele de Oliveira Maia, Guil- 2008). European Language Resources Association,
herme Lucas Moreira Dias Almeida, Daysemara Reykjavik, Iceland, pages 73–77.
maria Cotta, Ricardo Cunha Pedroso, and Aurélio
de Aquino araújo. 2018. Assessing if an automated Siddhartha Jonnalagadda, Luis Tari, Jorg Hakenberg,
method for identifying features in texts is better than Chitta Baral, and Graciela Gonzalez. 2009. To-
another: discussions and results. Technical report, wards effective sentence simplification for auto-
Department of Statistics, Universidade Federal de matic processing of biomedical text. In Proceed-
Minas Gerais. https://bit.ly/2xUD2BI. ings of NAACL HLT 2009: Short Papers. Associ-
293
ation for Computational Linguistics, Boulder, Col- Carolina Scarton, Alessio Palmero Aprosio, Sara
orado, pages 177–180. Tonelli, Tamara Martı́n Wanton, and Lucia Specia.
2017. MUSST: A Multilingual Syntactic Simplifi-
Walter Kintsch and David M. Welsch. 1991. The cation Tool. In The Companion Volume of the IJC-
construction–integration model: A framework for NLP 2017 Proceedings: System Demonstrations.
studying memory for text. In W. E. Hockley and Taipei, Taiwan, pages 25–28.
S. Lewandowsky, editors, Relating theory and data:
Essays on human memory, Hillsdale, NJ: Erlbaum, Advaith Siddharthan, Ani Nenkova, and Kath-
pages 367–385. leen McKeown. 2004. Syntactic simplifica-
tion for improving content selection in multi-
Sigrid Klerke, Héctor Martı́nez Alonso, and Anders document summarization. In Proceedings of the
Søgaard. 2015. Looking hard: Eye tracking for de- 20th International Conference on Computational
tecting grammaticality of automatically compressed Linguistics. Association for Computational Lin-
sentences. In NODALIDA. Linköping University guistics, Stroudsburg, PA, USA, COLING ’04.
Electronic Press / ACL, pages 97–105. https://doi.org/10.3115/1220355.1220484.
Vladimir Iosifovich Levenshtein. 1966. Binary Codes Linnea B. Timm. 2018. Looking at text simplification
Capable of Correcting Deletions and Insertions and - Using eye tracking to evaluate the readability of
Reversals. Soviet Physics Doklady 10 (8):707–710. automatically simplified sentences. Bachelor the-
sis, Institutionen fr datavetenskap, Linkpings univer-
Danielle S. McNamara, Arthur C. Graesser, Philip M. sitet.
McCarthy, and Zhiqiang Cai. 2014. Automated
Evaluation of Text and Discourse with Coh-Metrix. David Vickrey and Daphne Koller. 2008. Sentence
Cambridge University Press. simplification for semantic role labeling. In Pro-
ceedings of ACL-08: HLT. Columbus, Ohio, pages
Christina Niklaus, Bernhard Bermeitinger, Siegfried 344–352.
Handschuh, and André Freitas. 2016. A sentence
simplification system for improving relation extrac- Tu Thanh Vu, Giang Binh Tran, and Son Bao Pham.
tion. In Proceedings of COLING 2016, the 26th In- 2014. Learning to Simplify Children Stories with
ternational Conference on Computational Linguis- Limited Data, Springer, Bangkook, Thailand, pages
tics: System Demonstrations. Association for Com- 31–41.
putational Linguistics, Osaka, Japan, pages 170–
174. https://www.aclweb.org/anthology/C16-2036. Sander Wubben, Antal van den Bosch, and Emiel
Krahmer. 2012. Sentence simplification by mono-
Constantin Orăsan, Richard Evans, and Ruslan Mitkov. lingual machine translation. In Proceedings of the
2018. Intelligent text processing to help readers with 50th Annual Meeting of the Association for Compu-
autism. In Khaled Shaalan, Aboul Ella Hassanien, tational Linguistics (ACL-12). Association for Com-
and Mohammed F. Tolba, editors, Intelligent Natu- putational Linguistics, Jeju, Republic of South Ko-
ral Language Processing: Trends and Applications, rea, pages 1015–1024.
Springer, pages 713–740.
Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze
Martha Palmer, Daniel Gildea, and Paul Chen, and Chris Callison-Burch. 2016. Optimizing
Kingsbury. 2005. The proposition bank: statistical machine translation for text simplification.
An annotated corpus of semantic roles. TACL 4:401–415. https://bit.ly/2Sj5mag.
Computational Linguistics 31(1):71–106.
Menahem Yeari, Paul van den Broek, and Marja
https://doi.org/10.1162/0891201053630264.
Oudega. 2015. Processing and memory of central
Kishore Papineni, Salim Roukos, Todd Ward, and Wei- versus peripheral information as a function of read-
Jing Zhu. 2002. BLEU: a method for automatic ing goals: evidence from eye-movements. Reading
evaluation of machine translation. In Proceedings and Writing 28 (8):1071–1097.
of the 40th annual meeting for Computational Lin-
guistics. Association for Computational Linguistics,
Philadelphia, Pennsylvania, pages 311–318.
294
OlloBot - Towards A Text-Based Arabic Health Conversational Agent:
Evaluation and Results
295
Proceedings of Recent Advances in Natural Language Processing, pages 295–303,
Varna, Bulgaria, Sep 2–4 2019.
guage barriers. The approach targets demograph- of Arabic text produced. Chatbots could be used
ics who are not bilingual and can speak only Ara- as language learning tools, to access information,
bic language. This is important, since from Table- to visualise the context of a corpus and to give an-
1 we see the burden of NCDs in Arabic speaking swers to specific questions (Abu Shawar, 2005). A
countries. We propose OlloBot, an Arabic conver- work by Shawar et al., (Abu Shawar, 2005) built a
sational agent able to converse with users and han- general language chatbot that supported different
dle daily tasks about diet, physical activity, men- language among which is Arabic and English. The
tal wellness and coping. The bot can also track study also used different corpora structure, such as
users daily food and keep a record of their daily dialogue, monologue and structured text to build
dietary habits. To our knowledge, few studies ex- the dialogue model.
ists on conversational agents that supports Ara- A work by Ali et al., (Ali and Habash, 2016)
bic language, able to flexibly converse and han- presented BOTTA, an Arabic dialect chatbot that
dle dialogues. However, no study exists that have communicates with users using the Egyptian Ara-
considered the health benefit achievable with lan- bic dialect. Another work by Shawar et al.,
guage specific conversational agents. Few plat- (Shawar and Atwell, 2009) described a technique
forms provide support to Arabic language technol- to access Arabic information using chatbot with
ogy to build interactive and intelligent conversa- natural language processing (NLP) tools. The bot
tional agents. This might be due to the complexity provides responses to capture the logical ontology
of the language and the resource scarcity available of a given domain using a set of pattern-template
to support it. Our approach relies on IBM Wat- matching rules. The literature shows no studies
son Conversation API (IBM bluemix)1 to handle considering conversational agent as assistive tool
the dialogue structure and Telegram Bot Platform for Arabic speakers. Which could provide instant
to build the chatbot. We tested OlloBot with 43 access to health information and data, and assist
Arabic speaking users and presented the findings physicians in providing a follow up to the patients.
in this paper. Conversational agents are great in handling repet-
itive tasks which consumes most of the health-
2 Background Research care providers time. Sundermeyer et al., (Sunder-
Although cognitive behavioral therapeutic (CBT) meyer et al., 2014) build a two translation recur-
apps have demonstrated efficacy, still they are rent neural models. The first one is a word-based
characterized by low adherence rate (Fitzpatrick approach using word alignments, while the second
et al., 2017). Conversational agents, on the other presents phrase-based translation models that are
hand, may offer a convenient and engaging alter- more consistent with phrase-based decoding. The
native of giving support at any time. A work by models are capable of improving strong baselines
Graf et al., (Graf et al., 2015) built a chatbot that in BOLT task for Arabic to English.
assimilates into daily routines. The bot communi-
3 OlloBot Architecture
cates with user and gathers nutrition data. Keep-
ing interaction with the bot final is a good design To provide the required level of Natural Language
practice, however building a great dialogue flow Understanding (NLU), we built the Arabic dia-
is also essential to ensure smooth user interaction. logue states on IBM Watson Conversation2 , which
This includes the way the bot handles several user defines the conversation flow and dialogue states.
interactions and requests. Zaghouani et al., (Za- The platform provides Artificial Intelligence sup-
ghouani et al., 2015) presented a correction an- port to catch different user intents and entities.
notation guidelines to create a manually corrected The OlloBot starts with the user sending a mes-
nonnative (L2) Arabic corpus. The work extends sage to the bot (OlloBot) running on Telegram
a large scale Arabic corpus and its manual cor- Bot Platform3 . The application provides the user
rections to include manually corrected non-native with the topics they can chat about, then based
Arabic learner essays. The approach uses anno- on user selection the bot forwards the request to
tated corpus to develop components for automatic IBM Watson Conversation cloud. The conversa-
detection and correction of language error to help tion slots takes user input and provides them with
standard Arabic learners and improve the quality 2
https://www.ibm.com/watson/services/conversation/
1 3
https://www.ibm.com/cloud-computing/bluemix/it https://telegram.org/blog/bot-revolution
296
relevant answer after checking their intent, entity The chatbot gathers users interaction data, and per-
and condition of the conversion. The logical parts formance indicators from their interaction. These
of the dialogue are handled by a Node.js wrap- data are saved in a database in the form of acces-
per to handle unmatched dialogues by the dialogue sible reports by the caregivers. OlloBot acts as a
flow ( Figure-1 illustrates the high-level architec- health assistants, meaning that it offers help and
tural view of OlloBot). support rather than treatment.
297
the task they were trying to accomplish, therefore terday, today and tomorrow. The “Quantity” en-
tracking the dialogue flow is essential to offer a tity refers to the quantity measurements the user
more flexible conversation to the users. Finally, to might mention in the conversation (i.e., kg, g, tea
handle fallbacks where the bot has no clue about spoon, bread loaf), whereas the “Numbers” en-
the respond, we designed the “anything else” in- tity refers to the countable number the user might
tent, to handle unhandled intents. mention. We have also defined a comprehensive
list of common Arabic food items and included
3.2 Health Report them in the “Food” entity. Finally, we defined
The chatbot application provides a health report a list of any kind of sports in the “Sport” entity.
generated by users interaction with the chatbot. The entity list was accompanied by a list of syn-
This report contains user activity data, their over- onym to add flexibility for the entity detection.
all interaction with the application and indications This is important to detect user intents and sen-
about their overall health. The health measure is timents from their conversation. While there has
mainly about their diet, exercise and stress pat- been a lot of research on sentiment analysis in En-
tern. All these data are structurally generated by glish, the amount of research and datasets for Ara-
the chatbot and saved into the database. These data bic language is still limited. A work by Alayba et
can then be accessed by the caregiver to obtain rel- al., (Alayba et al., 2017) built a sentiment analy-
evant user specific data. sis dataset in Arabic from Twitter data. The study
applied machine learning algorithms (i.e., Naive
3.3 Entities Bayes, Support Vector Machine and Logistic Re-
Entities are the pieces of valuable information hid- gression) for the analysis. Another work by Ismail
den in users input. They’re important keywords et al., (Ismail and Ahmad, 2004) proposed a new
extracted from a sentence. For example, in the ut- type of recurrent neural network architecture for
terance “I want to talk about physical activity”, the speech recognition and Backpropagation Through
word “physical activity” is detected as an entity, Time (BPTT) learning algorithm to observe differ-
and hence the bot switches to the physical activ- ences in alphabet “alif” until “ya”.
ity topic. Entities focus on defining the topic the
user is talking about. This is important to provide 3.4 Dialogue Engine
the right respond to user questions. We defined a
big range of entities to build our dialogue model. The dialogue structure was designed for each of
These entities represented the four main topics and healthy diet, physical activity, and mental wellness
the daily food logging. In Table-3 we provide the topics by referencing the Cognitive-behavioural
entities listed to structure the dialogue flow. therapy (CBT) (Rothbaum et al., 2000). We based
the chatbot dialogue on techniques suggested by
Topic Entities the CBT.
Topics Meal times The dialogue was then developed on IBM Watson
Daily times Weekly times Conversation, where we listed each of the intents
Quantity Food and entities and built the dialogue structure and
Numbers Sport flow. The tasks were intentionally built to be sim-
ple to interact with, so to decrease the time and
Table 3: The Topic Entities amount of physical and mental efforts needed, and
increase user’s ability. For example, when asked
The above table lists the “Topics” entity, namely to log their food, users can either write the list or
diet, physical activity, mood, and coping that the send an emoji of the food item. Moreover, but-
bot converses about. The topics entity also covers ton replies were provided to further simplify the
users daily food logging. Other entities include data logging. Simplifying food tracking process
the “Meal times” which defines the meal periods will increase users ability and decrease the learn-
per day that includes: breakfast, lunch, snacks ing curve associated to health tracking. This is be-
and dinner. The “Daily times” refers to the peri- cause interaction with the bot shouldn’t be only
ods of the day, namely morning, noon, afternoon, conversational, since some interactions are better
evening and night. Whereas, “Weekly times” en- with Graphical UI and others with Conversational
tity includes the period of the day, namely yes- UI.
298
3.4.1 Healthy Diet ple, X2AI4 created a set of chatbots for mental
Around 2 billion people are overweight, but many health applications. Their flagship AI, Tess, helps
are ready to change (Van Itallie, 1985). However, patients in tandem with their doctors by providing
according to studies (Mann et al., 2007) tempo- resources on cognitive-behavioral therapy, medi-
rary fixes to old habits makes people to regain their cation side effects, and questionnaire automation.
weight. OlloBot acts as an interactive AI-powered Woebot5 is a mood tracking chatbot with person-
diet tracking bot that converses with individuals in ality. Backed by scientific research, Woebot can
a friendly way directly through the Telegram mes- help reduce depression, share CBT resources, and
saging application. The goal is to give instant ad- learn from conversations over time. Finally, Joy6
vices with each meal eaten and help improve the uses a chatbot approach to mental health. She of-
eating habits on the go. Once the conversation is fers options for both individuals and therapists.
executed, OlloBot starts conversing about user’s 3.4.4 Coping Skills
diet and asks them questions about their eating
Being mentally or emotionally healthy is more
habits and highlights the values associated with
than being free of depression, anxiety, or other
healthy diet. For example, OlloBot stresses the
psychological issues. Rather than the absence of
fact that following a diet rich in vegetables and
mental illness, mental health refers to the presence
fruits helps decrease escalation into overweight,
of positive characteristics. With OlloBot, we han-
obesity or even chronic conditions, and the detri-
dle this by providing coping skills and mindful-
mental effects associated otherwise.
ness support through clever motivational quotes,
3.4.2 Physical Activity relevant workout suggestions mixed with some
health facts to help users understand the benefit
Studies (Cooney et al., 2014; Mead et al., 2009; of health and wellness. The bot provides the uses
Artal et al., 1998; Byrne and Byrne, 1993) have with tips about coping with stress, diet, exercise,
shown that exercise can treat mild to moderate sleep, and mindfulness. The quotes and sugges-
depression as effectively as antidepressant medi- tions are all based on best recommendations pro-
cation and with no associated side-effects. Our vided by the World Health Organisation (WHO)
conversational agent chats with users about their (Michie et al., 2009). Coping skills are important,
physical activity and daily energy level, and can since finding a moment to take a few deep breaths
provide personalized plans and keep track of and quiet your mind is a great way to relieve stress
workout progress by storing relevant user inputs. and improve your overall health.
Whether the user wants to stay fit, loose weight,
or get toned, OlloBot can later provide an effi- 3.4.5 Daily Food Logging
cient and consistent workout plan while keeping While many food-logging tools are already on the
track of the progress. Although the bot is not market, most are either too complicated or bor-
considered as a tool to loose weight through ex- ing for the average individual. OlloBot keeps a
ercise, but rather a supportive tool to help individ- food diary, tracks calories, and provides basic nu-
uals track their exercise and improve in on the go, tritional tips based on user’s eating habits. We
making users conscious about their health habits wanted to make the food logging process sim-
also makes them more likely to set aside time for pler, more engaging, and more informative for the
physical activities in the future. users. The bot tracks user meals, by asking them
to insert each meal items. This helps to know more
3.4.3 Mental Wellness about users diet. With the NLP capabilities of-
Mental health refers to our overall psychological fered by IBM Watson Conversation, the bot sup-
well-being. This includes the way we feel about ports any food combinations. The user can even
ourselves, the quality of our relationships and the log their food by sending emojis of the food items.
ability to manage our feeling. OlloBot chats with Once the food logging is done, the bot asks the
users about mental health and asks them questions user about their preferred time to recheck with
about their stress, sleep and other measures rele- them. OlloBot is in its early stages, due to the time
vant to their mental wellbeing. There exists sev- 4
https://x2.ai/
eral mental health chatbots that provide support 5
https://woebot.io/
at different stages of mental illness. For exam- 6
http://www.hellojoy.ai/
299
Figure 2: Daily Food Logging Conversation with OlloBot.
required to build and integrate new features, then pects of the bot. The questionnaire is based on a
validate it through testing with real users. We un- standard framework, namely USE Questionnaire
derstand that existing food logging apps are now (Lund, 2001) to measure the usability of the chat-
far ahead of us in terms of calorie tracking, pre- bot. The framework tests four items of usability
cision, and various interface features. However, within a product. It consists of 30 questions to
we aim to close this gap and simplify the food test the usefulness, easy of use, ease of learning
tracking/logging and feedback providing process and satisfaction with the application. The ques-
in the future (see Figure-2 for the daily food log- tionnaires are organised in a scale of 1 - 5, where
ging with OlloBot). 1 is “strongly disagree” and 5 is “strongly agree”.
We describe the experiment settings and results in
4 OlloBot Platform the following sections.
The dialogue flow in OlloBot were handled with 5.1 Participants Demographics
IBM Watson Conversation, whereas the logical
The user demographics consisted of both male
part of the dialogue were handled with a Node.js
(n=26) and female (n=17) participants with an age
implementation. Our choice of IBM Watson was
range of 20 - 65 years old for both gender. We
due to its language support, to our knowledge it
checked the participants familiarity with tracking
was the only conversational agent dialogue build-
devices (e.g., any diet, sleep, physical activity, or
ing platform that supports Arabic. We deployed
mood tracking application) and chatbot applica-
the bot on Telegram Bot Platform and built the
tions (see Table-4 for the results). There was no
components to handle user requests and responses.
significant differences in both cases with either
We used various UI elements made available by
tracking devices (e.g., wearable, sensor, or mobile
Telegram and integrated them into our dialogue
applications) or chatbot applications. The major-
model.
ity of participants have shown no familiarity with
5 OlloBot - USE Questionnaire tracking devices (n=26) and chatbot applications
(n=27). We provided additional questions at the
After building and integrating OlloBot with Tele- end of the survey to evaluate their overall experi-
gram Bot Platform, we conducted a user exper- ence with the chatbot. A reimbursement of 5ewas
iment with 43 Arabic speaking users. The dia- given to all participants in the form of Amazon
logue and various questions the bot asks were all coupon.
based on a WHO questionnaire we extracted about
health and wellbeing (Kessler and Üstün, 2004; 5.2 Experiment
Parslow and Jorm, 2000; , WHO et al.(2017; Or- We distributed the chatbot to all the Arabic speak-
ganization et al., 2006). After testing the chatbot, ing participants recruited from Iraq. After carrying
we performed a survey analysis to test various as- out the experiment for 1 week, we collected data
300
Female 17 to remember each time. For example, the question
Gender
Male 26 “I quickly became skilful with it” checks for how
Mean 29.8 quickly a skill is learned using the chatbot.
Age
Std. Dev 9.28
No 26 5.2.4 Satisfaction
Tracking devices familiar
Yes 17 This step checks for user’s overall satisfaction
No 27 with the chatbot. It checked whether the users are
Chatbot familiar
Yes 16 satisfied and would recommend the tool to others
and how entertaining they perceived it. For exam-
Table 4: Participants Demographics. ple, the question “I would recommend it to a friend
or family member” checks whether the user would
through questionnaire to test the four main points recommend the bot to their relatives or friends.
relevant to the usability of OlloBot. We performed descriptive statistics on the results
and reported them in Table-5. One Sample t test
5.2.1 Usefulness showed that averages for all four scales were sta-
This item helps to test how useful the participants tistically significantly different from the middle
perceived the application. The usefulness includes value (value = 3, see Table-5). Further analy-
measuring whether the bot helps the user to be ses were performed considering gender, familiar-
more effective and productive and enhance their ity with tracking technology, and familiarity with
control over their daily life activities. For exam- chatbot as between factors. No differences among
ple, “I believe is effective to track my health” eval- the four usability dimensions were observed be-
uates whether the user thinks the bot is effective in tween male and female participants, and no dif-
tracking their health. ferences emerged between participants who fre-
quently use or not use tracking devices. A signifi-
Dimensions Mean Std. t value p value
cant difference for the “ease of learning” scale was
Dev. (df=42)
observed when comparing participants who were
Usefulness 3.381 0.373 6.695 p <.01
familiar in interacting with chatbot application and
Ease Of Use 3.626 0.3944 10.405 p <.01 participants who were not. The latter group re-
Learning 3.942 0.5341 11.564 p <.01 ported lower scores for learnability compared to
Satisfaction 3.505 0.3667 9.031 p <.01 the former group (t(41)= -2.46, p <.01).
Table 5: The Results from the USE Questionnaire. 5.2.5 Overall Experience
This part evaluated the overall experience with Ol-
5.2.2 Ease of Use loBot and is not part of the USE questionnaire. We
This point helps understand the ease of use aspect. tested user satisfaction with the reliability of Ol-
It measures easiness, simplicity and user friendli- loBot, their overall experience with the bot from
ness of the application (chatbot). This point also different perspectives (see Table-7 for the expe-
measures the steps and effort required to achieve rience comparison). We have checked whether
the goal set and whether its easy to recover from users like chatbots or rather use a mobile appli-
mistakes. For example, to measure the effort re- cation. Finally, we considered the list of most
quired for each step, we asked users to provide positive and negative aspects the users mentioned
their scale for “It requires the fewest steps possible during the survey. In addition, we asked the users
to accomplish what I want to do with it”. We per- about the features they would like to see/use in fea-
formed a descriptive statistics about the four items ture versions of OlloBot.
checked in the overall experience (see Table-6).
Figure-3 below lists each of the overall bot relia- 6 Discussion
bility (Figure-3a), and the participants overall ex-
perience (Figure-3b, Figure-3c and Figure-3d). This work was the first to evaluate the application
of NLP powered health tools into language and
5.2.3 Ease of Learning context scarce domains. The work involved ini-
This point measures the learnability of the tool. tial design phase, which involved researchers and
We measure whether its quick to learn and easy health experts in the context of healthy lifestyle
301
(a) Overall Reliability Satisfac-(b) Participants Overall experi-(c) Participants Overall experi-(d) Participants Overall experi-
tion. ence. ence. ence.
App vs. Chatbot Not at all satisfied vs. Extremely satisfied Terrible vs. Wonderful Frustrating vs. Satisfying Dull vs. Stimulating
Valid 43 43 43 43 43
Mean 3.860 3.163 3.767 3.581 3.512
Std. Dev. 0.9900 0.8710 0.7508 0.6980 0.70028
Min 1.000 1.000 2.000 2.000 2.000
Max 5.000 5.000 5.000 5.000 5.000
Overall Experience
Questions Scale = 1 Scale = 5
How satisfied are you with the reliability of this chatbot? Not at all satisfied Extremely satisfied
Can you rate your experience with the chatbot: Terrible Wonderful
Can you rate your experience with the chatbot: Frustrating Satisfying
Can you rate your experience with the chatbot: Dull Stimulating
Would you rather use the chatbot or prefer a mobile app to Use an app Use a chatbot
track your everyday health ?
302
References Gillian E Mead, Wendy Morley, Paul Campbell, Car-
olyn A Greig, Marion McMurdo, Debbie A Lawlor,
Bayan Aref Abu Shawar. 2005. A corpus based ap- et al. 2009. Exercise for depression. Cochrane
proach to generalise a chatbot system. Ph.D. thesis, Database Syst Rev 3.
University of Leeds.
Susan Michie, Charles Abraham, Craig Whittington,
Abdulaziz M Alayba, Vasile Palade, Matthew Eng-
John McAteer, and Sunjai Gupta. 2009. Effective
land, and Rahat Iqbal. 2017. Arabic language sen-
techniques in healthy eating and physical activity in-
timent analysis on health services. arXiv preprint
terventions: a meta-regression. Health Psychology
arXiv:1702.03197 .
28(6):690.
Dana Abu Ali and Nizar Habash. 2016. Botta: An
arabic dialect chatbot. In COLING (Demos). pages World Health Organization et al. 2006. Global strategy
208–212. on diet, physical activity and health: a framework to
monitor and evaluate implementation .
Michal Artal, Carl Sherman, and Nicholas A DiNubile.
1998. Exercise against depression. The Physician World Health Organization et al. 2010. WORLD
and sportsmedicine 26(10):55–70. HEALH REPORT (The): Health Systems Financ-
ing: the path to universal Coverage (Arabic). World
A Byrne and DG Byrne. 1993. The effect of Health Organization.
exercise on depression, anxiety and other mood
states: a review. Journal of psychosomatic research World Health Organization et al. 2012. World health
37(6):565–574. day 2012: ageing and health: toolkit for event orga-
nizers .
Gary Cooney, Kerry Dwan, and Gillian Mead. 2014.
Exercise for depression. Jama 311(23):2432–2433. Ruth A Parslow and Anthony F Jorm. 2000. Who uses
mental health services in australia? an analysis of
Kathleen Kara Fitzpatrick, Alison Darcy, and Molly data from the national survey of mental health and
Vierhile. 2017. Delivering cognitive behavior ther- wellbeing. Australian and New Zealand Journal of
apy to young adults with symptoms of depres- Psychiatry 34(6):997–1008.
sion and anxiety using a fully automated conver-
sational agent (woebot): A randomized controlled Barbara Olasov Rothbaum, Elizabeth A Meadows, Pa-
trial. JMIR Mental Health 4(2):e19. tricia Resick, and David W Foy. 2000. Cognitive-
behavioral therapy. .
Bettina Graf, Maike Krüger, Felix Müller, Alexander
Ruhland, and Andrea Zech. 2015. Nombot: sim- B Abu Shawar and Eric Atwell. 2009. Arabic question-
plify food tracking. In Proceedings of the 14th In- answering via instance based learning from an faq
ternational Conference on Mobile and Ubiquitous corpus. In Proceedings of the CL 2009 International
Multimedia. ACM, pages 360–363. Conference on Corpus Linguistics. UCREL. volume
386.
Eat Smart Play Hard, Healthy Kansas Schools, In-Class
Physical Activity Breaks, and Blast Off Game. 2012. Martin Sundermeyer, Tamer Alkhouli, Joern Wuebker,
Physical activity. Fitness and health: Internatio . and Hermann Ney. 2014. Translation modeling with
bidirectional recurrent neural networks. In EMNLP.
Saliza Ismail and A Ahmad. 2004. Recurrent neural
pages 14–25.
network with back propagation through time algo-
rithm for arabic recognition. Proceedings of the Theodore B Van Itallie. 1985. Health implications of
18th ESM Magdeburg, Germany pages 13–16. overweight and obesity in the united states. Annals
Ronald C Kessler and T Bedirhan Üstün. 2004. The of internal medicine 103(6 Part 2):983–988.
world mental health (wmh) survey initiative version World Health Organization (WHO et al. 2017. Obesity
of the world health organization (who) composite in- and overweight factsheet from the who. Health .
ternational diagnostic interview (cidi). International
journal of methods in psychiatric research 13(2):93– Wajdi Zaghouani, Nizar Habash, Houda Bouamor, Alla
121. Rozovskaya, Behrang Mohit, Abeer Heider, and Ke-
mal Oflazer. 2015. Correction annotation for non-
Hian Chye Koh, Gerald Tan, et al. 2011. Data mining
native arabic texts: Guidelines and corpus. In
applications in healthcare. Journal of healthcare in-
LAW@ NAACL-HLT. pages 129–139.
formation management 19(2):65.
Arnold M Lund. 2001. Measuring usability with the
use questionnaire12. Usability interface 8(2):3–6.
Traci Mann, A Janet Tomiyama, Erika Westling, Ann-
Marie Lew, Barbra Samuels, and Jason Chatman.
2007. Medicare’s search for effective obesity treat-
ments: diets are not the answer. American Psychol-
ogist 62(3):220.
303
Developing the Old Tibetan Treebank
304
Proceedings of Recent Advances in Natural Language Processing, pages 304–312,
Varna, Bulgaria, Sep 2–4 2019.
and Old Tibetan Chronicles, since these are language like Old Tibetan.
the best sources that we have at our disposal
in terms of length and linguistic variety (many 3 Pre-Processing
other Old Tibetan text are short inscriptions
The Old Tibetan texts we work with to start
or more fragmentary). The Old Tibetan An-
this corpus are already transcribed from the
nals are Tibet’s earliest extant history. The
original manuscripts or digitised images. For
Old Tibetan Chronicle, written in the early
the present paper we thus only address the is-
9th century, is more narrative and includes
sues concerning encoding of transcriptions and
historical accounts and songs related to the
transliterations and the issues of tokenising
Yarlung dynasty and the Tibetan empire.
a language without word or sentence bound-
In this paper we present our annota- aries.
tion procedure that addresses all issues of
pre-processing, segmentation, POS tagging 3.1 Transliteration Issues
and parsing in detail. Our semi-supervised
One of the first challenges we encountered
method, resulting in the first Old Tibetan
in creating the Old Tibetan corpus was the
Treebank, can furthermore serve as an ex-
conversion from Tibetan Unicode script (see
ample of how to overcome challenges of low-
Hill 2012) to the Wylie transliteration system.
resource and under-researched languages in
There are few reliable tools available and in
general.
addition, we have to take the peculiar ortho-
graphic features of Old Tibetan into consid-
2 The Annotation Procedure
eration. The Tibetan Unicode script for the
Since Old Tibetan is an extremely under- Old Tibetan documents was obtained from a
resourced language the procedure to develop modified version of the Wylie transliteration
an annotated corpus needs to be developed system that is used for the Old Tibetan Doc-
with great care. Additional steps are neces- uments Online (OTDO) website, through the
sary at each of the normal stages, from pre- BDRC conversion tool.2
processing to POS tagging and parsing and However, this tool only partially addresses
finally post-processing. In the pre-processing the issue, because we also want to transform
stage, for example, the normalisation is not a Old Tibetan into a form of Tibetan that looks
trivial task because of a range of issues with more similar to Classical Tibetan in terms of
the Tibetan script and the way it is digitised, orthography. Therefore, the Wylie transliter-
in either Unicode or a variety of transliter- ation used by the OTDO website had to be
ation formats. In addition to solving these modified. As an example the reverse ‘i’ vowel
script issues, our core solution is to trans- mark, ྀ - called gigu - is transliterated with ‘I’
form our Old Tibetan texts through a ‘conver- on the OTDO website. We substituted ‘I’ with
sion/normalisation process’ with a Constraint ‘i’, which is the standard Wylie transliteration
Grammar (Cg3) to a form of Tibetan that is for this character, as shown in (1):
closer to Classical Tibetan, for which at least
(1) rgyal po’I > rgyal po’i ‘of the king’
some NLP tools are available.
Before we can move on to the annotation 3.2 Normalisation
stage, we need to solve a further non-trivial is-
The Old Tibetan script furthermore presents
sue of finding word and sentence boundaries.
a set of features that need to be ‘normalised’
Since there are no Gold Standards or train-
or converted to a form that looks like Classical
ing data available for Old Tibetan, we resort
Tibetan. We therefore created a set of rules
to the little material and tools available for
translated into the Constraint Grammar
Classical Tibetan and then do a rigorous error
(Cg3) formalism. Most of the Cg3 rules
analysis checking specific Old Tibetan features
used to normalise Old Tibetan are simple
that we know differ from Classical Tibetan.
replacement rules. For example, In Old
Our annotation method is thus supervised in
Tibetan there are many instances of the
various ways to overcome the obstacles build-
ing a treebank of an extremely low-resourced 2
https://www.tbrc.org/
305
above-mentioned reverse gigu such as ཀྱྀ� kyI. ([\\u0F7C\\u0F7A\\u0F74\\u0F72\\u0F80]
These two forms of gigu, ི and ྀ are phonetically �?))>"r)(NOT 0 (split) or (genitive)
indistinguishable and mark no difference in
Classical Tibetan. The Cg3 SUBSTITUTE rule or (diphthongs));
to normalise the reverse gigu is: Through these conversions and normalisa-
tions, we could apply existing tools for Classi-
SUBSTITUTE (``([[\^{ }<]*)\u0F80(.*)"r) cal Tibetan to our Old Tibetan corpus to avoid
(“$1ི$2v) TARGET (σ) manually creating our treebank from scratch
completely. The full Cg3 grammar is discussed
Two additional problems encountered in the in detail in our forthcoming research.
normalisation of Old Tibetan are represented
3.3 Segmenting Sentences
by the alternation between aspirated and
unaspirated voiceless consonants and the dif- Segmenting sentences is necessary since there
ficulty of splitting merged syllables. This as- are no obvious sentence boundaries in Old Ti-
piration, however, was probably not phonemic betan. The Tibetan scripts does have a punc-
in Old Tibetan (Hill, 2007, 471). Therefore, a tuation marker that sometimes (but not al-
set of string replacement rules in the Cg3 for- ways) indicates meaningful phrases, a so-called
malism was created to normalise and convert shad, ། or double shad, །།. Since without any
these instances to their equivalent reading in further annotation, there is no way of know-
Classical Tibetan. ing where sentences begin or end, we used the
Furthermore in Classical Tibetan, syllables single and double shad as sentence boundaries
are separated by a punctuation marker called and automatically inserted utterance bound-
tsheg: ་. In Old Tibetan texts, syllable margins aries indicators (<utt>) after every instance.
are not so clear and syllables are often merged This greatly facilitates subsequent annotation
together with the following case marker or con- tasks that depend on sentence boundaries,
verb, e.g. Old Tibetan བཀུམོ bkumo > Clas. Tib. such as POS tagging and chunkparsing.
བཀུམ་མོ bkum mo ‘kill, destroy’:
3.4 Tokenisation
(2) བཀུམོ > བཀུམ་མོ The Tibetan script furthermore does not indi-
cate word boundaries. Tokenisation is there-
These types of merged syllables were also
fore a tremendous issue, not only for scholars
converted to their classical forms, using a
of Tibetan (who often disagree on what the
set of three regular expressions in the Cg3
word boundaries should be), but even more so
formalism through the rule SPLITCOHORT.
for any Tibetan NLP tasks. The Classical Ti-
Considering the complexity of the Tibetan
betan script does have a way of indicating syl-
syllable, in order to generate the rules, we
lable boundaries though, by using the above-
took the maximum number of its constituents
mentioned tsheg marker ་ , e.g. བྲག་མར transliter-
into account (in terms of vowels and conso-
ated brag mar ‘Dagmar’ with spaces between
nants) as well as their order.
every syllable according to the conventions of
the Wylie transliteration.
Generic Rule:
For Classical Tibetan, Meelen and Hill
([^aeiouI\s]+[aeiouI][^aeiouI\s]*) (2017) addressed this tokenisation issue by
([^aeiouI\s'])([aeiouI][^aeiouI\s']*) recasting it as a classification task with a
> $1$2 $2$3 memory-based tagger (Daelemans et al., 2003)
giving ‘beginning’, ‘middle’ or ‘end’ labels to
every syllable (automatically split based on
Cg3 rule: the aforementioned tsheg and shad markers.
SPLITCOHORT ( With our supervised learning method first nor-
"<$1>"v "$1$3�"v malising and then converting our Old Tibetan
"<$3$4>"v "$3$4"v corpus to a form of Tibetan that is much
)("<(.{2,6})(([^\\u0FB2\\u0FB1]) closer to Classical Tibetan, we were able to
306
use this existing segmentation tool for Classi- memory-based tagger4 on the >318k Clas-
cal Tibetan and extend and modify them after sical Tibetan Gold Standard, yielded better
manually correcting part of our Old Tibetan results compared to those of the large tag set
data. reported by (Meelen and Hill, 2017) (increase
from 95.0% to 96.3% in Global Accuracy;
4 POS Tagging Known Words increased from 96.8% to 97.8%;
Unknown Words from 53.4% to 59.7%).
Since there was no Old Tibetan POS-tagged All tags with a very low number of tokens in
Gold Standard either, here too we started from the out-of-vocabulary set (ranging from n = 1-
the Classical Tibetan training data3 and tag- 92) have a Precision and Recall close or equal
ging method developed by Meelen and Hill to zero. These items are always very short (one
(2017). We tested a number of ways to get or two characters only), which makes predict-
and improve results for the Old Tibetan cor- ing the tag for new items in this category an al-
pus, e.g. developing a new, reduced tag set, most impossible task for the tagger. With the
changing scripts (Unicode vs. Wylie) as well newly trained small tag set tagger, we tagged
as generating new taggers, based on the man- the Old Tibetan Annals and manually cor-
ually corrected Old Tibetan only and, finally, rected the first 3.5k tokens as a start. We then
adding the manually corrected Old Tibetan to evaluated the tagger again with another 10-
the existing Classical Tibetan Gold Standard. fold cross-validation, first on this small Old Ti-
betan corpus and then again adding this man-
4.1 Small vs Large Tag Set ually corrected Old Tibetan data to the ex-
The tag set used for the Classical Tibetan isting Classical Tibetan Gold Standard. This
Gold Standard, developed by Garrett et al. yielded a better Global Accuracy for the com-
(2014) is with 79 morpho-syntactic tags bination of Old and Classical Tibetan (96.1%)
rather large. This causes major issues for compared to Old Tibetan alone (92.8%). How-
the out-of-vocabulary items, especially for ever, the results for Unknown Words are signif-
languages without insightful morphological icantly lower (decrease from 71.1% to 58.5%).
suffixes like Tibetan. For this first attempt Since these two new Gold Standards differ
of developing an Old Tibetan Treebank, we significantly in size it is impossible to do a fair
therefore decided to reduce the amount of comparison until we manually correct more
tags to a small and simplified version of Old Tibetan. It is clear, however, that despite
the standard Universal Dependency POS our efforts to normalise and convert the Old
set, consisting of 15 tags only (De Marneffe Tibetan into a form of the language that looks
et al., 2014). We transformed the existing more like Classical Tibetan, it is still making
Classical Tibetan training data, which is a difference, shown in the lower accuracy (by
our Gold Standard, in the following way: more than 10%) of unknown words for this
interj > INTJ, punc > PUNCT, n.prop combined training data. Without adding the
> PROPN, skt, dunno > X, adj, num.ord Classical Tibetan training data, however, the
> ADJ, n.v.cop, v.cop, v.cop.neg > vocabulary list that the memory-based tagger
AUX, n.count, n.mass, n.rel > NOUN, builds would simply be too small to get any
num.card, numeral > NUM, cl.focus, good results on unseen data. Despite the 10-
cv.fin, cv.imp, cv.ques, neg > PART, fold cross-validation, the relatively high scores
p.indef, p.interrog, p.pers, p.refl > for the Old Tibetan corpus only are mislead-
PRON, d.dem, d.det, d.emph, d.indef, ing, because of the small size of the corpus.
d.plural, d.tsam > DET, and, finally, all Until we have more manually corrected Old
verb remaining verb forms in all tenses > Tibetan data, we therefore proceed with the
VERB, all remaining converbs > SCONJ, all Classical Tibetan Gold Standard and add an
post-positional case markers > ADP and all extra stage of error correction, see Section 6.
adverbs > ADV. A 10-fold cross-validation 4
These settings for Classical Tibetan are:
with the exact same parameter settings of the -p dwdwfWaw -P psssdwdwdwFawaw -M 1100 -n 5
-% 8 -O+vS -FColumns -G K: -a0 -k1 U: -a0 -mM
3
http://github.com/tibetan-nlp/soas-corpus/ -k17 -dIL.
307
4.2 Unicode vs Wylie Transliteration 4.3 Memory-Based vs Neural-Network
Tagging
The above-mentioned taggers were trained
and tested on Tibetan script in Unicode. The Finally, we tested a BiLSTM-CNN-CRF tag-
Unicode Tibetan script contains a lot of so- ger5 to see if it would yield better results
called ‘stacked’ characters that are centred than the memory-based tagger. We chose this
before, above and below one single root let- neural-network tagger, because it processes
ter. A typical example is བ བས་ , which is both word- and character-level representations
transliterated in the official Wylie system as automatically, using a combination of a bidi-
bsgrubs ‘achieved’. In Tibetan Unicode, the rectional Long-Short-Term-Memory (LSTM),
order of these stacked characters can differ a Convolutional Neural-Network (CNN) and a
depending on the exact combinations of con- Conditional Random Field (CRF). Although
sonants and vowels. This varying order of- this tagger requires no pre-processing of the
ten yields unexpected problems when process- data or any further feature engineering, the
ing Tibetan Unicode text as our NLP algo- results are better when the system can use
rithms do not recognise variants of the same word vectors for the specific language. Since
order in the same word as the same type. the current number of manually corrected to-
This then increases the number of types and kens in Old Tibetan is too small to train any
thus reduces the overall accuracy. For this neural-network-based tool, we again resorted
reason, we converted the Classical Tibetan to using Classical Tibetan instead. For Classi-
Gold Standard from Tibetan Unicode script cal Tibetan, we used FastText6 to create word
to Wylie transcription as well. Some exam- embeddings with the aim of improving the re-
ples of Tibetan Unicode with Wylie transliter- sults of the tagger with word vectors based
ations are: བཅོམ་ལྡན་འདས་ bcom-ldan-’das ‘Blessed on a large amount of Tibetan data digitised
One’, ཤཱཀྱ་སེང་གེ་ shAkya-seng-ge ‘Buddha’, ཕྱག་ by the BDRC7 and annotated by Meelen and
phyag ‘arm, prostration’. Hill 2017: the Annotated Corpus of Classical
In a 10-fold cross-validation of the Classi- Tibetan (ACTib) (version 1, (Meelen et al.,
cal Tibetan Gold Standard, this conversion 2017)). We then divided the above-mentioned
to Wylie yields slightly better results. Global >318k token Classical Tibetan Gold Stan-
Accuracy was 95.0% for Tibetan Unicode vs. dard in training, test and developments sets
96.5% for Wylie. We observed a major im- (80/10/10), trained a tagger with these word
provement in Unknown Words in particular embeddings and evaluated the results on the
from 53.4% in the Tibetan Unicode to 62.2% held-out test set. With its default settings,8
in the Wylie transliteration. Since the results this BiLSTM-CNN-CRF tagger yielded a re-
with the Wylie transliteration are slightly bet- sult of 95.8% Global Accuracy (F1 score).9
ter, especially for unknown, out-of-vocabulary These results are slightly better than those
items, converting all Unicode Tibetan to Wylie of the memory-based tagger (95% Global Ac-
transliteration would appear to be a logical curacy). They are reasonable, but could be
way forward. However, in practice, Unicode improved in a number of ways. Furthermore,
Tibetan script is far more widely used within at present they cannot easily be reproduced
the Tibetan community. To make the cor- for our small corpus of the Old Tibetan An-
pus more accessible, but also to get support nals written in a very different style and genre.
from members of this community who are will- 5
See https://github.com/achernodub/targer and
ing to correct segmentation and any further
Chernodub et al. (2019).
type of linguistic annotation, a Unicode Ti- 6
https://fasttext.cc/
betan version is indispensable. It is there- 7
https://www.tbrc.org/
8
fore important to develop segmenters, taggers Batch size = 10; 100 epochs; dropout ration = 0.5
with the Bi-RNN-CNN-CRF model.
and parsers that work well for both, or de- 9
Although it is not common practice (anymore) for
velop tools that can automatically convert the POS tagging evaluations, we calculated the F1 instead
Tibetan text (but not any type of annota- of normal accuracy to make it directly comparable to
the results presented by (Meelen and Hill, 2017). Ac-
tion also in roman script) back from its Wylie tual accuracies are slightly higher than the Global Ac-
transliteration to Unicode Tibetan script. curacies presented here.
308
These initial neural-network results thus look extremely high (98%, 100% and 99% respec-
promising, but need further extension and re- tively for n=5627 in the Wylie translitera-
finement. In forthcoming work we address tion evaluation of Classical Tibetan discussed
these issues by optimising the parameters, im- above). The advantage of keeping the agen-
proving the segmentation and, with that, cre- tive case marker tag is that for many transi-
ating better word embeddings (Hill et al., tive sentences at least, we will be able to au-
ming). tomatically detect the subject of the clause.
Since Old Tibetan was a pro-drop language
4.4 Summary of POS Tagging (i.e. pronouns need not necessarily be overtly
The below table summarises the results of our expressed, see Tournadre 2010, 101), it is not
tests and evaluations discussed in the previ- always possible to detect non-marked subjects
ous sections. There are some differences be- of verbs automatically, so a certain amount
tween the small and the larger tag sets and of manual correction is still always necessary.
between the Unicode Tibetan script and the Similarly, keeping the genitive case markers
Wylie transliteration, with the smaller tag sets (ཀྱི་ གྱི་ འི་/case.gen, see Tournadre and Dorje
and the Wylie transliteration getting better 2003, 102) has the advantage of getting much
results. The neural network tagger performs better automatically chunk-parsed results for
best overall with the larger tag set. With the complex nominals.
smaller tag set, the Wylie transliteration is We used the NLTK chunk-parser10 to com-
best for the smaller tag set. bine tagged tokens into phrases. Semi-
hierarchical structures were created by care-
Global
fully formulating all phrase formation rules
Accuracy
in the correct order, e.g. adjectival phrases
Clas. Tib. (318k; 15 tags) 96.3%
(ADJP) before noun phrases (NP) and deter-
Old Tib. (3.5k; 15 tags) 92.8%
miner phrases (DP) before pre/postpositional
Old & Clas. (321.5k; 15 tags) 96.1%
phrases (PP). A set of sample rules developed
Wylie translit. (318k; 15 tags) 96.5%
to generate a RegEx grammar for Old Tibetan
Unicode Tib. (318k; 79 tags) 95.0%
looks like this:
Wylie translit. (318k; 79 tags) 94.7%
NN-tagger (318k; 79 tags) 95.8% ADJP: {<ADJ><ADJ>?}
NP: {<NOUN|PROPN>}
NUMP: {<NUM><NUM>?}
5 Chunk-Parsing DP: {<DET>?<NP>?<ADJP|NUMP>?<DET>}
DP: {<NP><ADJP|NUMP><ADJP|NUMP>?}
To facilitate further future research, we also
DP: {<NP|DP><case.gen><NP|DP>}
developed a ‘hierarchical chunk-parse’ of our
SbjNP: {<NP|DP><case.agn>}
Old Tibetan corpus. This is a detailed, but
PP: {<DP|NP><ADP>}
rather shallow parse that aims to be as theory-
VP: {<VERB|AUX>?<VERB|AUX>}
neutral as possible. Constituents are com-
ADVP: {<ADV><ADV>?}
bined into phrases where necessary and un-
controversial, in a hierarchical fashion, e.g. Some sample results are shown in (3) and (4):
nouns can combine with adjectives and deter-
(3) (S(SbjNP(NP ད་རྒྱལ་མང་པོ་རྗེ/PROPN) ས་/case.agn)
miners into a Determiner Phrase (DP), which (PP (NP ཞིང་/NOUN) གྱི་/ADP)
can then combine with a post-positional case (NP ཕྱིང་རིལ་/NOUN) (VP བགྱིས/VERB))
marker into a Pre/Postpositional Phrase (PP). da rgyal mang po rje-s zhing gyi phying ril bgyis
With the small tag set, all case markers ‘Dargyal Mangporje carried out a ‘felt roll tax’.’
are automatically converted into adpositions. (4) (S(PP(DP(NP ཞང་ཞུང་ཡུལ་/PROPN) གྱི་/case.gen)
This includes the ‘Agentive Case’ (case.agn) (NP མངན་/NOUN)) དུ་/ADP)
that is used to indicate the subject of transitive (NP ག་གྱིམ་རྩན་རྨ་ཆུང་/PROPN) (VP བཅུག/V)
verbs. If instead we keep this agentive case
marker, our small tag set will be extended, zhang zhung yul gyi mngan du spug gyim rtsan rma
but since this marker is highly consistent in chung bcug
spelling, its Precision, Recall and f-score are 10
http://www.nltk.org
309
‘[He] installed Spug Gyimrtsan Rmachung as the fiscal 6.1 Correction & Error Analysis
governor of the land of Zhang-zhung.’
For the segmentation stage clear errors are in-
stances of case markers and converbs that are
By exploiting the language’s standard head- still attached to the tokens they modify, but
final word order, we can create subordinate these markers should each receive their own
clauses for phrases with nominalised verbs tag. Because of their consistent orthography,
ending in subordinate conjunctions. Similarly, they can often easily be split from their preced-
we can create relative clauses for nominalised ing token to facilitate POS tagging and pars-
verbs followed by the genitive, which functions ing. In addition, these homophonous forms
as a relative marker linking the following word could be checked after POS tagging: their tag
to the preceding relative clause.11 The re- should be a converb following a verb, but a
sults require only minimal manual correction case marker following a noun. Similarly, a
and are sufficiently theory-neutral to facilitate simple dictionary look-up script could ‘check’
morpho-syntactic research within a variety of whether the forms proposed by the segmenter
frameworks. The bracket notation is format- actually exist. In order to make this latter
ted according to the standard .psd guidelines loop-up task work well, however, we first need
and converted to .psdx (a TEI XML ver- to collate and convert Old and/or Classical
sion of .psd) so that they can be queried by dictionaries into a reliable and searchable for-
CorpusSearch,12 CorpusStudio13 or any other mat.
plain text or XML-based way of querying syn-
tactic data. These semi-hierarchical structures 6.1.1 Specific Old Tibetan Errors
are not only useful for historical syntacticians We have detected a number of specific Old Ti-
interested in comparing basic phrasal struc- betan errors as well. In example (5), for in-
ture in different languages, but they are also stance, we can identify some regular mistakes.
invaluable for students and scholars of Tibetan Adverbial expressions like དགུན dgun ‘in winter’,
to get a good insight into how the grammar of དབྱརད� dbyard ‘in summer’, have been tagged
the language has changed over time. Finally, as nouns in many instances, so we can search
this semi-hierarchical phrasal structure serves for these and other recurring adverbial expres-
as a great starting point for further Old Ti- sions and replace their incorrect nominal tags.
betan NLP challenges, such as creating more
meaningful word embeddings and developing (5) བཙནཔོ་ དབྱརད་ སྤེལ་ ན་ བཞུགས ཤིང་
tools for keyphrase extraction, document clus- NOUN ADV PROPN ADP VERB SCONJ
tering and topic modelling.14 “In summer, the emperor stayed in Spel.”
བྲག་མ ར་ ན་ བཞུགས །
11
See Meelen and Roux (fc) for further examples of (6)
semi-automatic syntactic annotation. NOUN ADP ADP VERB PUNCT
12
corpussearch.sourceforge.net/index.html
13 Lit: ‘cliff into/for in stayed’
https://dev.clarin.nl/node/4239
14
After finishing all manual corrections at the end of
the year, the entire annotated corpus and tools will be The correct analysis here instead should com-
made available through Github. bine the ར -r, which was originally tagged as an
310
adposition (ADP) with the preceding noun བྲག་མ extant tools for Classical Tibetan could be
brag ma ‘cliff’, resulting in the proper noun of tested. We then optimised and extended these
the place called ‘Dagmar’:15 tools in various ways and finally developed a
chunk-parser to create the first Old Tibetan
(7) བྲག་མར་ ན་ བཞུགས ། Treebank as an indispensable tool for philolo-
PROPN ADP VERB PUNCT
gists, linguist, but also for scholars in Tibetan
“[he] resided in Dagmar”
studies and the Tibetan communities, as it fa-
This correction, as many others occurring with cilitates the development of good Tibetan dic-
proper nouns, cannot be done automatically tionaries and other Tibetan NLP tools.
since the error patterns are not regular. Some-
times Dagmar, a toponym, is tagged correctly
References
as a proper noun, however, dagma + r ‘into
a cliff’ is also a possible segmentation, in Chernodub, A., Oliynyk, O., Heidenreich, P., Bon-
which case the correct POS tags would be darenko, A., Hagen, M., Biemann, C., and
Panchenko, A. (2019). Targer: Neural argu-
NOUN + ADP. Since the Tibetan script does not ment mining at your fingertips. In Proceedings
identify capital letters, it is difficult for any of the 57th Annual Meeting of the Association
NLP tool to make the right decision in these of Computational Linguistics (ACL’2019), Flo-
cases. It would also be difficult to look up rence, Italy.
ambiguous forms like these in a comprehen- Daelemans, W., Zavrel, J., van den Bosch, A., and
sive, searchable Old Tibetan proper noun lex- Van der Sloot, K. (2003). Mbt: Memory-based
icon (which we are currently developing), as tagger. Reference Guide: ILK Technical Report-
the alternative reading is still possible. This ILK, pages 03–13.
issue is exacerbated by the fact that Tibetan De Marneffe, M.-C., Dozat, T., Silveira, N., Haver-
proper nouns are almost exclusively also nor- inen, K., Ginter, F., Nivre, J., and Manning,
mal nouns, mainly referring to natural phe- C. D. (2014). Universal stanford dependen-
cies: A cross-linguistic typology. In LREC, vol-
nomena, e.g. Nyima ‘sun, Nyima’. ume 14, pages 4585–4592.
6.2 De-Normalisation Eckhoff, H., Bech, K., Bouma, G., Eide, K., Haug,
D., Haugen, O. E., and Jøhndal, M. (2018). The
Since in the pre-processing stage we converted PROIEL treebank family: a standard for early
and normalised our Old Tibetan to ‘Classical attestations of Indo-European languages. Lan-
Tibetan’ orthography, in the post-processing guage Resources and Evaluation, 52(1):29–65.
stage we need to reverse the Cg3 normalization
El-Haj, M., Kruschwitz, U., and Fox, C. (2015).
rules and apply them to the normalised text. Creating language resources for under-resourced
This task is straightforward since the Cg3 nor- languages: methodologies, and experiments
malisation grammar has been created with this with Arabic. Language Resources and Evalua-
de-normalisation process in mind. Through se- tion, 49(3):549–580.
lecting and deselecting the OT and σ tags re- Galves, C. (2018). The Tycho Brahe Corpus
spectively, we converted our Old Tibetan cor- of Historical Portuguese. Linguistic Variation,
pus back to its original form after annotation. 18(1):49–73.
311
Hill, N. W. (2007). Aspirated and unaspirated
voiceless consonants in Old Tibetan. Languages
and Linguistics, 8(2):471–493.
Hill, N. W. (2010). An overview of Old Ti-
betan synchronic phonology. Transactions of
the philological society, 108(2):110–125.
Hill, N. W. (2012). A note on the history and
future of the ‘Wylie’ system. Revue d’Etudes
Tibétaines, 23:103–105.
Meelen, M. and Hill, N. (2017). Segmenting and
POS tagging Classical Tibetan using a memory-
based tagger. Himalayan Linguistics, 16(2):64–
89.
Meelen, M., Hill, N. W., and Handy, C. (2017).
The Annotated Corpus of Classical Tibetan
(ACTib), Part I - Segmented version, based on
the BDRC digitised text collection, tagged with
the Memory- Based Tagger from TiMBL.
Meelen, M. and Roux, E. (fc). Meta-dating the
ACTib.
Rögnvaldsson, E., Ingason, A. K., Sigurðsson,
E. F., and Wallenberg, J. (2012). The Icelandic
Parsed Historical Corpus (IcePaHC). In LREC,
pages 1977–1984.
Taylor, A. and Kroch, A. S. (1994). The Penn-
Helsinki Parsed Corpus of Middle English. Uni-
versity of Pennsylvania.
Tournadre, N. (2010). The Classical Tibetan cases
and their transcategoriality: From sacred gram-
mar to modern linguistics. Himalayan Linguis-
tics, 9(2):87–125.
Tournadre, N. and Dorje, S. (2003). Manual of
standard Tibetan: Language and civilization:
Introduction to standard Tibetan (spoken and
written) followed by an appendix on classical lit-
erary Tibetan. Snow Lion Publications.
312
Summarizing Legal Rulings: Comparative Experiments
313
Proceedings of Recent Advances in Natural Language Processing, pages 313–322,
Varna, Bulgaria, Sep 2–4 2019.
main topics that should appear in the summary. tation is to split the text and generate a vocabu-
Another important issue with manually created lary from the training data keeping the most fre-
summaries is the lack of standardization. Each quent tokens. Even considering that the vocabu-
specialist from each Court has their own writing lary would be extracted from a single language, if
style. A standardized way of writing is desir- we consider that texts have upper and lower case
able as it would provide more homogeneous sum- characters, numbers, dates, etc. the vocabulary
maries. For these reasons, the use of abstractive usually becomes quite large, frequently with hun-
approaches is especially appealing in this area. dreds of thousands of different tokens.
The summarization of legal texts differs re- A large vocabulary is a problem because the
markably from mainstream work in text summa- output layer of the neural network must have its
rization, which is mostly devoted to summarizing size. This means that the probability of choosing
news articles, headlines or tweets. Legal texts are the correct output diminishes as the vocabulary in-
generally lengthier and typically contain complex creases. Also, even if infrequent tokens are repre-
vocabulary and expressions. Also, the order of the sented in the vocabulary, its infrequent use would
words in some expressions can make a big differ- not be sufficient for the model to learn when to use
ence in their meaning, e.g.,“denied an appeal that it correctly.
had accepted” is very different from “accepted an One approach to deal with the size of the vo-
appeal that had denied”. cabulary is to convert all characters to lowercase
In this paper, we investigate the suitability of and replace numbers and dates with zero represen-
extractive and abstractive approaches in summa- tations. With these modifications, the vocabulary
rizing legal rulings. Given that the vast majority of becomes smaller, but the model will become less
works in this area focused solely on news datasets, capable of generating outputs in the same way as
we believe that testing summarization on a new humans would.
domain is important given the different nature of
Even with the text simplifications, the problem
the input documents. Fourteen approaches were
with out of vocabulary words (OOV) remains be-
tested over a real dataset containing 10K rulings
cause not all possible words will be represented.
from the Brazilian Supreme Court (Feijó and Mor-
So, it is usual to represent any token that is not
eira, 2018). Thus, another contribution is using a
present in the vocabulary by a reserved OOV to-
language that is not typically included in summa-
ken. This token would be used for uncommon
rization experiments.
names, dates, and numbers.
The remainder of this paper is organized as fol-
Another approach is to use one token per char-
lows. Section 2 discusses text representations and
acter. This greatly reduces the vocabulary size
introduces the problem of out of vocabulary to-
only using lower and upper case characters, num-
kens. Section 3 revises recent works on abstractive
bers, and symbols. Although the vocabulary is
summarization. Section 4 describes the dataset
smaller, probably not more than few hundred to-
and how it was prepared to be used in the experi-
kens, it would require each word to be represented
ments. Section 5 presents a simple heuristic-based
by several tokens, leading to very long document.
extractive summarizer that we proposed to serve
This is problematic because the model will require
as a baseline. Sections 6 and 7 describes the al-
a large memory to be able to generate the summary
gorithms and models that were investigated here.
without repeating already generated tokens.
Section 8 discusses our results and findings. Fi-
nally, Section 9 concludes this paper and points The alternative to mitigate the disadvantages of
out possibilities for future work. these two approaches is to use sub-word units as
the token representation (Schuster and Nakajima,
2 Text Representation 2012; Chitnis and DeNero, 2015; Sennrich et al.,
2015; Kudo, 2018). The idea is that a token would
Representing texts using a sequence of words is represent a common pattern seen in the training
useful for exchanging information between hu- data rather than words or characters. This oper-
mans. However, when using neural models, we ates with a fixed vocabulary size and assigns a to-
need to convert the text into numbers that could be ken for the most common patterns found. With
used as input into the neural network. this method, the OOV problem is reduced as one
The usual way of generating a text represen- word now can be represented by a combination of
314
sub-words. The problem of longer sequences is marization in the sense that both require some text
also addressed because a token now can represent comprehension before an output can be generated.
several characters. When translating, a model does not have an exact
alignment for each word read in the source to the
3 Related Work word generated in the output. This happens be-
cause a source word may be represented by more
Neural Networks are models capable of learning (or fewer) words in the translation. Also, the order
very complex functions. In the last few years, of the tokens can be different.
there has been significant interest in applying them
Wu et al. (2016) describe the architecture
for natural language tasks such as automatic trans-
used for the Google Neural Machine Translation
lation and summarization.
(GNMT) system. In that work, they used a LSTM
One of the first issues that needs to be ad-
network with 8 encoders and 8 decoders using
dressed is that Neural Networks require that both
attention and residual connections. In the beam
inputs and outputs have a fixed a length, and that
search, they used length-normalization and ap-
is not the case when dealing with text, because
plied a coverage penalty to favor an output that
each document (or sentence) can have a differ-
is able to cover more words from the source se-
ent length. In order to overcome this limitation,
quence.
Sutskever et al. (2014) introduced a general end-
to-end approach capable of representing sequence- Vaswani et al. (2017) improved the GNMT sys-
to-sequence models using LSTM (Long Short- tem replacing entirely the recurrence and convolu-
Term Memory) cells. Nevertheless, both input and tions by an attention-based model known as Trans-
output must still have fixed lengths large enough former. The model is quite complex and relies on
to fit. The network is trained to output the end- many training variables, but it has the advantage
of-sentence (EOS) token when the output is large of allowing more parallel computation. With this
enough. With this approach, both source input and modification, the model is able to use many GPUs
generated sequence may logically have different and train a lot faster.
lengths and do not require any type of alignment See et al. (2017) addressed two common prob-
between input and output. lems found in the application of RNNs in the con-
Bahdanau et al. (2014) proposed a model for text of summarization. First, the problem of rare
learning to align the input with the generated out- words that were not present in the vocabulary was
put. They realized that the approach for encoding solved by having hybrid pointer networks that are
the whole text before starting to decode requires capable of using the source word as output when
that the whole idea is represented by just one vec- the attention weight is high enough. The second
tor. So, they proposed to use auxiliary vectors to problem is using a coverage vector that represents
represent the alignment of the input in relation to the weighted sum of the attention vectors. Then,
the generated output. Their approach significantly to increase coverage, their model is trained to pe-
improved the quality of the generated outputs and nalize every time the attention vectors are high in
became known as Attention vectors. the same regions, encouraging the model to better
Following the same ideas, Luong et al. (2015) distribute the attention over the source input.
explored different versions for the attention mech- In our work, the problem of rare words has been
anism. They also noticed that as the length of the diminished using sub-word encoding, as discussed
input increases, the attention vector has a lot more in Section 2. The coverage mechanism, proposed
difficulty in learning the weights. So they evalu- to deal with repetition problem, is complex to
ated the impact of using a local attention mecha- train. It is used after the training to condition the
nism to look just for a smaller portion of the source decoder to avoid generating attention vector using
at each time step. the same positions that were already used. Paulus
The attention mechanism works well for short et al. (2017) proposed trigram avoidance during
texts, but struggles to focus on relevant informa- the beam search. This approach is much simpler
tion when applied to long documents, common in and easy to apply but does not really fix the prob-
the legal domain. lem, as it still will happen, it only masks its effects
Neural Machine Translation (NMT) is a for the evaluation.
sequence-to-sequence task that is similar to sum- There is a growing interest in using reinforce-
315
ment learning approaches (Paulus et al., 2017; source input.
Li et al., 2018; Celikyilmaz et al., 2018) to im-
prove summarization performance. In general, re- 4.2 Official Rouge Script
inforcement learning is employed when an non-
The standard evaluation metric for text summa-
differentiable operation is being used. These
rization is called Recall-Oriented Understudy for
methods have the disadvantage of being hard to
Gisting Evaluation (ROUGE). The general idea
tune and generally slow to converge.
of this metric is to count the number overlapping
Chen and Bansal (2018) proposed a method for
units between one or more reference summaries
using both extracting and abstracting approaches.
and the machine-generated summary. It is ex-
Their idea was to select salient sentences and
pected, that a high-quality summary should use the
rewrite them abstractively. Reinforcement learn-
same words found in the reference summaries and
ing was used to combine these two neural net-
preferably in the same order.
works. Their idea seems to be a good approach
for working with long documents. Although, they The results reported here include Precision, Re-
did not report results for long documents datasets. call, and F-measure for ROUGE-1, ROUGE-2,
We intend to further explore the application of this and ROUGE-L metrics. A confidence of 95%
technique to long documents in the legal domain. was adopted. We used the official ROUGE (Lin,
2004) 1.5.5 script in our experiments. Most pa-
4 Materials and Methods rameters were set to their default values. The
pyrouge package, which is a wrapper to the of-
Our summarization experiment in the legal area ficial script, was used to provide the required out-
uses rulings written in Portuguese. There are some put text (in HTML) and the configuration file. The
peculiarities when using a language different from only change was to remove the call to the English
English, e.g., we need to check if the standard Porter Stemmer, since our texts are in Portuguese.
summarization evaluation (designed for English)
can be directly applied. 4.3 Text Preprocessing
316
5 A Simple Extractive Summarization Limit Mean Min Max Std Dev
Baseline
No 190 20 1,909 179.46
When generating summaries, the source input 600 180 20 600 131.89
length and the desired summary length are re- 450 173 20 450 112.02
quired. Ideally, these two lengths should be au- 300 158 20 300 82.37
tomatically determined by the algorithm, but that 150 120 20 120 36.85
is not how these standard extractive approaches
Table 1: Lengths of the reference summaries that
work. In order to establish these lengths, we de-
compose our Test Set.
fined a heuristic-based baseline extractive method.
317
Length R1-F R1-P R1-R R2-F R2-P R2-R RL-F RL-P RL-R
600+600 33.99 33.46 43.16 12.20 12.34 15.59 19.44 19.10 25.36
450+450 34.07 34.40 41.32 12.06 12.50 14.72 19.48 19.64 24.24
300+300 34.47 34.84 40.25 12.08 12.43 14.19 19.74 19.88 23.58
150+150 34.47 34.32 37.46 11.88 11.84 13.01 20.49 20.29 22.61
Table 2: ROUGE scores for different source lengths. The results show that the F-measure is reasonably
stable across different lengths.
318
decoding, and it is configured to ignore outputs a training phase for the neural network models,
that were shorter than the minimum length. ROUGE results are reported only for the test set.
Both NMT and Transformer models use word One advantage of extractive algorithms is that
embedder of size 512. Each model was evalu- they do not require prior training, and they can be
ated until its training loss was no longer dimin- applied directly over the test data. On the other
ishing. We report ROUGE results with minimum hand, after the time-consuming training, the ab-
decoding lengths of 100 and 120 tokens. Recall stractive approaches can create the summary a lot
that we are using SentencePiece and each decoded faster.
word may be represented by more than one token. Table 4 shows reasonably good results for both
So, the generated output may contain fewer words NMT and Transformer models. There was a small
than this minimum length. In all reported results, advantage for the standard Transformer model
we show the mean length of the output considering when compared to its modified version with the
generated tokens separated by spaces. Average Attention Network. They both have
Two NMT configurations were used. NMT- reached very similar results and have converged
Small uses 2-layers, unidirectional LSTM with in about 40K steps.
512 units, and it has converged in 15,000 steps. Since the Transformer model has many vari-
NMTMedium uses 4-layers, bidirectional LSTM, ables, it requires a lot of memory to run. So, the
with 512 units and it has converged in 26,000 batches need to be smaller. As a consequence, it
steps. Transformer model uses the configuration needed more steps to converge. Despite that, we
as originally proposed by Vaswani et al. (2017). observed that it trains faster than standard RNNs.
TransformerAAN uses cumulative average atten- As we are using a concurrent environment, our
tion network in the decoder as described in (Zhang measures of the time taken for training were not
et al., 2018). The objective is to reduce the re- accurate, so we could not report them.
quired training and inference time.
Summarization results in other datasets are not
directly comparable to our results. Still, they
8 Results and Discussion
may serve as reference. Zhang et al. (2019)
The performance of the extractive algorithms reports that the current state-of-the-art for the
shown in Table 3 was disappointing. With the CNN/Daily Mail dataset (Nallapati et al., 2016)
exception of SumBasic, all other algorithms have reaches scores of ROUGE-1 41.71, ROUGE-2
performed worse than our simple Baseline by at 19.49 and ROUGE-L 38.79.
least 0.6 points in ROUGE-L. In some cases, the The summaries generated by the abstractive ap-
performances were not far from the random base- proaches were promising. They look similar to
line. A possible explanation for such poor re- those produced by humans. In most generated
sults is the limitation of this approach of generat- summaries, the main topic was correctly captured
ing summaries using only complete sentences that by the summarizer. Nevertheless, as shown in Fig-
were present in the source text. Looking at the ure 1 there are still some cases in which the sum-
generated summaries, most of them have selected marizer barely captured any meaning of the text,
just one very long sentence while others used a few generating summaries that had almost no relation
random disconnected sentences. with the expected output. In these cases, the ex-
Our experiments varying the lengths of the in- tractive approach would probably have done bet-
put method (Table 2) have shown that even with ter. In other cases, the general meaning was cor-
larger source inputs, which could contain more to- rectly captured, but the output had repeating ex-
kens that should be present in the output, the per- pressions. We believe this may have been caused
formance was decreasing. by the minimum length restriction.
The baseline results provided by Feijó and Mor- Legal operators rely on summaries to their jobs,
eira (2018) for the extractive methods ranged be- since it is impossible to read the full contents of
tween 11 and 16 points in terms of ROUGE-L. each decision to find precedents for their cases.
Those results cannot be directly compared to the Missing or referring to an incorrect precedent may
results in our experiments because they had re- cause the petition to be denied and the case would
moved stopwords and they reported results for the be lost. Thus, considering the results seen so far,
entire dataset. Since in this paper, we require neither approach delivers results that could safely
319
Algorithm R1-F R1-P R1-R R2-F R2-P R2-R RL-F RL-P RL-R
Random 31.52 34.42 34.81 10.55 11.81 11.49 17.88 19.67 19.99
Baseline 34.47 34.84 40.25 12.08 12.43 14.19 19.74 19.88 23.58
Luhn 33.16 33.17 39.08 11.06 11.25 13.09 18.77 18.67 22.65
LexRank 34.06 34.06 40.07 11.65 11.85 13.69 19.16 19.04 23.06
LSA 32.31 32.26 38.04 10.44 10.62 12.23 17.88 17.76 21.50
KLSum 31.96 32.42 37.14 11.45 11.74 13.30 18.24 18.38 21.66
SumBasic 34.51 34.41 40.74 12.32 12.49 14.46 18.76 18.69 22.43
TextRank1 33.09 33.07 38.99 10.85 11.07 12.73 18.78 18.67 22.60
TextRank2 33.66 34.14 39.10 12.00 12.31 13.97 19.16 19.24 22.82
Model Len R1-F R1-P R1-R R2-F R2-P R2-R RL-F RL-P RL-R
NMTSmall 130 38.86 44.75 40.42 21.28 23.14 22.89 30.22 33.99 32.02
NMTMedium 130 43.25 49.25 44.80 25.41 27.60 27.05 33.91 37.78 35.69
Transformer 134 44.27 49.38 46.24 26.50 28.36 28.26 35.27 38.52 37.36
TransformerAAN 137 43.67 48.38 45.90 25.60 27.15 27.43 34.47 37.38 36.74
NMTSmall 141 38.37 42.48 41.54 20.77 21.77 23.29 29.55 31.92 32.68
NMTMedium 140 41.56 46.44 44.00 23.43 24.95 25.56 32.01 34.87 34.58
Transformer 145 43.91 47.34 47.76 25.95 26.93 28.84 34.55 36.48 38.16
TransformerAAN 147 43.39 46.46 47.37 25.23 25.93 28.13 33.90 35.51 37.60
Table 4: ROUGE scores using abstractive models. The mean length is shown because it affects the
scores.
320
References Aria Haghighi and Lucy Vanderwende. 2009. Explor-
ing content models for multi-document summariza-
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- tion. In Proceedings of Human Language Tech-
gio. 2014. Neural machine translation by jointly nologies: The 2009 Annual Conference of the North
learning to align and translate. arXiv preprint American Chapter of the Association for Compu-
arXiv:1409.0473 . tational Linguistics. Association for Computational
Federico Barrios, Federico López, Luis Argerich, and Linguistics, pages 362–370.
Rosa Wachenchauzer. 2016. Variations of the simi- Guillaume Klein, Yoon Kim, Yuntian Deng, Jean
larity function of textrank for automated summariza- Senellart, and Alexander Rush. 2017. Open-
tion. arXiv preprint arXiv:1602.03606 . NMT: Open-source toolkit for neural machine
Mišo Belica. 2018. sumy: Module for automatic sum- translation. In Proceedings of ACL 2017, Sys-
marization of text documents and HTML pages. tem Demonstrations. Association for Computa-
https://github.com/miso-belica/sumy. tional Linguistics, Vancouver, Canada, pages 67–72.
https://www.aclweb.org/anthology/P17-4012.
Asli Celikyilmaz, Antoine Bosselut, Xiaodong He, and
Yejin Choi. 2018. Deep communicating agents Taku Kudo. 2018. Subword regularization: Improv-
for abstractive summarization. arXiv preprint ing neural network translation models with mul-
arXiv:1803.10357 . tiple subword candidates. CoRR abs/1804.10959.
http://arxiv.org/abs/1804.10959.
Yen-Chun Chen and Mohit Bansal. 2018. Fast ab-
stractive summarization with reinforce-selected Piji Li, Lidong Bing, and Wai Lam. 2018. Actor-
sentence rewriting. CoRR abs/1805.11080. critic based training framework for abstractive sum-
http://arxiv.org/abs/1805.11080. marization. arXiv preprint arXiv:1803.11070 .
Rohan Chitnis and John DeNero. 2015. Variable- Chin-Yew Lin. 2004. Rouge: A package for auto-
length word encodings for neural translation models. matic evaluation of summaries. Text Summarization
In Proceedings of the 2015 Conference on Empiri- Branches Out .
cal Methods in Natural Language Processing. pages H. P. Luhn. 1958. The automatic creation of liter-
2088–2093. ature abstracts. IBM J. Res. Dev. 2(2):159–165.
H. P. Edmundson. 1969. New methods in au- https://doi.org/10.1147/rd.22.0159.
tomatic extracting. J. ACM 16(2):264–285. Minh-Thang Luong, Hieu Pham, and Christo-
https://doi.org/10.1145/321510.321519. pher D. Manning. 2015. Effective approaches to
Günes Erkan and Dragomir R. Radev. 2004a. Lexrank: attention-based neural machine translation. CoRR
Graph-based lexical centrality as salience in text abs/1508.04025. http://arxiv.org/abs/1508.04025.
summarization. J. Artif. Int. Res. 22(1):457–479. R. Mihalcea and P. Tarau. 2004. TextRank: Bringing
http://dl.acm.org/citation.cfm?id=1622487.1622501. order into texts. In Proceedings of EMNLP-04and
Günes Erkan and Dragomir R Radev. 2004b. Lexrank: the 2004 Conference on Empirical Methods in Nat-
Graph-based lexical centrality as salience in text ural Language Processing.
summarization. Journal of Artificial Intelligence Ramesh Nallapati, Bing Xiang, and Bowen
Research 22:457–479. Zhou. 2016. Sequence-to-sequence rnns for
Diego de Vargas Feijó and Viviane Pereira Moreira. text summarization. CoRR abs/1602.06023.
2018. Rulingbr: A summarization dataset for legal http://arxiv.org/abs/1602.06023.
texts. In Aline Villavicencio, Viviane Moreira, Al- Ani Nenkova and Lucy Vanderwende. 2005. The im-
berto Abad, Helena Caseli, Pablo Gamallo, Carlos pact of frequency on summarization. Microsoft Re-
Ramisch, Hugo Gonçalo Oliveira, and Gustavo Hen- search, Redmond, Washington, Tech. Rep. MSR-TR-
rique Paetzold, editors, Computational Processing 2005 101.
of the Portuguese Language. Springer International
Publishing, Cham, pages 255–264. Romain Paulus, Caiming Xiong, and Richard Socher.
2017. A deep reinforced model for abstractive sum-
Yihong Gong and Xin Liu. 2001. Generic text summa- marization. arXiv preprint arXiv:1705.04304 .
rization using relevance measure and latent semantic
analysis. In Proceedings of the 24th annual inter- Radim Řehůřek and Petr Sojka. 2010. Software Frame-
national ACM SIGIR conference on Research and work for Topic Modelling with Large Corpora. In
development in information retrieval. ACM, pages Proceedings of the LREC 2010 Workshop on New
19–25. Challenges for NLP Frameworks. ELRA, Valletta,
Malta, pages 45–50.
Google/SentencePiece. 2019. Unsupervised text
tokenizer for neural network-based text gen- Mike Schuster and Kaisuke Nakajima. 2012. Japanese
eration. https://github.com/google/ and korean voice search. In 2012 IEEE Interna-
sentencepiece. Accessed: 2019-05-25. tional Conference on Acoustics, Speech and Signal
Processing (ICASSP). IEEE, pages 5149–5152.
321
Abigail See, Peter J Liu, and Christopher D Man- Le, Mohammad Norouzi, Wolfgang Macherey,
ning. 2017. Get to the point: Summarization Maxim Krikun, Yuan Cao, Qin Gao, Klaus
with pointer-generator networks. arXiv preprint Macherey, Jeff Klingner, Apurva Shah, Melvin
arXiv:1704.04368 . Johnson, Xiaobing Liu, Lukasz Kaiser, Stephan
Gouws, Yoshikiyo Kato, Taku Kudo, Hideto
Rico Sennrich, Barry Haddow, and Alexandra Birch. Kazawa, Keith Stevens, George Kurian, Nishant
2015. Neural machine translation of rare words with Patil, Wei Wang, Cliff Young, Jason Smith, Jason
subword units. arXiv preprint arXiv:1508.07909 . Riesa, Alex Rudnick, Oriol Vinyals, Greg Corrado,
Macduff Hughes, and Jeffrey Dean. 2016. Google’s
Josef Steinberger and Karel Jezek. 2004. Using latent neural machine translation system: Bridging the gap
semantic analysis in text summarization and sum- between human and machine translation. CoRR
mary evaluation. Proc. ISIM 4:93–100. abs/1609.08144. http://arxiv.org/abs/1609.08144.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Biao Zhang, Deyi Xiong, and Jinsong Su. 2018.
Le. 2014. Sequence to sequence learning Accelerating neural transformer via an aver-
with neural networks. CoRR abs/1409.3215. age attention network. CoRR abs/1805.00631.
http://arxiv.org/abs/1409.3215. http://arxiv.org/abs/1805.00631.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Haoyu Zhang, Yeyun Gong, Yu Yan, Nan Duan, Jian-
Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
jun Xu, Ji Wang, Ming Gong, and Ming Zhou.
Kaiser, and Illia Polosukhin. 2017. Atten-
2019. Pretraining-based natural language gen-
tion is all you need. CoRR abs/1706.03762.
eration for text summarization. arXiv preprint
http://arxiv.org/abs/1706.03762.
arXiv:1902.09243 .
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V.
322
Entropy as a Proxy for Gap Complexity in Open Cloze Tests
323
Proceedings of Recent Advances in Natural Language Processing, pages 323–327,
Varna, Bulgaria, Sep 2–4 2019.
ideally predict the quality of the gaps during the Equation 1, where P (xi ) stands for the probabil-
generation process. In this regard, Skory and Es- ity of event xi , i.e. the probability that each word
kenazi (2010) observe that Shannon’s information in the vocabulary occurs in the evaluated context.
theory (Shannon, 1948) could be used to estimate
n
the reading difficulty of answers to a gap based X
H(X) = − P (xi ) log2 P (xi ) (1)
on their probability of occurrence. Thus, for the
i=1
sentence “She drives a nice ”, the word
“car” would be the most likely answer (lowest In this work, we use entropy to assign a score
readability level) while words such as “taxi”, to each gap based on the number of valid words
“tank” and “ambulance” would be at increasingly that could fill in the slot given the surrounding con-
higher levels. text. As a result, gaps with many possible answers
Research on predicting the difficulty of cloze will yield higher entropy than those with fewer an-
tests is also directly relevant to this work. Bein- swers.
born et al. (2014) built models to predict the diffi-
culty of C-tests (i.e. gaps with half of the required 4 Experiments
word removed) at the gap and test level and later
We followed Malafeev’s (2014) approach and used
extended their approach to cover closed cloze tests
open cloze tests from Cambridge English exami-
(Beinborn et al., 2015; Beinborn, 2016). More re-
nations as our gold standard data, since they are
cently, Pandarova et al. (2019) presented a diffi-
manually created by experts in the field of EFL
culty prediction model for cued gap-fill exercises
testing. We collected the sample open cloze tests
aimed at practising English verb tenses while Lee
for KET, FCE, CAE and CPE exams that are fea-
et al. (2019) investigated how difficulty predic-
tured in their respective online handbooks1 (one
tions could be manipulated to adapt tests to a target
per exam together with their answers). These ex-
proficiency level. Unlike our work, however, all
ams correspond respectively to levels A2, B2, C1
these approaches are supervised and not applied
and C2 in the Common European Framework of
to open cloze tests.
Reference for Languages (CEFR). An open cloze
test is not included in the PET (B1) exam, which is
3 Entropy
why it has not been included in our experiments.
In this paper, we build on the assumption that the For each exam, we restored the original text by
complexity of a gap is correlated to the number using the answers provided (using the first alter-
of possible answers determined by the surround- native if there were many) and created 10 differ-
ing context and the likelihood of each answer. ent variations of the open cloze tests by inserting
As noted by Pino et al. (2008), high-quality open gaps randomly throughout the text. We created the
cloze questions should sufficiently narrow the con- same number of gaps as in the original tests.
text of each gap in order to avoid multiple valid an- For each original and automatically generated
swers, which would make the exercise too broad test, we compute entropy per gap using a 5-gram
in scope and therefore ineffective. We thus as- language model trained on the 1 Billion Word
sume that gaps with more restricted context elic- WMT 2011 News Crawl corpus2 using KenLM
iting very specific answers should be more useful (Heafield, 2011). We use the language model bidi-
than broad gaps with very general answers, so the rectionally, taking 3 words to the left and right
less “branching” that a gap allows, the better. of each gap to predict the probability of the next
This property can be modelled by entropy, and previous words respectively. Since we obtain
which quantifies the amount of information con- a probability for all the words in our vocabulary
veyed by an event. Intuitively, entropy can be con- (> 82, 200 words) given the left and right con-
sidered a measure of disorder, uncertainty or sur- text individually, we multiply the probabilities for
prise. If the probability of an event is very high, each word to get a unified “bidirectional” proba-
entropy will be low (i.e. there is less surprise about bility (see Figure 1). Given that this can lead to
what will happen) while events with low probabili- infinitesimal probabilities that can affect computa-
ties will lead to higher entropy. Shannon’s entropy, 1
https://www.cambridgeenglish.org/
a common formulation to measure the number of exams-and-tests/
bytes needed to encode information, is shown in 2
https://www.statmt.org/lm-benchmark/
324
Electronics firms, for example, expect to have only six months CEFR Number Avg. gap
after they have introduced a new product before a rival Exam
level of gaps entropy
company produces a efficient or cheaper alternative.
−→ ←− KET A2 11 1.29 ± 0.69
wide more more FCE B2 9 2.33 ± 1.28
variety very energy CAE C1 9 2.69 ± 1.22
lot less is
quarter is as
CPE C2 9 5.16 ± 3.38
line and less
... ... ...
Table 1: Characterisation of our gold standard data.
325
Average gap entropy per test
Exam
1 2 3 4 5 6 7 8 9 10
KET 5.40 3.47 3.63 3.73 4.22 4.18 4.60 4.33 4.74 4.34
FCE 4.53 7.13 6.12 4.30 2.23 4.01 2.45 3.93 4.18 5.67
CAE 4.70 4.66 3.57 2.68 3.26 4.38 2.79 5.08 5.07 4.82
CPE 6.58 3.91 2.72 4.43 5.02 4.18 5.83 5.46 4.02 3.26
Table 2: Average gap entropy for the automatically generated tests. Values lower than the gold standard are marked
in bold.
326
Kenneth Heafield. 2011. KenLM: Faster and smaller Claude Elwood Shannon. 1948. A mathematical the-
language model queries. In Proceedings of the ory of communication. The Bell System Technical
Sixth Workshop on Statistical Machine Translation, Journal, 27(3):379–423.
pages 187–197, Edinburgh, Scotland. Association
for Computational Linguistics. Adam Skory and Maxine Eskenazi. 2010. Predicting
cloze task quality for vocabulary training. In Pro-
Ji-Ung Lee, Erik Schwan, and Christian M. Meyer. ceedings of the NAACL HLT 2010 Fifth Workshop
2019. Manipulating the difficulty of c-tests. CoRR, on Innovative Use of NLP for Building Educational
abs/1906.06905. Applications, pages 49–56, Los Angeles, California.
Association for Computational Linguistics.
John Lee and Stephanie Seneff. 2007. Automatic
generation of cloze items for prepositions. In Simon Smith, P. V. S. Avinesh, and Adam Kilgarriff.
Eighth Annual Conference of the International 2010. Gap-fill tests for language learners: Corpus-
Speech Communication Association, pages 2173– driven item generation. In Proceedings of ICON-
2176, Antwerp, Belgium. 2010: 8th International Conference on Natural Lan-
guage Processing, pages 1–6. Macmillan Publishers.
Y. Lin, L. Sung, and M. Chen. 2007. An auto-
matic multiple-choice question generation scheme Eiichiro Sumita, Fumiaki Sugaya, and Seiichi Ya-
for english adjective understanding. In Workshop mamoto. 2005. Measuring non-native speak-
on Modeling, Management and Generation of Prob- ers’ proficiency of english by using a test with
lems/Questions in eLearning, pages 137–142. 15th automatically-generated fill-in-the-blank questions.
International Conference on Computers in Educa- In Proceedings of the Second Workshop on Building
tion (ICCE 2007). Educational Applications Using NLP, EdAppsNLP
05, pages 61–68, Stroudsburg, PA, USA. Associa-
Alexey Malafeev. 2014. Language exercise generation: tion for Computational Linguistics.
Emulating cambridge open cloze. Int. J. Concept.
Struct. Smart Appl., 2(2):20–35. Yuni Susanti, Hitoshi Nishikawa, Takenobu Tokunaga,
and Hiroyuki Obari. 2016. Item difficulty analysis
Edison Marrese-Taylor, Ai Nakajima, Yutaka Matsuo, of english vocabulary questions. In CSEDU 2016
and Ono Yuichi. 2018. Learning to automatically - Proceedings of the 8th International Conference
generate fill-in-the-blank quizzes. In Proceedings on Computer Supported Education, volume 1, pages
of the 5th Workshop on Natural Language Pro- 267–274.
cessing Techniques for Educational Applications,
pages 152–156, Melbourne, Australia. Association
for Computational Linguistics.
Ruslan Mitkov and Le An Ha. 2003. Computer-aided
generation of multiple-choice tests. In Proceedings
of the HLT-NAACL 03 Workshop on Building Edu-
cational Applications Using Natural Language Pro-
cessing, pages 17–22.
Irina Pandarova, Torben Schmidt, Johannes Hartig,
Ahcène Boubekki, Roger Dale Jones, and Ulf
Brefeld. 2019. Predicting the difficulty of exercise
items for dynamic difficulty adaptation in adaptive
language tutoring. International Journal of Artifi-
cial Intelligence in Education.
Juan Pino and Maxine Eskenazi. 2009. Measuring hint
level in open cloze questions. In Twenty-Second In-
ternational FLAIRS Conference, pages 460–465.
Juan Pino, Michael Heilman, and Maxine Eskenazi.
2008. A selection strategy to improve cloze question
quality. Intelligent Tutoring Systems for Ill-Defined
Domains: Assessment and Feedback in Ill-Defined
Domains., page 22.
Keisuke Sakaguchi, Yuki Arase, and Mamoru Ko-
machi. 2013. Discriminative approach to fill-in-the-
blank quiz generation for language learners. In Pro-
ceedings of the 51st Annual Meeting of the Associa-
tion for Computational Linguistics (Volume 2: Short
Papers), pages 238–242, Sofia, Bulgaria. Associa-
tion for Computational Linguistics.
327
Song Lyrics Summarization Inspired by Audio Thumbnailing
328
Proceedings of Recent Advances in Natural Language Processing, pages 328–337,
Varna, Bulgaria, Sep 2–4 2019.
2 Summarization Methods (Arora and Ravindran, 2008) to model each sen-
tence’s distribution over latent topics. Another
This section reports on the related work on both type of summarization methods is graph-based
text and audio summarization methods. summarizers. Three of the most popular graph-
based summarizers are TextRank (Mihalcea and
2.1 Text Summarization
Tarau, 2004), LexRank (Erkan and Radev, 2004),
In the literature, there are two different families and (Parveen et al., 2015). These methods work
of approaches for automatic text summarization: by constructing a graph whose nodes are sentences
extraction and abstraction (Allahyari et al., 2017). and whose graph edge weights are sentence sim-
Extractive summarization methods identify impor- ilarities. Then, the sentences that are central to
tant elements of the text and generate them verba- the graph are found by computing the PageRank
tim (they depend only on extraction of sentences (Page et al., 1999). Contrarily to all previously
or words from the original text). In contrast, ab- described methods, systems using supervised ma-
stractive summarization methods interpret and ex- chine learning form another type of summarizers.
amine the text to generate a new shorter text that For instance, (Fattah, 2014) treats extractive sum-
conveys the most critical information from the marization as a binary classification task, where
original text. Even though summaries created by they extract indicator features from sentences of
humans are usually not extractive, most of the gold summaries and learn to detect the sentences
summarization research has focused on extractive that should be included in a summary.
methods. Purely extractive summaries often give
better results (Nallapati et al., 2016), due to the Context-Specific Summarization. If specific
fact that latter methods cope with more complex knowledge about the application scenario or the
problems such as semantic representation, infer- domain of the summarized text is available,
ence and natural language generation. Existing ab- generic summarization methods can be adapted to
stractive summarizers often rely on an extractive take into account the prior information. In query-
pre-processing component to produce the abstract based summarization (Otterbacher et al., 2005;
of the text (Berg-Kirkpatrick et al., 2011; Knight Wang et al., 2016), the user’s query is taken into
and Marcu, 2000). Consequently, in this paper we account when generating a summary. Summa-
focus on extractive summarization methods, also rization of a scientific paper can be improved by
given the fact that lyrics i) strongly use figurative considering the citations of it, as in (Delort et al.,
language which makes abstractive summarization 2003). However, to the best of our knowledge no
even more challenging; and ii) the choice of the summarization methods have been proposed for
words by the composer may also have an impor- the domain of song texts. In this paper we present
tance for capturing the style of the song. a summarization method that uses prior knowl-
In the following, we focus on unsupervised edge about the text it summarizes to help generic
methods for text summarization, the ones targeted summarizers generate better summaries.
in our study (no available gold-standard of human-
produced summaries of song texts exists). Most Evaluation Criteria and Methods. Summaries
methods have in common the process for sum- should i) contain the most important information
mary generation: given a text, the importance of from input documents, ii) not contain redundant
each sentence of that text is determined. Then, the information, iii) be readable, hence they should
sentences with highest importance are selected to be grammatical and coherent (Parveen and Strube,
form a summary. The ways different summarizers 2015). While a multitude of methods to iden-
determine the importance of each sentence may tify important sentences has been described above,
differ: Statistics-based summarizers extract indi- several approaches aim to make summaries less
cator features from each sentence, e.g. (Fattah and redundant and more coherent. The simplest way
Ren, 2009) use among others the sentence position to evaluate summaries is to let humans assess
and length and named entities as features. Topic- the quality, but this is extremely expensive. The
based summarizers aim to represent each sentence factors that humans must consider when giving
by its underlying topics. For instance, (Hennig, scores to each candidate summary are grammati-
2009) apply Probabilistic Latent Semantic Anal- cality, non redundancy, integration of most impor-
ysis, while Latent Dirichlet Allocation is used in tant pieces of information, structure and coherence
329
(Saggion and Poibeau, 2013). The more common S consisting of n lines of text, S = (x1 , ..., xn ),
way is to let humans generate possibly multiple we define the task of extractive lyrics summariza-
summaries for a text and then automatically assess tion as the task of producing a concise summary
how close a machine-made summary is to the hu- sum of the song text, consisting of a subset of the
man gold summaries computing ROUGE scores original text lines: sum(S) ⊆ S, where usually
(Lin, 2004), which boils down to measuring n- |sum(S)| << |S|. We define the goal of a sum-
gram overlaps between gold summaries and auto- mary as to preserve key information and the over-
matic summary. More recently there have been all meaning of a song text. To address this task, we
attempts to rate summaries automatically with- apply the following methods from the literature:
out the need for gold summaries (Nenkova et al., the popular graph-based summarizer TextRank;
2011). The key idea is that a summary should be an adaptation of a topic-based method (TopSum).
similar to the original text in regard to characteris- Moreover, we introduce a method inspired by au-
tic criteria as the word distribution. (Mackie et al., dio thumbnailing (which we dub Lyrics Thumb-
2014) find that topic words are a suitable metric to nail) which aims at creating a summary from the
automatically evaluate micro blog summaries. most representative parts of the original song text.
While for TextRank we rely on the off-the-shelf
2.2 Audio Summarization implementation of (Barrios et al., 2016), in the fol-
Lyrics are texts that accompany music. Therefore, lowing we describe the other two methods.
it is worthwhile to see if methods in audio sum-
marization can be transferred to lyrics summariza- 3.1 TopSum
tion. In audio summarization the goal is to find the
most representative parts in a song, in Pop songs We implement a simple topic-based summariza-
those are usually the chorus and the bridge, in in- tion model that aims to construct a summary
strumental music the main theme. The task of cre- whose topic distribution is as similar as possible
ating short audio summaries is also known as au- to that of the original text. Following (Kleedorfer
dio thumbnailing (Bartsch and Wakefield, 2005; et al., 2008), we train a topic model by factorizing
Chai and Vercoe, 2003; Levy et al., 2006), as the a tf-idf-weighted term-document matrix of a song
goal is to produce a short representation of the text corpus (see Section 4.2) using non-negative
music that fits onto a thumbnail, but still covers matrix factorization into a term-topic and a topic-
the most representative parts of it. In a recent ap- document matrix. Given the learnt term-topic ma-
proach of audio thumbnailing (Jiang and Müller, trix, we compute a topic vector t for each new doc-
2015), the authors generate a Double Thumbnail ument (song text). In order to treat t as a (pseudo-
from a musical piece by finding the two most rep- ) probability distribution overP latent topics ti , we
resentative parts in it. For this, they search for normalize t by applying λt.t/ ti ∈t ti to it. Given
candidate musical segments in an a priori unseg- the distributions over latent topics for each song
mented song. Candidate musical segments are de- text, we then incrementally construct a summary
fined as sequences of music that more or less ex- by greedily adding one line from the original text
actly repeat themselves. The representativeness of at a time (same mechanism as in KLSum algo-
each candidate segment to the whole piece is then rithm in (Haghighi and Vanderwende, 2009)); that
estimated by their fitness metric. They define the line x∗ of the original text that minimizes the dis-
fitness of a segment as a trade-off between how ex- tance between the topic distribution tS of the orig-
actly a part is repeated and how much of the whole inal text S and the topic distribution of the incre-
piece is covered by all repetitions of that segment. mental summary sum(S):
Then, the audio segments along with their fitness
allow them to create an audio double thumbnail
x∗ = argmin {W (tS , tsum(S)+x )}
consisting of the two fittest audio segments. x∈(S\sum(S))
3 Lyrics Summarization
W is the Wasserstein distance (Villani, 2008)
Song texts are arranged in segments and lines. For and is used to measure the distance between two
instance the song text depicted in Figure 1 con- probability distributions (an alternative to Jensen-
sists of 8 segments and 38 lines. Given a song text Shannon divergence (Louis and Nenkova, 2013)).
330
Figure 1: Song text of “Let’s start a band” by Amy MacDonald along with two example summaries.
331
they find to capture the “hook” of the music more with unique text lines as input (i.e. rendundant
likely.2 lines are deleted). This is done to produce less re-
dundant summaries, given that for instance, Tex-
4 Experimental Setting tRank scores each duplicate line the same, hence
We now describe the WASABI dataset of song it may create summaries with all identical lines.
lyrics (Section 4.1), and the tested configurations TopSum can suffer from a similar shortcoming: if
of the summarization methods (Section 4.2). there is a duplicate line close to the ideal topic dis-
tribution, adding that line again will let the incre-
4.1 Dataset mental summary under construction stay close to
From the WASABI corpus (Meseguer-Brocal the ideal topic distribution. All models were in-
et al., 2017) we select a subset of 190k unique song structed to produce summaries of 4 lines, as this
texts with available genre information. As the cor- is the estimated median chorus length in our cor-
pus has spurious genres (416 different ones), we pus (see Section 4.1). The summary lines were
focus on the 10 most frequent ones in order to arranged in the same order they appear in the orig-
evaluate our methods dependent on the genre. We inal text.4 We use the TextRank implementation5
add 2 additional genres from the underrepresented of (Barrios et al., 2016) without removing stop
Rap field (Southern Hip Hop and Gangsta Rap). words (lyrics lines in input can be quite short,
The dataset contains 95k song lyrics. therefore we avoid losing all content of the line if
To define the length of sum(S) (see Section removing stop words). The topic model for Top-
3), we rely on (Bartsch and Wakefield, 2005) that Sum is built using non-negative matrix factoriza-
recommend to create audio thumbnails of the me- tion with scikit-learn6 (Pedregosa et al., 2011) for
dian length of the chorus on the whole corpus. 30 topics on the full corpus of 190k lyrics.7 For
We therefore estimate the median chorus length on the topical distance, we only consider the distance
our corpus by computing a Lyrics Single Thumb- between the 3 most relevant topics in the original
nail on each text, and we find the median chorus text, following the intuition that one song text usu-
length to be 4 lines. Hence, we decide to gener- ally covers only a small amount of topics. The
ate summaries of such length for all lyrics and all Lyrics Thumbnail is computed using String-based
summarization models to exclude the length bias distance between text segments to facilitate clus-
in the methods comparison3 . As the length of the tering. This similarity has been shown in (Watan-
lyrics thumbnail is lower-bounded by the length abe et al., 2016) to indicate segment borders suc-
of the chorus in the song text, we keep only those cessfully. In our implementation, segments are
lyrics with an estimated chorus length of at least 4. clustered using the DBSCAN (Ester et al., 1996)
The final corpus of 12 genres consists of 50k lyrics algorithm.8 We also produce two summaries by
with the following genre distribution: Rock: 8.4k, combining TextRank + TopSum and TextRank +
Country: 8.3k, Alternative Rock: 6.6k, Pop: 6.9k, TopSum + Lyrics Thumbnail, to test if summaries
R&B: 5.2k, Indie Rock: 4.4k, Hip Hop: 4.2k, can benefit from the complementary perspectives
Hard Rock: 2.4k, Punk Rock: 2k, Folk: 1.7k, the three different summarization methods take.
Southern Hip Hop: 281, Gangsta Rap: 185. Model Combination For any lyrics line, we can
4.2 Models and Configurations obtain a score from each of the applied methods.
TextRank provides a score for each line, TopSum
We create summaries using the three summariza- provides a distance between the topic distributions
tion methods described in Section 3, i.e. a graph- of an incremental summary and the original text,
based (TextRank), a topic-based (TopSum), and and f it provides the fitness of each line. We treat
fitness-based (Lyrics Thumbnail) method, plus our summarization methods as blackboxes and use
two additional combined models (described be- a simple method to combine the scores the dif-
low). While the Lyrics Thumbnail is generated ferent methods provide for each line. Given the
from the full segment structure of the lyrics in-
4
cluding its duplicate lines, all other models are fed In case of repeated parts, the first position of each line
was used as original position.
5
2
They also experiment with other methods to create a https://github.com/summanlp/textrank
6
thumbnail, such as section initial or section ending. https://scikit-learn.org
3 7
We leave the study of other measures to estimate the loss=’kullback-leibler’
8
summary length to future work. eps=0.3, min samples=2
332
original text separated into lines S = (x1 , ..., xn ), questionnaire and setting up the experiment. 26
a summary is constructed by greedily adding one participants - 12 nationalities, 18 men, 8 women,
line x∗ at a time to the incremental summary aged from 21 to 59 - were taking a questionnaire
sum(S) ⊆ S such that the sum of normalized (Google Forms), consisting of rating 30 items with
ranks of all scores is minimal: respect to the criteria defined before on a Likert
[ X scale from 1 (low) to 5 (high). Each participant
x∗ = argmin { RA (x)} was presented with 5 different summaries - each
x A
produced by one of the previously described sum-
Here x ∈ (S \ sum(S)) and A ∈ marization models - for 6 different song texts. Par-
{TextRank, TopSum, fit}. The normalized ticipants were given example ratings for the dif-
rank RA (x) of the score that method A assigns to ferent criteria in order to familiarize them with the
line x is computed as follows: first, the highest procedure. Then, for each song text, the original
scores9 are assigned rank 0, the second highest song text along with its 5 summaries were pre-
scores get rank 1, and so forth. Then the ranks are sented in random order and had to be rated ac-
linearly scaled to the [0,1] interval, so each sum cording to the above criteria. For the criterion of
P
of ranks A RA (x) is in [0,3]. Meaning, we asked participants to give a short ex-
planation in free text for their score. The selected
Model Nomenclature For abbreviation, we call
6 song texts10 have a minimum and a median cho-
the TextRank model henceforth Mr , the Top-
rus length of 4 lines and are from different genres,
Sum model Ms , the fitness-based summarizer Mf ,
i.e. Pop/Rock (4), Folk (1) and Rap (1), similar
model combinations Mrs and Mrsf , respectively.
to our corpus genre distribution. Song texts were
5 Evaluation selected from different lengths (18-63 lines), gen-
ders of singer (3 male, 3 female), topics (family,
We evaluate the quality of the produced lyrics life, drugs, relationship, depression), and mood
summary both soliciting human judgments on the (depressive, angry, hopeful, optimistic, energetic).
goodness and utility of a given summary (Section The artist name and song title were not shown to
5.1), and through an automatic evaluation of the the participants.
summarization methods (Section 5.2) to provide a
comprehensive evaluation. Results Figure 2 shows the ratings obtained for
each criterion. We examine the significant dif-
5.1 Human Evaluation ferences between the models performances by
We performed human evaluation of the different performing a paired two-tailed t-test. The sig-
summarization methods introduced before by ask- nificance levels are: 0.05∗ , 0.01∗∗ , 0.001∗∗∗ , and
ing participants to rate the different summaries n.s. First, Informativeness and Meaning are rated
presented to them by specifying their agreement / higher∗∗ for the combined model Mrs compared
disagreement according to the following standard to the single models Mr and Ms . Combining
criteria (Parveen and Strube, 2015): all three models improves the summaries further:
Informativeness: The summary contains the both for Informativeness and Meaning the model
main points of the original song text. Mrsf is rated higher∗∗∗ than Mrs . Further, sum-
Non-redundancy: The summary does not con- maries created by Mrsf are rated higher∗∗∗ in Co-
tain duplicate or redundant information. herence than summaries from any other model -
Coherence: The summary is fluent to read and except from Mf (n.s. difference). Summaries
grammatically correct. are rated on the same level (n.s. differences) for
Plus one additional criterion coming from our def- Non-redundancy in all but the Mr and Mf sum-
inition of the lyrics summarization task: maries, which are perceived as lower∗∗∗ in Non-
Meaning: The summary preserves the meaning redundancy than all others. Note, how the model
of the original song text. Mrsf is more stable than all others by exhibit-
ing lower standard deviations in all criteria except
An experimental psychologist expert in Human
10
Computer Interaction advised us in defining the “Pills N Potions” by Nicki Minaj, “Hurt” by Nine Inch
Nails, “Real to me” by Brian McFadden, “Somebody That I
9
In the case of topical distance, a “higher score” means a Used To Know” by Gotye, “Receive” by Alanis Morissette,
lower value. “Let’s Start A Band” by Amy MacDonald
333
Figure 2: Human ratings per summarization model in terms of average and standard deviation.
Non-redundancy. The criteria Informativeness and Topical: similarity between the topic distribu-
Meaning are highly correlated (Pearson correla- tions of original and summary. Restricted to the
tion coefficient 0.84). Correlations between other 3 most relevant topics of the original song text.
criteria range between 0.29 and 0.51. We give results relative to the similarity of the best
Overall, leveraging the Lyrics Fitness in a song performing model (=100%).
text summary improves summary quality. Espe- Coherence: average similarity between word
cially with respect to the criteria that, we believe, distributions in consecutive sentences of the sum-
indicate the summary quality the most - Informa- mary, cf. (ShafieiBavani et al., 2018). We give re-
tiveness and Meaning - the Mrsf method is signif- sults relative to the coherence of the original song
icantly better performing and more consistent. text (=100%).
Figure 1 shows an example song text and exam- Lyrics fitness: average line-based fitness f it (cf.
ple summaries from the experiment. Summary 1 is Section 3) of the lines in the summary. We give
generated by Mf and consists of the chorus. Sum- results relative to the Lyrics fitness of the original
mary 2 is made by the method Mrsf and has rel- song text (=100%).
evant parts of the verses and the chorus, and was
rated much higher in Informativeness and Mean- Results When evaluating each of the 12 genres,
ing. We analyzed the free text written by the par- we found two clusters of genres to behave very
ticipants to comment on the Meaning criterion, but similarly. Therefore, we report the results for these
no relevant additional information was provided two groups: the Rap genre cluster contains Hip
(the participants mainly summarized their ratings). Hop, Southern Hip Hop, and Gangsta Rap. The
Rock / Pop cluster contains the 9 other genres.
5.2 Automatic Evaluation Results of the different automatic evaluation met-
rics are shown in Table 1. Distributional Seman-
We computed four different indicators of summary
tics metrics have previously been shown (Louis
quality on the dataset of 50k songs described in
and Nenkova, 2013; ShafieiBavani et al., 2018)
Section 4.1. Three of the criteria use the similar-
to highly correlate with user responsiveness judg-
ity between probability distributions P, Q, which
ments. We would expect correlations of this met-
means we compute the Wasserstein distance be-
ric with Informativeness or Meaning criteria there-
tween P and Q (cf. Section 3.1) and apply
fore, as those criteria are closest to responsiveness,
λx. x−1 to it.11 The criteria are:
but we have found no large differences between
Distributional Semantics: similarity between
the different models for this criterion. The sum-
the word distributions of original and summary, cf.
maries of the Ms model have the highest similar-
(Louis and Nenkova, 2013). We give results rela-
ity to the original text and the Mf have the low-
tive to the similarity of the best performing model
est similarity of 90%. The difference between the
(=100%).
highest and lowest values are low.
11
This works as we always deal with distances > 0. For the Topical similarity, the results are mostly
334
Evaluation criterion Genre Mr Ms Mrs Mf Mrsf original text
Rock / Pop 92 100 97 90 93
Distributional
Rap
P 94 100 99 86 92 n/a
Semantics [%]
92 100 98 90 93
Rock / Pop 44 100 76 41 64
Topical [%] Rap
P 58 100 80 48 66 n/a
46 100 77 42 64
Rock / Pop 110 95 99 99 100
Coherence [%] Rap
P 112 115 112 107 107 100
110 97 101 100 101
Rock / Pop 71 53 63 201 183
Lyrics
Rap
P 0 0 0 309 249 100
fitness [%]
62 47 55 214 191
Table 1: Automatic evaluation results for the 5 summarization models and 2 genre clusters. Distributional Seman-
tics and Topical are relative to the best model (=100%), Coherence and Fitness to the original text (=100%).
in the same order as the Distributional Semantics strikes a balance between Topical similarity (64%)
ones, but with much larger differences. While the and fitness (191%). Hence, Mrsf succeeds in cap-
Ms model reaches the highest similarity, this is a turing lines from the most relevant parts of the
self-fulfilling prophecy, as summaries of Ms were lyrics, such as the chorus, while jointly represent-
generated with the objective of maximizing topical ing the important topics of the song text.
similarity. The other two models that incorporate
Ms (Mrs and Mrsf ), show a much higher topical
similarity to the original text than Mr and Mf .
6 Conclusion
Coherence is rated best in Mr with 110%. All
In this paper we have defined and addressed
other models show a coherence close to that of the
the task of lyrics summarization. We have ap-
original text - between 97% and 101%. We be-
plied both generic unsupervised text summariza-
lieve that the increased coherence of Mr is not lin-
tion methods (TextRank and a topic-based method
guistically founded, but merely algorithmic. Mr
we called TopSum), and a method inspired by au-
produces summaries of the most central sentences
dio thumbnailing on 50k lyrics from the WASABI
in a text. The centrality is using the concept of
corpus. We have carried out an automatic evalua-
sentence similarity. Therefore, Mr implicitly op-
tion on the produced summaries computing stan-
timizes for the automatic evaluation metric of co-
dard metrics in text summarization, and a human
herence, based on similar consecutive sentences.
evaluation with 26 participants, showing that using
Sentence similarity seems to be insufficient to pre-
a fitness measure transferred from the musicology
dict human judgments of coherence in this case.
literature, we can amend generic text summariza-
As might be expected, methods explicitly in- tion algorithms and produce better summaries.
corporating the Lyrics fitness produce summaries
In future work, we will model the importance
with a fitness much higher than the original text -
of a line given the segment to avoid cutting off im-
214% for the Mf and 191% for the Mrsf model.
portant parts of the chorus, as we sometimes ob-
The methods not incorporating fitness produce
served. Moreover, we plan to address the chal-
summaries with much lower fitness than the origi-
lenging task of abstractive summarization over
nal - Mr 62%, Ms 47%, and Mrs 55%. In the Rap
song lyrics, with the goal of creating a summary
genre this fitness is even zero, i.e. summaries (in
of song texts in prose-style - more similar to what
median) contain no part of the chorus.
humans would do, using their own words.
Overall, no single automatic evaluation crite-
rion was able to explain the judgments of our
human participants. However, considering Topi- Acknowledgement
cal similarity and fitness together gives us a hint.
The model Mf has high fitness (214%), but low This work is partly funded by the French Research
Topical similarity (42%). The Ms model has the National Agency (ANR) under the WASABI
highest Topical similarity (100%), but low fitness project (contract ANR-16-CE23-0017-01).
(47%). Mrsf might be preferred by humans as it
335
References Mohamed Abdel Fattah and Fuji Ren. 2009. Ga, mr,
ffnn, pnn and gmm based models for automatic text
Mehdi Allahyari, Seyed Amin Pouriyeh, Mehdi Assefi, summarization. Comput. Speech Lang., 23(1):126–
Saeid Safaei, Elizabeth D. Trippe, Juan B. Gutier- 144.
rez, and Krys Kochut. 2017. Text summarization
techniques: A brief survey. CoRR, abs/1707.02268. Aria Haghighi and Lucy Vanderwende. 2009. Explor-
ing content models for multi-document summariza-
Rachit Arora and Balaraman Ravindran. 2008. Latent tion. In Proceedings of Human Language Tech-
dirichlet allocation based multi-document summa- nologies: The 2009 Annual Conference of the North
rization. In Proceedings of the second workshop on American Chapter of the Association for Computa-
Analytics for noisy unstructured text data, pages 91– tional Linguistics, pages 362–370. Association for
97. ACM. Computational Linguistics.
Federico Barrios, Federico López, Luis Argerich, and Ruifang He and Xingyi Duan. 2018. Twitter summa-
Rosa Wachenchauzer. 2016. Variations of the simi- rization based on social network and sparse recon-
larity function of textrank for automated summariza- struction. In AAAI.
tion. CoRR, abs/1602.03606.
Leonhard Hennig. 2009. Topic-based multi-document
Mark A. Bartsch and Gregory H. Wakefield. 2005. Au- summarization with probabilistic latent semantic
dio thumbnailing of popular music using chroma- analysis. In Proceedings of the International Con-
based representations. Trans. Multi., 7(1):96–104. ference RANLP-2009, pages 144–149.
Taylor Berg-Kirkpatrick, Dan Gillick, and Dan Klein. Meishan Hu, Aixin Sun, Ee-Peng Lim, and Ee-Peng
2011. Jointly learning to extract and compress. In Lim. 2007. Comments-oriented blog summarization
Proceedings of the 49th Annual Meeting of the Asso- by sentence extraction. In Proceedings of the Six-
ciation for Computational Linguistics: Human Lan- teenth ACM Conference on Conference on Informa-
guage Technologies - Volume 1, HLT ’11, pages tion and Knowledge Management, CIKM ’07, pages
481–490, Stroudsburg, PA, USA. Association for 901–904, New York, NY, USA. ACM.
Computational Linguistics.
Nanzhu Jiang and Meinard Müller. 2015. Estimat-
David Brackett. 1995. Interpreting Popular Music. ing double thumbnails for music recordings. In
Cambridge University Press. 2015 IEEE International Conference on Acous-
tics, Speech and Signal Processing (ICASSP), pages
Wei Chai and Barry Vercoe. 2003. Music thumbnail- 146–150.
ing via structural analysis. In Proceedings of the
eleventh ACM international conference on Multime- Florian Kleedorfer, Peter Knees, and Tim Pohle. 2008.
dia, pages 223–226. Oh oh oh whoah! towards automatic topic detection
in song lyrics. In Ismir, pages 287–292.
Jianpeng Cheng and Mirella Lapata. 2016. Neural
Kevin Knight and Daniel Marcu. 2000. Statistics-
summarization by extracting sentences and words.
based summarization - step one: Sentence compres-
In Proceedings of the 54th Annual Meeting of the As-
sion. In Proceedings of the Seventeenth National
sociation for Computational Linguistics (Volume 1:
Conference on Artificial Intelligence and Twelfth
Long Papers), pages 484–494. Association for Com-
Conference on Innovative Applications of Artificial
putational Linguistics.
Intelligence, pages 703–710. AAAI Press.
Jean Yves Delort, Bernadette Bouchon-Meunier, and Mark Levy, Mark Sandler, and Michael Casey. 2006.
Maria Rifqi. 2003. Enhanced web document sum- Extraction of high-level musical structure from au-
marization using hyperlinks. In Proceedings of the dio data and its application to thumbnail generation.
Fourteenth ACM Conference on Hypertext and Hy- In 2006 IEEE International Conference on Acous-
permedia, HYPERTEXT ’03, pages 208–215, New tics Speech and Signal Processing Proceedings, vol-
York, NY, USA. ACM. ume 5, pages V–V. IEEE.
Günes Erkan and Dragomir R Radev. 2004. Lexrank: Chin-Yew Lin. 2004. Rouge: A package for automatic
Graph-based lexical centrality as salience in text evaluation of summaries. In Text Summarization
summarization. Journal of artificial intelligence re- Branches Out.
search, 22:457–479.
Annie Louis and Ani Nenkova. 2013. Automatically
Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei assessing machine summary content without a gold
Xu, et al. 1996. A density-based algorithm for standard. Computational Linguistics, 39(2).
discovering clusters in large spatial databases with
noise. In Kdd, volume 96, pages 226–231. Stuart Mackie, Richard McCreadie, Craig Macdonald,
and Iadh Ounis. 2014. On choosing an effective au-
Mohamed Abdel Fattah. 2014. A hybrid machine tomatic evaluation metric for microblog summarisa-
learning model for multi-document summarization. tion. In Proceedings of the 5th Information Interac-
Applied intelligence, 40(4):592–600. tion in Context Symposium, pages 115–124. ACM.
336
Qiaozhu Mei and ChengXiang Zhai. 2008. Generating Scikit-learn: Machine learning in Python. Journal
impact-based summaries for scientific literature. In of Machine Learning Research, 12:2825–2830.
ACL.
Horacio Saggion and Thierry Poibeau. 2013. Auto-
Gabriel Meseguer-Brocal, Geoffroy Peeters, Guil- matic Text Summarization: Past, Present and Fu-
laume Pellerin, Michel Buffa, Elena Cabrio, Cather- ture, pages 3–21. Springer, Berlin, Heidelberg.
ine Faron Zucker, Alain Giboin, Isabelle Mirbel, Ro-
main Hennequin, Manuel Moussallam, Francesco Elaheh ShafieiBavani, Mohammad Ebrahimi, Ray-
Piccoli, and Thomas Fillon. 2017. WASABI: a Two mond Wong, and Fang Chen. 2018. Summariza-
Million Song Database Project with Audio and Cul- tion evaluation in the absence of human model sum-
tural Metadata plus WebAudio enhanced Client Ap- maries using the compositionality of word embed-
plications. In Web Audio Conference 2017 – Col- dings. In Proceedings of the 27th International Con-
laborative Audio #WAC2017, London, United King- ference on Computational Linguistics, pages 905–
dom. Queen Mary University of London. 914. Association for Computational Linguistics.
Rada Mihalcea and Paul Tarau. 2004. Textrank: Bring- Cédric Villani. 2008. Optimal transport: old and new,
ing order into text. In Proceedings of the 2004 Con- volume 338. Springer Science & Business Media.
ference on Empirical Methods in Natural Language
Processing. Lu Wang, Hema Raghavan, Vittorio Castelli, Radu
Florian, and Claire Cardie. 2016. A sentence
Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2016. compression based framework to query-focused
Summarunner: A recurrent neural network based se- multi-document summarization. arXiv preprint
quence model for extractive summarization of docu- arXiv:1606.07548.
ments. In AAAI.
Kento Watanabe, Yuichiroh Matsubayashi, Naho Orita,
Ani Nenkova, Kathleen McKeown, et al. 2011. Auto- Naoaki Okazaki, Kentaro Inui, Satoru Fukayama,
matic summarization. Foundations and Trends
R in Tomoyasu Nakano, Jordan Smith, and Masataka
Information Retrieval, 5(2–3):103–233. Goto. 2016. Modeling discourse segments in lyrics
using repeated patterns. In Proceedings of COLING
Jahna Otterbacher, Güneş Erkan, and Dragomir R 2016, the 26th International Conference on Compu-
Radev. 2005. Using random walks for question- tational Linguistics: Technical Papers, pages 1959–
focused sentence retrieval. In Proceedings of the 1969.
conference on Human Language Technology and
Empirical Methods in Natural Language Process-
ing, pages 915–922. Association for Computational
Linguistics.
Lawrence Page, Sergey Brin, Rajeev Motwani, and
Terry Winograd. 1999. The pagerank citation rank-
ing: Bringing order to the web. Technical report,
Stanford InfoLab.
Daraksha Parveen, Hans-Martin Ramsl, and Michael
Strube. 2015. Topical coherence for graph-based ex-
tractive summarization. In Proceedings of the 2015
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 1949–1954.
Daraksha Parveen and Michael Strube. 2015. Inte-
grating importance, non-redundancy and coherence
in graph-based extractive summarization. In Pro-
ceedings of the 24th International Conference on
Artificial Intelligence, IJCAI’15, pages 1298–1304.
AAAI Press.
Samuel Pecar. 2018. Towards opinion summarization
of customer reviews. In Proceedings of ACL 2018,
Student Research Workshop, pages 1–8. Association
for Computational Linguistics.
Fabian Pedregosa, Gael Varoquaux, Alexandre Gram-
fort, Vincent Michel, Bertrand Thirion, Olivier
Grisel, Mathieu Blondel, Peter Prettenhofer, Ron
Weiss, Vincent Dubourg, Jake Vanderplas, Alexan-
dre Passos, David Cournapeau, Mathieu Brucher,
Mathieu Perrot, and Edouard Duchesnay. 2011.
337
Comparing Automated Methods to Detect Explicit Content in Song Lyrics
338
Proceedings of Recent Advances in Natural Language Processing, pages 338–344,
Varna, Bulgaria, Sep 2–4 2019.
available related works and shed light on the inher- for clean lyrics. They compare different methods
ent hardness and subjectivity of the task at hand. to generate such a lexicon. The achieved perfor-
The paper is organized as follows: in Section 2 mances using solely dictionary lookup range from
we survey the state of the art in explicit lyrics de- 49% F1 for a man-made dictionary to 75.6% F1
tection. In Sections 3 and 4 we introduce the clas- when using relative class frequencies. Note, that
sification methods we apply, and the comparative the latter performance is achieved with a dictio-
experimentation. Conclusions end the paper. nary of only 25 words. They work with a corpus
NOTE: This paper contains examples of lan- of Korean lyrics (see Figure 1, K19). Unlike pre-
guage which may be offensive to some readers. vious work, they apply a recursive neural network,
They do not represent the views of the authors. resulting in 76.6% F1 , slightly higher than the sim-
ple dictionary lookup. They find performance to
2 Related Work increase to 78.1% when combining the vector rep-
Only a few works on the problem of explicit lyrics resentation of the RNN with a one-hot vector indi-
detection exist. (Bergelid, 2018) consider a dataset cating for each profane word from the dictionary if
of English lyrics (see Table 1, B18) to which they the lyric contains it. They argue to use the RNN to
apply classical machine learning algorithms such find such cases where the expliciteness arises from
as Support Vector Machine (SVM) and Random the context and not from a dictionary check. How-
Forest (RF). As features they extract either (i) tf- ever, no examples of finding this phenomenon are
idf weighted bag-of-word (BOW) representations presented.
of each song text or (ii) represent the lyrics with
3 Methods for Explicit Lyrics Detection
paragraph vectors (Le and Mikolov, 2014). The
explicit labels are obtained from Soundtrack Your In this work, we compare a range of classification
Brand 4 . They find the RF with tf-idf BOW to per- methods for the task of explicit lyrics detection.
form best, especially in combination with a ran- Common to all methods is that they classify a full
dom undersampling strategy to the highly imbal- song into one of two mutually exclusive classes -
anced dataset. They also experiment with adding explicit or clean (=not explicit). This means, the
lyrics metadata to the feature set, such as the artist decision if a song text is explicit is taken glob-
name, the release year, the music energy level, and ally. We assess the performance of different clas-
the valence/positiveness of a song. This results in sification methods ranging from simple dictionary
marginal improvements for some of their models. lookup / lexicon checking to general purpose deep
(Chin et al., 2018) apply explicit lyrics detec- learning language understanding models. We try
tion to Korean song texts. They also use tf-idf to identify contextual effects by applying a method
weighted BOW as lyrics representation and aggre- that outputs the “importance” for each word (see
gate multiple decision trees via boosting and bag- Section 3.4).
ging to classify the lyrics for explicit content. On
their corpus (see Figure 1, C18) they report 78% 3.1 Dictionary-Based Methods
F1 using the bagging method. Note, that bagging The most straightforward way to implement an
with decision trees is similar to the Random Forest automated explicit content detection method, is
method used by (Bergelid, 2018). Interestingly, checking against a dictionary of explicit words.
they also report a baseline for dictionary lookup, The dictionary can be man-made or automatically
i.e. given a profanity dictionary the song text is created from example explicit and clean lyrics.
classified as explicit if and only if one of its words Then, a classifier uses this dictionary to predict the
occurs in the profanity dictionary. With such a class of an unseen song text.
baseline they obtain 61% F1 .
More recently, (Kim and Mun, 2019) proposed 3.1.1 Dictionary Creation
a method to create explicit words dictionaries au- It is possible to use handcrafted dictionaries such
tomatically by weighting a vocabulary according as Noswearing 5 . Performance using an automat-
to all words’ frequencies in the explicit class vs. ically created lexicon has previously been shown
the clean class, accordingly. For instance the word (Kim and Mun, 2019) to improve over the manu-
“fuck” is typical for explicit lyrics and atypical ally created dictionary. We therefore consider only
4 5
https://www.soundtrackyourbrand.com https://www.noswearing.com/
339
the case of the machine-made dictionary in this One of the most successful transformer-based
work. We generate a dictionary of words that are models proposed in the last few years is BERT
indicative of explicit lyrics. We define the impor- (Devlin et al., 2018). This model is composed of
tance I of a word w for explicit lyrics by the fre- multiple transformers connected by residual con-
quency f (w, ex) of w in explicit lyrics compared nections. Pre-trained models are provided by the
to its frequency f (w, cl) in clean lyrics: authors, and they are used in our work to perform
I(w) = f (w, ex)/f (w, cl) explicit language detection in lyrics, without re-
training the full model.
We filter out unique and too common words and
restrict the number of terms to 1,000 to avoid over- 3.4 Textual Deconvolution Saliency
reliance on terms that are very corpus specific. We use the Textual Deconvolution Saliency (TDS)
The dictionary Dn of the n words most impor- model of (Vanni et al., 2018), which is a Convolu-
tant for explicit lyrics, is now straightforwardly tional Neural Network (CNN) for text classifica-
defined as containing the n words with the high- tion. It is a simple model containing an embedding
est I score. layer for word representations, a convolutional
3.1.2 Dictionary Lookup layer with max pooling and two fully connected
Given a dictionary Dn , this method simply checks layers. The interesting part about this model is that
if a song text S contains any of the explicit terms they manage to reverse the convolution. Given the
defined in Dn . Then, S is classified as explicit iff learned feature map (the output of the convolution
it contains at least one explicit term from Dn . before max pooling) of the CNN, they upsample it
to obtain a 3-dimensional sample with dimensions
3.1.3 Dictionary Regression (#words, embedding size, #filters). The TDS for
This method uses BOW made from Dn as the each word is now defined as the sum along the
feature set of a classifier. We used a logistic re- embedding axes of the output of the deconvolu-
gression, but RF or SVM have been used alike in tion. The TDS represents the importance of each
(Bergelid, 2018). word of the input with respect to the learned fea-
ture maps. We use this model with the goal to find
3.2 Tf-idf BOW Regression
local explanations for the global decision of the
Similar to the Dictionary Regression, but the classification as explicit or clean. Such explana-
BOW contains the whole vocabulary of a train- tions can arise from contexts or phrases that the
ing sample instead of only the explicit terms. The model assigns a high importance.
word features are weighted with the well-known
tf-idf weighting scheme. 4 Experimental Setting and Evaluation
3.3 Transformer Language Model We compare the different methods as introduced
Recently, approaches based on self-attention in the previous section to the task of explicit lyrics
(Vaswani et al., 2017) have been proposed and detection. We attempt a comparison to the related
have proven effective for natural language under- work as well, although due to different datasets
standing tasks. These models are structured as an comparing the reported scores directly is problem-
encoder-decoder, and they are trained on unsuper- atic. We finally analyze the classification qualita-
vised tasks (such as masked language modelling) tively with examples, and demonstrate the intrin-
in order to learn dense representations of sentences sic hardness and subjectivity of the explicit lyrics
or documents. These models differ from more tra- detection task.
ditional recurrent neural networks in different as- Abbreviations used: to refer to related works in
pects. In particular, while recurrent models can Table 1 and 3, we use the following abbrevia-
process sequences (in NLP, typically word embed- tions. B18 stands for (Bergelid, 2018), C18 is
dings) in order, transformers use a joint model of (Chin et al., 2018), K19 means (Kim and Mun,
the right and left context of each word in order 2019), while Ours is this work.
to encode an entire sequence or document. Ad-
ditionally, transformers are typically less compu- 4.1 Dataset
tationally expensive than recurrent models, espe- The WASABI database (Meseguer-Brocal et al.,
cially when trained on a GPU accelerator. 2017) contains song-wise labels for explicit lyrics,
340
Work total explicit ratio language 4.3 Results
B18 25,441 3,310 13.0% English
C18 27,695 1,024 3.7% Korean Overall, the results of the different classification
K19 70,077 7,468 10.7% Korean methods we tried are all close to each other. The
WAS 179,391 17,808 9.9% English simple dictionary lookup with 32 words performs
comparably to the deep neural network with 110M
Table 1: Overview of our dataset WAS (# songs) and parameters (BERT base model). As baseline, we
comparison to the related works. include the majority class classifier that always
predicts the clean class. Furthermore, all related
works show similar tendencies of performance on
such as explicit, unknown, no advice available, or
their respective datasets. The results of all the dif-
clean (=not explicit). These labels are provided by
ferent methods we applied are depicted in Table 2
the music streaming service Deezer 6 . We selected
and described in the following.
a subset of English song texts from the corpus
The majority class classifier delivers a perfor-
which are tagged as either explicit or clean. We
mance of 47.4% F1 , which is the only outlier in the
filtered out duplicate lyrics and such that contain
sense that this is far below any other model. The
less than 10 tokens. Finally, our dataset (WAS)
dictionary lookup with a vocabulary of the 32 most
comprises of 179k lyrics, with a ratio of explicit
indicative explicit words obtains a balanced per-
lyrics of 9.9%. The details and comparison with
formance as precision and recall are close to each
related works datasets are depicted in Table 1.
other, the overall performance is 77.3% F1 . The
For training any of the models described in the
dictionary regression performs somewhat better in
previous section, we once randomly split the data
terms of f-score (78.5% F1 ), achieving this with
into training-development-test sets with the com-
the highest overall recall of 81.5%, but it has lower
mon 60%-20%-20% ratio. We tuned the hyperpa-
precision. The tf-idf BOW regression performs
rameters of the different classification algorithms
very similarly to the dictionary regression. This
on the development set to then test with the best
proves that a limited number of words influences
performing parameters on the test set. As evalua-
the overall performance of the models, and that
tion metrics we use precision (P), recall (R), and
they do not need to consider the whole vocabulary,
f-score (F1 ). Unless stated otherwise, the scores
just the most offensive words. The increased vo-
are macro-averaged over the two possible classes.
cabulary of 929k unigrams and bigrams is gigan-
4.2 Hyperparameters tic compared to the explicit words dictionary (32
words). As most of these n-grams may be noise to
For the dictionary-based methods, we found the
the classifier, this could explain the slight decrease
ideal dictionary size to be 32 words for the lookup
in performance over the dictionary regression. Fi-
and 128 words for the regression. The Tf-idf BOW
nally, the neural-network-based methods behave
regression performed best when the full vocab-
a bit differently: the BERT language model is
ulary of unigrams and bigrams was used. We
clearly better in precision (84.4%) over all other
used the sklearn implementation of logistic regres-
models - the second best is TDS with 81.2%.
sion with the class weighting scheme ’balanced’
However, BERT performs the worst in recall with
to account for the class imbalance in the dataset.
only 73.7%. The overall performance of BERT is
We used TDS with max sequence length 512 and
average with 77.7% F1 . Finally, TDS performs
dropout probability 50%. As is the default with
best in terms of 79.6% F1 . We tested if TDS out-
TDS, corpus-specific word vectors were trained
performing BERT was due to TDS using domain-
using Word2Vec (Mikolov et al., 2013) with di-
specific word vectors trained on our corpus (BERT
mensionality 128. The BERT model comes pre-
is trained on books and Wikipedia). This was
trained and no further pre-training was performed.
not the case as TDS performed almost identically,
We used the smaller of the two published mod-
when using generic word vectors (GloVe, 200d):
els. BERT then was finetuned to our task using
80.4% P, 78.7% R, 79.5% F1 .
max sequence length 256 and batch size 16, other-
A closer look at the classification performance
wise default parameters for text classification task
shows that the F1 scores for the minority class (ex-
learning.
plicit lyrics) is highest with TDS (63%) and lowest
6
https://www.deezer.com with the dictionary lookup (58.9%). The majority
341
Model P R F1 more likely ones found in explicit terms (C18) im-
Majority Class 45.0 50.0 47.4 proves performance slighly. Using random under-
Dictionary Lookup 78.3 76.4 77.3 sampling to fight the imbalanced class problem
Dictionary Regression 76.2 81.5 78.5 (B18) increases performance drastically, however
Tf-idf BOW Regression 75.6 81.2 78.0 makes the problem somewhat different from the
TDS Deconvolution 81.2 78.2 79.6 imbalanced problem. The final takeaway is that
BERT Language Model 84.4 73.7 77.7 deep models do not necessarily outperform shal-
low models. Neither HAN, TDS, nor BERT de-
Table 2: Performances of our different models on the
liver much higher scores than the dictionary-based
WAS dataset. Values in percent.
or the BOW method.
342
as its meaning is not necessarily explicit. (3) Four riods and cultures comes from words themselves:
texts appeared to have been mislabelled since no an inoffensive word can become offensive in slang
explicitness could be found. We found for three or common language. One such example can
of them that the album the song is contained in is be found in Johnny Cash’s The Christmas Guest:
tagged as explicit. In cases as these, inheriting the “When the cock was crowing the night away - The
label from the album is wrong, but it seems this is Lord appeared in a dream to me”. Here, cock
exactly what had happened here. In one Raggae means male chicken, as opposed to the offensive
lyric, in particular, we found no explicit content, meaning that is now arguably more common.
so we suspect the song was mislabelled. We finally want to raise attention to the problem
Since we found some annotation to be problem- of genre confounding. We found that the genre
atic, we will discuss difficulties that arise from an- Hip Hop contributed by far the most to all ex-
notating explicitness in lyrics. plicit lyrics - 33% of all Hip Hop lyrics. Since
only about 5% of the whole corpus are tagged
4.4.2 How Hard is this Task?
as Hip Hop, this genre is highly overrepresented.
As stated in the introduction, the explicit label is This raises the question in how far our task is con-
voluntary and we will argue that it is also some- founded with genre classification. When inspect-
what subjective in its nature. There are lyrics ing the explicit terms dictionaries we have created,
which are not tagged as explicit although they have we clearly see that genre bias reflected. The dic-
profanity in them. Consider for example the song tionary of 32 terms that we used for the dictio-
Bitch by Meredith Brooks. While it already con- nary lookup method consists approximately half
tains profanity in the title, it does not carry the of terms that are quite specific to the Rap genre,
explicit label and one can argue that in the con- such as glock, gat, clip (gun-related), thug, beef,
text of the song, the term “bitch” is used as a con- gangsta, pimp, blunt (crime and drugs). Finally,
trastive term and to raise attention to the struggle the terms holla, homie, and rapper are arguably
the songwriter sees in her life, torn between po- no causes for explicit lyrics, but highly correlated
tentially conflicting expectations of society (“I’m with explicit content lyrics. Biasing an explicit
a little bit of everything - All rolled into one - I’m lyrics detection model away from genres is an in-
a bitch, I’m a lover - I’m a child, I’m a mother - teresting future direction of work.
I’m a sinner, I’m a saint - I do not feel ashamed”).
Another example is Check Your Head by 5 Conclusion
Buckcherry where it says “Ooh and you still bitch
Classifying song lyrics as explicit or clean is an
about your payments” where “bitch” is used as a
inherently hard task to accomplish since what is
verb and one can argue that the acceptance in this
considered offensive strongly depends on cultural
verb form is higher than in the noun form. A sim-
aspects that can change over time. We showed
ilar case where the part of speech influences the
that shallow models solely based on a dictionary
perceived level of profanity is Hail Hail Rock ’n’
of profane words achieve a performance compara-
Roll by Discipline. It contains the line “the band
ble to deep neural networks. We argued that even
starts to play loud as fuck”.
the hand-labelling is highly subjective, making it
We encounter a different kind of problem when
problematic to automatically detect if a song text
dealing with substance abuse or other drug-related
should be tagged as explicit or clean.
content. It is evident that the legal status of the
substances mentioned plays a major role in how We propose as a possible simplification and ob-
such content is labelled. This is further com- jectification to study the local detection of explicit
plicated by the fact that legislation about sub- content. If we present an authority a report on
stances can vary wildly between different coun- found trigger words, found contextual sexual con-
tries. The labels applied to this content are not tent, and alike, they can come to their own subjec-
culture-invariant, and furthermore changes in the tive conclusion about the final label of the text.
societal view can lead to labels that are not rele-
Acknowledgement
vant anymore. This, like other examples, shows
why the labels applied to lyrics are subject to This work is partly funded by the French Research
change in different cultures and time periods. National Agency (ANR) under the WASABI
Another aspect that is very sensitive to time pe- project (contract ANR-16-CE23-0017-01).
343
References Million Song Database Project with Audio and Cul-
tural Metadata plus WebAudio enhanced Client Ap-
Valerio Basile, Cristina Bosco, Elisabetta Fersini, Deb- plications. In Web Audio Conference 2017 – Col-
ora Nozza, Viviana Patti, Francisco Manuel Rangel laborative Audio #WAC2017, London, United King-
Pardo, Paolo Rosso, and Manuela Sanguinetti. 2019. dom. Queen Mary University of London.
Semeval-2019 task 5: Multilingual detection of hate
speech against immigrants and women in twitter. In Tomas Mikolov, Kai Chen, Greg Corrado, and Jef-
Proceedings of the 13th International Workshop on frey Dean. 2013. Efficient estimation of word
Semantic Evaluation, pages 54–63. representations in vector space. arXiv preprint
arXiv:1301.3781.
Linn Bergelid. 2018. Classification of explicit music
content using lyrics and music metadata. Ji Ho Park and Pascale Fung. 2017. One-step and two-
step classification for abusive language detection on
Cristina Bosco, Felice Dell’Orletta, Fabio Poletto, twitter. In Proceedings of the First Workshop on
Manuela Sanguinetti, and Maurizio Tesconi. 2018. Abusive Language Online, pages 41–45. Association
Overview of the EVALITA 2018 hate speech de- for Computational Linguistics.
tection task. In Proceedings of the Sixth Evalua- Laurent Vanni, Mélanie Ducoffe, Carlos Aguilar, Fred-
tion Campaign of Natural Language Processing and eric Precioso, and Damon Mayaffre. 2018. Textual
Speech Tools for Italian. Final Workshop (EVALITA deconvolution saliency (tds): a deep tool box for lin-
2018) co-located with the Fifth Italian Conference guistic analysis. In Proceedings of the 56th Annual
on Computational Linguistics (CLiC-it 2018), Turin, Meeting of the Association for Computational Lin-
Italy. guistics (Volume 1: Long Papers), pages 548–557.
Hyojin Chin, Jayong Kim, Yoonjong Kim, Jinseop Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Shin, and Mun Y Yi. 2018. Explicit content de- Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz
tection in music lyrics using machine learning. In Kaiser, and Illia Polosukhin. 2017. Attention is all
2018 IEEE International Conference on Big Data you need. In I. Guyon, U. V. Luxburg, S. Bengio,
and Smart Computing (BigComp), pages 517–521. H. Wallach, R. Fergus, S. Vishwanathan, and R. Gar-
IEEE. nett, editors, Advances in Neural Information Pro-
cessing Systems 30, pages 5998–6008. Curran As-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and sociates, Inc.
Kristina Toutanova. 2018. Bert: Pre-training of deep
bidirectional transformers for language understand- Michael Wiegand, Melanie Siegel, and Josef Ruppen-
ing. arXiv preprint arXiv:1810.04805. hofer. 2018. Overview of the germeval 2018 shared
task on the identification of offensive language. In
Elisabetta Fersini, Paolo Rosso, and Maria An- Proceedings of GermEval 2018, 14th Conference on
zovino. 2018. Overview of the task on auto- Natural Language Processing (KONVENS 2018).
matic misogyny identification at ibereval 2018. In
Marcos Zampieri, Shervin Malmasi, Preslav Nakov,
IberEval@SEPLN, volume 2150 of CEUR Work-
Sara Rosenthal, Noura Farra, and Ritesh Kumar.
shop Proceedings, pages 214–228. CEUR-WS.org.
2019. Semeval-2019 task 6: Identifying and cate-
gorizing offensive language in social media (offen-
Darja Fišer, Ruihong Huang, Vinodkumar Prab- seval). CoRR, abs/1903.08983.
hakaran, Rob Voigt, Zeerak Waseem, and Jacqueline
Wernimont. 2018. Proceedings of the 2nd workshop
on abusive language online (alw2). In Proceedings
of the 2nd Workshop on Abusive Language Online
(ALW2). Association for Computational Linguistics.
344
Linguistic Classification:
Dealing Jointly with Irrelevance and Inconsistency
Laura Franzoi Andrea Sgarro Anca Dinu Liviu P. Dinu
Faculty of Mathematics DMG Faculty of Faculty of Mathematics
and Computer Science University of Trieste Foreign Languages and Computer Science
University of Bucharest sgarro@units.it and Literatures University of Bucharest
laura.franzoi@ University of Bucharest liviu.p.dinu@
gmail.com ancaddinu@ gmail.com
gmail.com
Abstract
This paper is based on previous work on fuzzy
In this paper we present new methods for string distances and linguistic classification
language classification which put to good started in (Franzoi and Sgarro, 2017a,b; Franzoi,
use both syntax and fuzzy tools, and are 2017), and inspired by the path-breaking ideas
capable of dealing with irrelevant linguis- put forward back in 1967 (Muljačić, 1967) by the
tic features (i.e. features which should not Croat linguist Ž., Muljačić. The technical tool
contribute to the classification) and even which will be used in this paper is the general
inconsistent features (which do not make Steinhaus transform, or biotope transform, ap-
sense for specific languages). We intro- plied to crisp strings which are however affected
duce a metric distance, based on the gener- by irrelevance and inconsistency, as happens with
alized Steinhaus transform, which allows data due to the linguist G. Longobardi and his
one to deal jointly with irrelevance and in- school. Fuzziness in linguistics has been seldomly
consistency. To evaluate our methods, we treated (Franzoi and Sgarro, 2017a,b; Dinu et al.,
test them on a syntactic data set, due to 2018), as compared to crisp approaches.
the linguist G. Longobardi and his school.
We obtain phylogenetic trees which some- In his 1967 paper Muljačić, even if only rather
times outperform the ones obtained by implicitly, had introduced what appears to us as
Atkinson and Gray (Gray and Atkinson, a natural fuzzy generalization of crisp Hamming
2003; Bouckaert et al., 2012). distances between binary strings of fixed length
n, and this only two years after Zadeh’s seminal
1 Introduction
work (Zadeh, 1965): the aim was showing that
According to Ethnologue (Eth, 2018), there are Dalmatic, now an extinct language, is a bridge
around 7000 living natural languages in the world, between the Western group of Romance languages
and one of the most interesting topics (not only in and the Eastern group, mainly Romanian. The
the academic field, but also in the general public) situation is the following: Romance languages
is their classification. While the comparative L, Λ, . . . are each described by means of n fea-
method was the main method of classifying tures, which can be present or absent, and so are
natural languages until the 90s, the last decades encoded by string s(L) = x = xi . . . xn , where
brought an increasing number of computational xi is the truth value of the proposition feature i is
approaches for estimating the historical evolution present in language L; however, presence/absence
of languages and their relationships. Most of the is sometimes only vaguely defined and so each
computational historical linguistics approaches x = xi is rather a truth value x ∈ [0, 1] in a
rely on the use of lexical items. In contrast, multi-valued logic as is fuzzy logic; x = xi is
very few of them take into account syntactic crisp only when either x = 0 = false = absent or
aspects. Moreover, fuzzy tools and information x = 1 = true = present, else x is strictly fuzzy.
theory were employed quite sparsely in language So, the mathematical objects one deals with are
classification tasks (Ciobanu et al., 2018), in spite strings x, y, . . . of length n, each of the n compo-
the inherent fuzzy nature of the natural language nents being a real number in the interval [0, 1], and
data. moreover distances between such objects, since
345
Proceedings of Recent Advances in Natural Language Processing, pages 345–352,
Varna, Bulgaria, Sep 2–4 2019.
the classifications are all distance-based. In what tolussi et al., 2011; Longobardi et al., 2016); lan-
follows, rather than Muljačić distance, we need guages are represented through yes-no strings of
string distances obtained by use of the Steinhaus length 53, each string position corresponding to a
transform, cf. (Dinu et al., 2018), and the gen- syntactic feature which can be present or absent.
eralized Steinhaus transform; they are all metric In his notation Longobardi uses + if a feature is
distances, in particular they verify the triangle present, - if it is absent, 0 if it is undefined; in
equality. Unlike the case of Muljačić distances, our case, cf. Tables 1, 2, we write 1 if a feature
which span the interval [0, n], these distances is present, 0 if it is absent, * if it is undefined. Ac-
are normalized to the interval [0, 1]. Steinhaus tually, due to a complex network of logical impli-
transforms allow one to deal with irrelevance cations which constrain features, some positions
and inconsistency in linguistics, as we already might be undefined (logically inconsistent). For
argued in (Dinu et al., 2018), and not only with example, in Longobardi’s classification, feature 34
vagueness, or fuzziness, as in Muljačić case, cf. is defined if and only if feature 8 is set to + and ei-
(Muljačić, 1967; Franzoi and Sgarro, 2017a); the ther feature 9 is set to + or feature 18 is not set to
reason to use the generalized Steinhaus transform, + (or both); otherwise it will be “neutralized” (in-
as we do here, is that it allows one to deal jointly consistent)1 . This property does not hold true for
with both irrelevance and inconsistency. Ptg (Portuguese), OE (Old English) and Ice (Ice-
landic).
All this establishes an extremely complex network
Based on arguments defended by the linguist G.
of logical dependencies in Longobardi’s data, and
Longobardi and his school, cf. (Bortolussi et al.,
makes it necessary, if one wants to cover also this
2011; Longobardi et al., 2016, 2013, 2015), if a
new intriguing facet, to suitably generalize crisp
feature i has a low truth value in two languages
Hamming distances, or crisp Jaccard distances, re-
L and Λ, then that feature is scarcely relevant: in
spectively: in Longobardi’s approach, cf. (Borto-
fact, in the practice of linguistics the values 0 and
lussi et al., 2011; Longobardi et al., 2016, 2013,
1 have a very asymmetric use, and the fact that
2015), the two distances for ternary strings one de-
languages L and Λ both have zero in a position
fines and uses are quite useful, but unfortunately
i means that such an irrelevant feature i should
they violate the triangle property, and so are not
not really contribute to the distance between the
metric. In this paper we propose one metric alter-
two languages. Technically, one should move
native based on the generalized Steinhaus trans-
from Hamming distances to (normalized) Jaccard
form (or generalized biotope transform): the star
distances. To achieve the goal, the convenient tool
∗ will be replaced by the totally ambiguous truth
we have used was the Steinhaus transform, cf.
value 12 , and the pivot strings in the transform will
(Dinu et al., 2018), which is known to preserve 1
be given by the set compound by the all- string,
metricity and which is general enough so as to 2
amply cover also the fuzzy situation: one starts i.e. the totally ambiguous string z = 21 , . . . , 12
from a distance like Muljačić distance dM (x, y), (which stands for inconsistency) and all-0 string
and obtains its Steinhaus transform, in this case a z = (0, . . . , 0), i.e. the totally false string, which
fuzzy Jaccard distance dJ (x, y) for fuzzy strings 1
Feature 34 stands for checking possessives: it opposes
x and y; starting from the usual crisp Hamming languages like French, wherein possessives occur without
any visible article (mon livre vs. le mon livre), to those like
distance the transform gives the usual crisp
Italian, in which a visible determiner is possible and nor-
Jaccard distance. mally required instead (il mio libro vs. mio libro). This fea-
ture seems to conceptually and typologically depend on full
grammaticalization of definiteness (feature 8). Also, it is rel-
In general, to apply a Steinhaus transformation evant only in languages with strong Person in D (feature 9)
or without strong article (feature 18), because otherwise the
one needs a pivot string, which in the Jaccard case language would have GenS with determiner-like function, cf.
is the all-0 string z = 0 = (0, . . . , 0). In the (Longobardi et al., 2013). Feature 8 asks if a language gener-
transform, actually, any other string z might be alizes the overt marking of definiteness to all relevant cases.
Feature 9 (Strong Person) defines whether attraction to the
used, cf. (Dinu et al., 2018), as we do here so as to D area of referential nominal material (e.g. proper names) is
cover the case of logical inconsistency, as appears overt (e.g. Romance) or not (e.g. English). Feature 18 (Strong
in the data due to G. Longobardi: his school is in- Article) is presence of an indefinite article, i.e. of an obliga-
tory marker on singular indefinite count argument nominals,
volved in an ambitious and innovative project on distinct from those used for definite and mass indefinite, cf.
language classification based on syntax, cf. (Bor- (Longobardi et al., 2013).
346
stands for irrelevance. The idea is to play down since, as soon proved, cf. e.g. (Franzoi and Sgarro,
not only the contribution of 0’s and 12 ’s separately, 2017a), for two truth values x and y one has
as we have done in (Dinu et al., 2018), but rather (x ∧ y) ∨ (x ∧ y) equal to f (xi ) ∨ f (yi ) or to 1 −
the contribution of both 0’s and 12 ’s jointly. It [f (xi ) ∨ f (yi )] according whether there is conso-
will turn out that in this case, which is not gen- nance or dissonance. This distance, called hence-
uinely fuzzy, rather than to Muljačić distances, the forth Muljačić distance (and called Sgarro dis-
generalized Steinhaus transform had been better tance in (Deza and Deza, 2009), cf. also (Sgarro,
applied to the usual taxicab distance (Manhattan 1977)) is simply a natural generalization of crisp
distance, Minkowski distance), re-found when the Hamming distances to a fuzzy setting. As for
standard fuzzy logical operators of min and max alternative logical operators for conjunctions and
for conjunctions and disjunctions are replaced by disjunctions (different T-norms and T-conorms,
Łukasiewicz T-norms and T-conorms, cf. (Franzoi for which cf. e.g. (Dubois et al., 2000)), they have
and Sgarro, 2017b; Dinu et al., 2018). been discussed in (Franzoi and Sgarro, 2017b).
The paper is divided as follow: in Section 2 we From a metric point of view, the only attractive
shortly re-take both fuzzy Hamming distances, or choice, beside fuzzy Hamming distances, turned
Muljačić distances, and taxicab distances stress- out to be Łukasiewicz T-norms for conjunctions
ing how the latter relate to Łukasiewicz T-norms; and the corresponding T-conorms for disjunctions:
in Section 3 we introduce Steinhaus transform and . .
we apply it to taxi-cab or Łukasiewicz distances; x>y = (x + y − 1) ∨ 0, x⊥y = (x + y) ∧ 1
in Section 4 we introduce the general Steinhaus
One soon checks that in this case, rather curiously,
transform to deal with irrelevance and inconsis-
(x>y)⊥(x>y) turns out to be simply |x − y|,
tency jointly and we comment on our linguistic re-
and so the string distance one obtains is nothing
sults; in Section 5 we sum up our results.
else but the very well-known taxicab distance
P
2 Fuzzy Hamming Distances vs. dT (x, y) = i |xi − yi |, which in our context,
Łukasiewicz or Taxicab Distances when it is applied to fuzzy strings of length n,
might be also legitimately called Łukasiewicz
We need some notations and definitions: we distance.
. .
set x ∧ y = min [x, y], x ∨ y = max [x, y] and
.
x = 1−x; these are the truth values of conjunction If we consider the fuzziness f (x)=d(x,
˙ x) of a
AND, disjunction OR and negation NOT, w.r. to logical value x and if we use the Muljačić distance,
propositions with truth values x and y in standard then we get fM (x) = x ∧ (1 − x); if we use in-
fuzzy logic, a relevant form of multi-valued logic; stead the Łukasiewicz distance, then the fuzziness
x ∈ [0, 1]. Define the fuzziness of the truth value is always 0.
.
x to be f (x) = x ∧ (1 − x). For the truth values x However, if we consider another equally legit-
and y in [0, 1] we say that x and y are consonant imate definition of fuzziness, namely “ambigu-
if either x ∨ y ≤ 12 or x ∧ y ≥ 21 , else they are ity - crispness”, which can be formalized as 21 −
dissonant; let D and C denote the set of dissonant
d x, 12 , then if we use the Muljačić distance the
and consonant positions i, respectively. We define
new fuzziness is 0, but if we use the
Łukasiewicz
the following distance for strings x, y ∈ [0, 1]n :
distance it is fT (x) = 12 −dT x, 12 = x∧(1−x):
.
dM (x, y) = the result of the competition Muljačić distance vs.
Łukasiewicz distance turns out to be a tie. In the
X X
[1 − [f (xi ) ∨ f (yi )]] + [f (xi ) ∨ f (yi )] next Section we explain why, with Longobardi’s
i∈D i∈C data, we decided to resort to taxicab distances.
(1) The distance in (1) is a fuzzy metric distance,
This expression stresses the link with crisp Ham- cf. (Sgarro, 1977; Franzoi and Sgarro, 2017a),
ming distances for binary strings ∈ {0, 1}n , but its from which a standard metric distance is soon ob-
meaning is better understood due to the following tained by imposing that self-distances dM (x, y)
fact: each of the n additive terms summed is the should be 0, while, unless x is crisp (i.e. belong to
truth value of the statement: {0, 1}n , the set of the 2n binary strings of length
[( feature fi is present in L and absent in Λ) or n), the value given by (1) would be strictly posi-
(feature fi is absent in L and present in Λ)] tive.
347
As for taxicab or Łukasiewicz distances, the self- sets due to Muljačić and Longobardi. So far we
distance dT (x, y) is always zero even when the ar- have dealt separately with irrelevance and incon-
gument x is not crisp, a possibly unpleasant fact sistency, but a question arises spontaneously: can
in a fuzzy context (but not in ours), as argued in we consider jointly both irrelevance and inconsis-
(Franzoi and Sgarro, 2017b). tency? Does a mathematical tool which takes into
account both of them exist? The answer is yes and
3 Steinhaus Transforms the tool we are looking for is the generalized Stein-
In the general situation, one has objects x, y, . . ., haus transform or generalized biotope transform,
not necessarily strings, a metric distance d(x, y) cf. (Deza and Deza, 2009).
between objects, and a special object z called the Prompted by arguments defended by G. Longo-
“pivot-object”. The Steinhaus transform, cf. (Deza bardi and his school, cf. (Bortolussi et al., 2011;
and Deza, 2009), itself proven to be a metric dis- Longobardi et al., 2016, 2013, 2015), the novelty
tance, is: of this section is that, since in the language classi-
fications features can be irrelevant or inconsistent,
. 2d(x, y) we want to consider both aspects together.
Sd (x, y) =
d(x, y) + d(x, z) + d(y, z) As we said above the idea is to play down not
set equal to zero when x = y = z. only the contribution of 0’s, as in the case of
In our case the objects are strings and pivots z will irrelevance, but also the contribution of the 12 -
always be constant strings z = (z, . . . , z), zi = z, positions. Unlike ours, Longobardi’s non-metric
∀i, z ∈ [0, 1]. distance gets rid of irrelevant and inconsistent po-
If one starts with the crisp Hamming distance, one sitions in quite a drastic way, possibly a serious
obtains the usual crisp Jaccard distance (distances draw-back, as we comment in our Conclusions.
from the pivot are then Hamming weights); start- The generalized Steinhaus transform, or general-
ing with the more general fuzzy Hamming dis- ized biotope transform, is:
tance, or Muljačić distance, one has an appropri-
2d(x, y)
ate Jaccard-like generalization, which weighs only Sd (x, y) =
“little” a position where both x and y are “almost d(x, y) + inf z∈M (d(x, z) + d(y, z))
(2)
0”, and which accounts for irrelevance in itself, but
where M is the set of pivots we are considering,
not for inconsistency, as instead we need.
cf. (Deza and Deza, 2009).
If the term dM (x, z) is equal to the fuzzy Hamming
. P We tackle Longobardi’s data (or rather to a sample
weight w(x) = i xi for z = 0, it is equal to n2
of his languages, since the data he and his school
independent of x when z = 12 , a constant pivot
are providing are steadily improving and extend-
string which we shall need to deal with inconsis-
ing), data which are not really fuzzy, even if we
tency. The fact that dM (x, z) with z = 12 is inde-
have decided to “simulate” logical inconsistency
pendent of x is a serious drawback, indeed. This
by total fuzziness. In this case the number of fea-
is why in the case of Longobardi’s data, we have
tures is 53, and the languages are: Sic = Sicilian,
applied the Steinhaus transform, rather than to
Cal = Calabrese as spoken in South Italy, It = Ital-
the fuzzy Hamming distance or Muljačić distance,
ian, Sal = Salentin as spoken in Salento, South
directly to the taxicab distance or Łukasiewicz
Italy, Sp = Spanish, Fr = French, Ptg = Portuguese,
distance dT (x, y). In this case, in the denomi-
Rm = Romanian, Lat = Latin, ClG = Classical At-
nator of the corresponding Steinhaus transform,
tic Greek, NTG = New Testament Greek, BoG =
the fuzzy Hamming weight w(x) is replaced by
P
1
Bova Greek as spoken in the village of Bova, Italy,
dT (x, z) = i xi − 2 . In the next Section, more Gri = Grico, a variant of Greek spoken in South
ambitiously, we shall deal jointly with both irrele- Italy, Grk = Greek, Got = Gothic, OE = Old En-
vance and inconsistency. glish, E = English, D = German, Da = Danish, Ice
= Icelandic, Nor = Norwegian, Blg = Bulgarian,
4 Dealing with Irrelevance and
SC = Serbo Croatian, Slo = Slovenian, Po = Pol-
Inconsistency
ish, Rus = Russian, Ir = Gaelic, Wel = Welsh, Far
In (Franzoi and Sgarro, 2017a,b; Franzoi, 2017; = Farsi, Ma = Marathi, Hi = Hindi, Ar = Arabic,
Dinu et al., 2018) one has presented new methods Heb = Hebrew or ’ivrit, Hu = Hungarian, Finn =
for language classification, testing them on data Finnish, StB = Standard Basque, WB = Western
348
Basque, Wo = Wolof as spoken mainly in Senegal. Italian is more integrated with the Ibero-Romance
For comparison reasons, we have selected a part of languages (i.e. Portuguese ans Spanish), which
Longobardi’s ndata set
o compound by 38 languages; are clustered together like in the standard lan-
taking M = 0, 21 in (2), the UPGMA tree we guage classifications. The three Italian dialects
obtain is given in the following figure: (i.e. Salentine, Sicilian and Calabrese) are exter-
nal to this cluster in our case in Fig. 1, while in
the original Longobardi’s tree (Fig. 2) they are
integrated with Italian and then the entire group
is linked with French and after with the Ibero-
Romance group. In both trees the Romanian is
grouped with Romance languages, but is the most
exterior with the languages from this group. In
both trees the Celtic languages Gaelic (Ir) and
Welsh (Wel) and Germanic languages are grouped
together, but in the Longobardi’s tree in Fig. 2 the
Celtic group is more integrated with the Germanic
group. There are two main differences between
the two trees: the first one is that in Longobardi’s
tree in Fig. 2 Bulgarian is grouped with Slavic lan-
guages; the second one is the moving of the entire
Slavic group from a closet proximity with the Ger-
manic group (in the Longobardi’s tree) to a more
distance linkage with them in our case.
Our classification compares with the one obtained
by Longobardi’s school with these data, cf. com-
Figure 1: Generalized Steinhaus transform with ments in the Conclusion, where we argue why our
taxi-cab distance and Longobardi’s data distance is quite promising for the new and ambi-
tious data Longobardi’s school are now providing.
while the Longobardi’s original tree is the follow- Actually, our distance compares rather well also
ing one: with the classification obtained by Q. D. Atkin-
son and R. D. Gray, cf. (Gray and Atkinson, 2003;
Bouckaert et al., 2012).
349
First of all for the classification we have used Lon-
gobardi’s dataset, while Atkinson and Gray have
used their own dataset. If we look to Marathi and
Hindi we can notice that they are grouped together
in both trees; also Polish, Russian, Serbo Croa-
tian and Slovenian are grouped together in both
trees; the same is for New Testament Greek, Greek
and Classical Attic Greek. Also the Celtic lan-
guages (i.e. Gaelic and Welsh) and Germanic lan-
guages are grouped together. Our misclassifica-
tion of Bulgarian is not that worring, since Longo-
bardi covers only the syntax of the noun, and the
Bulgarian noun is well-known to behave in quite a
non-Slavic way, due possibly to its Balcanian sub-
stratum.
5 Conclusions
350
linguistic, first of all in coding theory cf. (Franzoi
Table 2: Longobardi original data
and Sgarro, 2017a), or even in bioinformatics.
Irrelevance and inconsistency appear to be fea-
ft. Blg SC Slo Po Rus Ir Wel Far Ma Hi Ar Heb Hu Fin StB wB Wo
tures which are dealt with quite sparsely, if ever,
outside Longobardi’s school; actually, these flexi- 1. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
ble features might prove to be quite useful not only 2. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
3. 1 1 1 1 1 1 1 0 1 1 1 1 0 0 0 1 1
in linguistic classification phylogeny, cf. (Franzoi 4. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1
5. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0
and Sgarro, 2017a,b), but also in the investigation 6. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 * * *
of the history of texts. So far, we are just provid- 7. 1 0 0 0 0 1 1 0 0 0 1 1 1 0 0 0 1
8. 1 * * * * 1 1 * * * 1 1 1 * * * 1
ing technical tools to be used in Longobardi’s re- 9. 1 * * * * 0 0 * * * 1 1 1 * * * *
10. 0 0 0 0 0 0 0 0 0 0 0 0 0 1 * * *
search, which, in its turn, is methodically matched 11. 0 * * * * 0 0 * * * 0 0 0 * 0 1 1
with the current state of the art, cf. (Bortolussi 12.
13.
1
*
*
*
*
*
*
*
*
*
0
*
0
*
*
*
*
*
*
*
0
*
0
*
0
*
*
*
*
*
*
*
0
*
et al., 2011; Longobardi et al., 2016, 2013, 2015; 14. 0 * * * * 0 0 * * * 1 0 0 * * * 1
15. 0 0 0 0 0 0 0 0 0 0 1 0 0 0 * * *
Longobardi, 2017; Kazakov et al., 2017). 16. 1 1 1 1 1 1 0 0 1 1 1 1 0 0 * * *
17. 0 0 0 0 0 0 0 1 1 1 0 0 1 0 0 0 1
18. 0 * * * * 0 0 * * * 0 0 * * * * *
19. * * * * * * * 1 0 0 * * 0 * * * 0
20. * * * * * * * * * * 0 0 0 * 1 1 *
21. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0
22.
Table 1: Longobardi original data 23.
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
1
1
1
1
1
1
1
1
0
1
1
1
*
1
*
1
*
1
24. * * * * * 0 * * * * * * * * * * *
25. 1 1 1 1 1 1 1 1 0 0 1 1 1 1 0 0 1
ft. Sic Cal It Sal Sp Fr Ptg Rm Lat CIG NtG BoG Gri Grk Got OE E D Da Ice Nor 26. * * * * * * * * 1 1 * * * * 0 0 *
27. 0 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 *
28. 0 0 0 0 0 0 0 0 0 0 1 1 0 0 1 1 *
1. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 29. 1 0 0 0 0 1 1 1 1 1 1 1 0 0 1 1 0
2. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 30. 0 * * * * 0 0 0 0 0 0 0 * * 0 0 *
3. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
31. 1 * * * * 1 1 1 * * 1 1 * * * * *
4. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
32. 0 1 1 1 1 1 1 1 0 0 1 1 0 0 0 0 1
5. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
6. 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 33. 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0
7. 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 34. 0 * * * * 1 1 * * * 0 0 0 * * * 0
8. 1 1 1 1 1 1 1 1 * 1 1 1 1 1 0 1 1 1 1 1 1 35. 1 1 1 1 1 0 * * 1 1 0 0 * 0 * * 0
9. 1 1 1 1 1 1 1 1 * 1 1 1 1 1 * 0 0 0 0 0 0 36. 0 * * * * 0 * * 1 1 * * * * * * *
10. 0 0 0 0 0 * 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 37. 1 1 1 0 1 * 0 0 0 0 * * * * 0 0 *
11. 0 0 0 0 0 0 0 0 * 0 0 0 0 0 * 0 0 0 0 0 0 38. 0 0 0 0 0 0 0 0 0 0 1 1 * * 0 0 *
12. 0 0 0 0 0 0 0 1 * 0 0 0 0 0 * 0 0 0 1 1 1 39. 1 1 1 1 1 0 0 1 1 1 1 0 1 1 1 0 1
13. * * * * * * * * * * * * * * * * * * 0 1 1 40. 0 0 0 0 0 * * 0 0 0 1 * 0 0 0 * 0
14. 0 0 0 0 0 0 0 0 * 0 0 0 0 0 * 0 0 0 0 0 0 41. 1 * * * * * * * * * 0 * 0 * 1 * 1
15. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 42. 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0
16. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
43. 0 0 0 0 0 0 0 1 0 0 * * 0 0 0 0 0
17. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
44. 0 0 0 0 0 1 1 * 0 0 * * 0 0 * * *
18. 1 1 1 1 1 * 1 1 * 0 0 1 1 1 * 0 1 1 1 0 1
19. * * * * * * * * * * * * * * * * * * * * * 45. 0 0 0 0 0 * * * 0 0 * * 0 0 * * *
20. 0 0 0 0 1 0 1 * * 1 1 0 0 0 * * * * * * * 46. 0 0 0 0 0 * * * 0 0 * * 0 0 * * *
21. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 47. * 1 1 1 1 * * * * * * * * * * * 1
22. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 48. 1 * * * * * * * 1 1 * * 1 1 * * *
23. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 1 1 1 49. 0 0 0 0 0 * * * 0 0 * * 0 0 * * *
24. * * * * * * * * * * * * * * * 1 .* 1 1 1 1 50. 0 0 0 1 1 * * * 0 0 * * 0 0 * * *
25. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 51. * * * * * 0 0 * * * 1 1 * * * * 0
26. * * * * * * * * * * * * * * * * * * * * * 52. * * * 0 0 0 0 1 * * 0 0 * * 0 0 1
27. 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 53. 0 0 0 0 0 * * 0 * * * * 1 1 1 1 *
28. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
29. 1 1 1 1 1 1 1 0 1 1 0 0 0 0 0 0 1 1 1 1 1
30. 0 0 0 0 0 0 0 * 1 1 * * * * * * 0 0 0 0 0
31. 1 1 1 1 1 1 1 * 1 1 * * * * * * 1 1 1 1 1
32.
33.
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
*
0
*
0
1
0
1
0
1
0
1
0
1
0
1
0
0
0
1
0
0
0
1
0
0
0
References
34. 0 0 0 0 1 1 * 0 * 0 0 0 0 0 * * * * * * *
35. 1 1 1 1 1 1 1 1 1 1 1 0 0 0 1 1 * 1 1 1 1 2018. Ethnologue. https://www.ethnologue.com/.
36. 0 0 0 0 0 0 0 * 0 0 * * * * * * * 0 0 0 0
37. 0 0 0 0 0 0 0 0 0 0 0 * * * 0 0 0 0 0 0 0
38. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 L. Bortolussi, A. Sgarro, G. Longobardi, and
39.
40.
*
0
*
0
*
0
*
0
*
1
*
0
*
0
*
1
1
1
1
1
1
1
*
0
*
0
*
1
1
1
1
0
*
0
*
0
*
0
1
0
*
0
C. Guardiano. 2011. How many possible languages
41. 1 1 1 1 1 1 1 1 * 0 0 1 1 0 * 1 1 1 1 * 1 are there? Biology, Computation and Linguistics
42.
43.
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
pages 168–179.
44. 1 1 0 1 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0
45. * * 1 * 1 1 1 1 0 0 0 * * 0 0 0 0 0 0 0 0 R. Bouckaert, P. Lemey, M. Dunn, S. J. Greenhill,
46. * * * * * * * * 0 0 0 * * 0 0 0 0 0 0 0 0
47. * * * * * * * * 0 0 1 * * 1 1 1 * 1 * 1 * A. V. Alekseyenko, A. J. Drummond, R. D. Gray,
48. * * * * * * * * 1 1 * * * * * * 0 * 0 * 0 M. A. Suchard, and Q. D. Atkinson. 2012. Mapping
49. * * * * * * * * 1 1 1 * * 1 1 0 0 0 0 0 0
50. * * * * * * * * * * * * * * * 0 0 0 0 0 0 the origins and expansion of the indo-european lan-
51. 0 0 0 0 0 0 0 0 * 1 1 0 0 1 1 * 0 * * * * guage family. Science 337(6097):957–960.
52. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 * * * *
53. * * 1 * 1 1 1 1 0 0 0 * * 0 * 1 1 1 1 1 1
A. M. Ciobanu, L. P. Dinu, and A.Sgarro. 2018. To-
wards a map of the syntactic similarity of languages.
In CICLing 2017. volume LNCS 10761, pages 1–15.
351
M. M. Deza and E. Deza. 2009. Encyclopedia of G. Longobardi, C. Guardiano, G. Silvestri, A. Boat-
Distances. Springer Dordrecht Heidelberg, London tini, and A. Ceolin. 2013. Toward a syntactic phy-
New York. logeny of modern indo-european languages. Jour-
nal of Historical Linguistics 3:11:122–152.
A. Dinu, L. P. Dinu, L. Franzoi, and A. Sgarro. 2018.
Steinhaus transforms of fuzzy string distances in Ž. Muljačić. 1967. Die Klassifikation der romanischen
computational linguistics. In Information Process- Sprachen. Rom. Jahrbuch 18 pages 23–37.
ing and Management of Uncertainty in Knowledge-
Based Systems. Theory and Foundations - 17th In- A. Sgarro. 1977. A fuzzy hamming distance. Bullettin
ternational Conference, IPMU 2018, Cádiz, Spain, Math. de la Soc. Sci. Math. de la R. S. de Roumanie
June 11-15, 2018. volume Proceedings, Part I, pages 69(1-2):137–144.
171–182.
L. A. Zadeh. 1965. Fuzzy sets. Information and Con-
D. Dubois, H. T. Nguyen, and H. Prade. 2000. Pos- trol 8(3):338–353.
sibility theory, probability and fuzzy sets: Misun-
derstanding, bridges and gaps. In Fundamentals
of Fuzzy Sets. Kluwer Academic Publishers, pages
343–438.
L. Franzoi. 2017. Jaccard-like fuzzy distances for com-
putational linguistics. In 19th International Sympo-
sium on Symbolic and Numeric Algorithms for Sci-
entific Computing, SYNASC 2017, Timişoara, Roma-
nia, September 21-24, 2017. pages 196–202.
L. Franzoi and A. Sgarro. 2017a. Fuzzy hamming dis-
tinguishability. In 2017 IEEE International Confer-
ence on Fuzzy Systems, FUZZ-IEEE 2017, Naples,
Italy, July 9-12, 2017. pages 1–6.
L. Franzoi and A. Sgarro. 2017b. Linguistic classifi-
cation: T-norms, fuzzy distances and fuzzy distin-
guishabilities. In Knowledge-Based and Intelligent
Information & Engineering Systems: Proceedings of
the 21st International Conference KES-2017, Mar-
seille, France,
6-8 September 2017.. pages 1168–1177.
R. D. Gray and Q. D. Atkinson. 2003. Language-
tree divergence times support the anatolian theory
of indo-european origin. Nature 426:435–439.
D. Kazakov, G. Cordoni, A. Ceolin, M. A. Ir-
imia, S. Kim, D. Michelioudakis, N. Radkevich,
C. Guardiano, and G. Longobardi. 2017. Machine
learning models of universal grammar parameter de-
pendencies. Proceedings of Knowledge Resources
for the Socio-Economic Sciences and Humanities
associated with RANLP-17 pages 31–37.
G. Longobardi. 2017. Principles, parameters, and
schemata: A radically underspecified ug. Linguis-
tic Analysis 41(3–4):517–557.
G. Longobardi, A. Ceolin, L. Bortolussi, C. Guardiano,
M. A. Irimia, D. Michelioudakis, N. Radkevich, and
A. Sgarro. 2016. Mathematical modeling of gram-
matical diversity supports the historical reality of
formal syntax. University of Tübingen, online pub-
lication system Tübingen DEU pages 1–4.
G. Longobardi, S. Ghirotto, C. Guardiano, F. Tassi,
A. Benazzo, A. Ceolin, and G. Barbujan. 2015.
Across language families: Genome diversity mirrors
language variation within europe. American Journal
of Phisical Anthropology 157:630–640.
352
Corpus Lexicography in a Wider Context
Chen Gafni
Bar-Ilan University
chen.gafni@gmail.com
Abstract
2 Preliminaries
This paper describes a set of tools that
offers comprehensive solutions for corpus The described system is implemented in the Child
lexicography. The tools perform a range of Phonology Analyzer software (CPA; Gafni, 2015) 1,
tasks, including construction of corpus which was built in MS Excel due to its popularity
lexicon, integrating information from and user-friendly interface. 2 Nevertheless, the
external dictionaries, internal analysis of concepts behind the system are general and can be
the lexicon, and lexical analysis of the
implemented in various environments. The
corpus. The set of tools is particularly
useful for creating dictionaries for under-
software can analyze corpora stored in various file
resourced languages. The tools are formats, including Excel and plain-text files, as
integrated in a general-purpose software well as several special formats used in linguistic
that includes additional tools for various research: Praat’s TextGrids, CHAT transcription
research tasks, such as linguistic files, EAF annotation files, and XML schema for
development analysis. Equipped with a TalkBank data. The software converts analyzed
user-friendly interface, the described corpora into Excel format and adds all analysis
system can be easily incorporated in products to the Excel file.
research in a variety of fields.
2.1 Organizing the Data
1 Introduction
The described tools (“macros”) require that the
Corpus lexicography, a key component in modern corpus text be stored in a vector format. The text
dictionary compilation, has become increasingly can be converted to a vector format using CPA’s
powerful and efficient due to the development of “Data preparation” macro with the “Corpus
various tools such as Word Sketch (Kilgarriff and tokenization” option, which segments the text into
Tugwell, 2002) and TickBox Lexicography words.
(Kilgarriff et al., 2010). While corpus lexicography Segmentation is performed on the basis of blank
deserves further development in its own right, it is spaces and additional word-dividing characters,
worthwhile considering it as an integral part of which can be defined in CPA’s “Word dividers”
wider scientific missions. This paper describes table (Figure 1). There are two types of word
several tools for corpus lexicography, whose dividers, which can be used for separating words
design takes into consideration their contribution to even at the absence of a blank space: punctuation
linguistic research. The tools are integrated in a marks (e.g., comma) are deleted during
general-purpose linguistic software, where they segmentation, while final letters are not (final
can be readily applied in language acquisition letters are special letter forms appearing only at
studies, psycholinguistics, and other fields of word endings. See some examples from Hebrew in
research. Figure 1).
353
Proceedings of Recent Advances in Natural Language Processing, pages 353–359,
Varna, Bulgaria, Sep 2–4 2019.
2.2 Longitudinal and Multi-Level Corpora
The system was designed especially for analyzing
longitudinal data from language acquisition
studies. Such corpora typically list utterances made
by a child alongside the hypothesized intended
target utterances (a corpus containing such paired
utterances is a multi-level corpus). The age of the
child is also recorded for each utterance for the
purpose of developmental analysis (Figure 2).
Corpus tokenization (see above) processes all the Figure 1: Word dividing symbols
above-mentioned information. This further allows
analyzing both the input (target) and output
(production) lexicons of the child from a
developmental perspective.
354
information needs to be added manually. However,
if there is an available resource for the particular
language, the “Import dictionary” macro can
import the lexical information from the external
resource.
This macro receives as input the corpus lexicon
and an external dictionary in a table format. The Figure 5: An external words dictionary
macro copies information from the external
dictionary to the lexicon for every lexicon entry For example, a condition value “Noun” indicates
found in the dictionary. The macro can be used for that the affix applies to nouns. (c) Function: the
importing information about lexical words (Figure name of the lexical field modified by the affix. For
5), as well as grammatical affixes (Figure 6). example, a value of “Definiteness” indicates that
The macro was specially designed to handle the affix specifies the definiteness value of the
lexical entries containing explicit morphological hosting word. (d) Value: the value assigned to the
boundaries. For such lexical entries (e.g., le#ˈkof lexical field specified in the “Function” field. For
‘to.monkey’ (Hebrew)), the macro first searches example, a value of “Def” indicates that the affix
the full entry in the words dictionary. If not found, marks the hosting word as being definite.
the macro then searches the individual morphemes.
If a morpheme is found in the dictionary (e.g., kof),
the macro imports the information for that entry to
the corpus lexicon (Figure 7).
When importing information from an affix Figure 6: An external affixes dictionary
dictionary, the macro can use the external list of
affixes to remove irrelevant entries from the corpus For each affix in the affixes dictionary, it is
affixes lexicon (i.e., stems included in the affixes possible to define multiple feature quadruplets
lexicon during its construction). (e.g., “Tier 1”, “Condition 1”, …, “Tier 2”,
An affixes dictionary is constructed in a similar “Condition 2”, etc.). This option is useful for
way to a words dictionary; it contains a list of handling affixes that can affect multiple word
affixes with additional columns providing classes (e.g., nouns and adjectives) or have
information about these affixes. However, the multiple functions (e.g., express possession and
additional fields have a functional role: they mark tense).
specify how the affix modifies the properties of
affixed words. This information can be used for
modifying polymorphemic entries in the lexicon
(see 5.1).
Figure 7: Imported lexical information based
on the stems of prefixed words
5 Analyzing the Lexicon
The “Morphological analysis” macro finds
5.1 Morphological Analysis
lexical entries containing affixes and modifies their
If the corpus lexicon contains polymorphemic properties according to the details of the affix. If an
words with explicit marking of morphological affix modifies a lexical field not defined in the
boundaries, and an affixes dictionary is available lexicon, the macro adds that field to the lexicon
(see 4), the “Morphological analysis” macro can (Figure 8).
import information from the affixes dictionary to
the corpus lexicon.
Each entry in the affixes dictionary should have
the following fields (Figure 6): (a) Tier: a name of
a field in the words lexicon. For example, a “POS”
value in the tier field (stands for “Part-of-speech”) Figure 8: Lexical entries of prefixed words
after morphological analysis
indicates that the affix applies to lexical items in a
specific lexical category. (b) Condition: a possible
value of the lexical field specified in the tier field.
355
5.2 Lexicon Summary the contributing corpora (e.g., if an item appears 10
This macro generates a summary table of the times in one corpus and 20 times in another corpus,
lexicon. The summary table includes a list for each the merged lexicon will record 30 tokens for that
lexical field (e.g., “POS”) that specifies the various item). In addition, the merged entry will contain the
values of the field (e.g., “Noun”, “Verb”). For each lexical properties collected from all contributing
value, the list indicates the number of corpus tokens entries. In case of conflicting inputs (e.g., an item
and types. The number of types is the number of is classified as a noun in one lexicon and as a verb
items in the lexicon with the relevant value (e.g., in another), the merged entry will indicate all
the number of noun types), and the number of possible values for that property (e.g., Noun /
corpus tokens is calculated from the “Count” field Verb). The merging macro can also add labels
in the lexicon (Figure 9). indicating the source(s) (i.e., the name of the input
lexicon) of each entry.
7 Lexical Development
Assessing the size of the child’s lexicon is an
important part of longitudinal language acquisition
studies, from both theoretical and clinical
perspectives. In particular, there is evidence that
aspects of grammatical development are tightly
correlated with vocabulary size (Bates and
Goodman, 1997).
Figure 9: Lexicon summary by part-of-speech The “Lexical development” macro analyzes
lexical growth in corpora that record the age of
6 Integrating Lexicons production of every utterance. Using the age of first
attempt to produce target words (see 3), the macro
Efficient integration of information is essential for divides the child’s lexicon into stages of lexical
compiling a dictionary based on data from multiple development (Figure 10). The first stage is marked
resources. The “Merge worksheets” macro is a by the acquisition of the first 10 words, the second
general utility macro that integrates the contents of by a total lexicon size of 50 words, and then an
multiple spreadsheets in a file. Thus, it requires additional 50 words for every subsequent stage
lexicons generated from different corpora to be (Adam and Bat-El, 2009).
stored in one file (this can be done either manually
or automatically with the “Merge workbooks” CPA
macro).
The “Merge worksheets” macro has several
operation modes, one of which is designed
specifically to integrate lexicon tables. The macro
receives as input any number of spreadsheets. It
creates a single lexicon3 containing information
from all input lexicons. The merged lexicon
contains the union of lexical fields in all input
lexicons (i.e., a lexical field will be included in the Figure 10: Lexical development
merged lexicon if it appears at least in one input
lexicon). Stages of lexical development are aligned with
The entries in the merged lexicon are sorted recording sessions, such that if a theoretical stage
alphabetically. If a lexical entry appears in multiple boundary is reached in mid-session, the actual
input lexicons, the duplicate entries are merged. boundary will be assigned either to that session or
The merged entry summarizes token counts from to the preceding session (whichever is closer). For
356
Figure 11: Lexicosyntactic query form
example, if the child has reached 49 cumulative entry has no lemma specified, the surface form of
target types at the end of session 1 and 58 the entry will be taken as the lemma.
cumulative target types at the end of session 2, the
theoretical landmark of 50 words will be assigned 8 Lexical Queries
to session 1. If the lexicon grows rapidly, such that
more than one theoretical stage is passed in a single Once lexical properties are specified in the lexicon,
session, the macro will “skip” intermediate stages this information can be used to analyze the corpus.
and assign only the last stage to that session. For CPA has a set of macros that can extract linguistic
example, if the size of the lexicon jumps from 100 information from the corpus on various levels of
words (stage 3) to 200 words (stage 5) in a single analysis, via a user-friendly query form (Figure 11).
session, that session will be marked as the end point One of these macros, “Lexicosyntactic query”,
of stage 5, skipping stage 4. The macro also queries the corpus at the word and utterance levels.
provides a more fine-grained account, indicating Specifically, “Content Lexicosyntactic queries”
the number of new words added to the lexicon in can find occurrences of lexical properties and
every session and the total lexicon size after every sequences of lexical properties in the corpus. For
session. example, the query [Verb] [Noun] will find all
By default, lexical development is calculated instances of verbs followed by nouns in the corpus.
based on the full list of lexical entries. However, Similarly, the query [Verb,1,SG] [Noun,SG] will
this list is organized by word form (types), such that find all instances of verbs conjugated in the first
words that are interrelated via inflectional person singular followed by singular nouns.
morphology (e.g., cat–cats) are listed as separate The scope of queries can be constrained by age
entries. Relying on plain surface forms can result in or stage of lexical development. Thus, for example,
over-estimation of lexicon size. This can be it is possible to get all verbs attempted by a child at
avoided by analyzing lexical development by a given age/lexical stage or range of ages/lexical
lemma/lexeme. When this option is chosen, the stages. This option allows for investigation of
macro analyzes the lemma field of the lexicon lexical development at a more fine-grained level.
rather than the word field. The lemma field Queries over single-item sequences (e.g.,
indicates the lemma of each lexical entry (e.g., the [Verb,SG]) calculate the number of tokens and
lemma of cat and cats is cat). The lemma types and can also return a list of items that
information can be supplied manually or imported matched the query (Figure 12). Queries over multi-
from an external dictionary (see 4). When a lexical item sequences (e.g., [Verb] [Noun]) do not return
357
looking for non-existing words, which likely
resulted from typos. For the lexicon of produced
forms (output lexicon), quality check mainly
involves examining tokens with unusual structure
that deviates from the phonology of the ambient
language. Thus, in the proposed approach, one
estimates the potential of lexical entries to contain
Figure 12: Corpus instances of singular errors, and then focuses on suspicious forms. This
masculine nouns (source) paired with the is more effective than examining corpus subsets
corresponding forms produced by an infant randomly.
(reference). The described tools can be integrated in any task
involving corpus analysis. For example, the CPA
specific items, but rather a list of indices of rows in
software includes an n-gram frequency calculator,
the corpus where such sequences are found.
which can calculate corpus-weighted mean n-gram
In addition to lexicosyntactic queries, CPA has
frequencies over a list of strings (in this context, n-
similar query macros for analyzing the
gram refers to a sequence of letters or phones
phonological properties of corpora. These queries,
within words). This is useful for creating controlled
too, can be correlated with lexical development.
sets of stimuli for psycholinguistic experiments.
Additional, more advanced macros can be used to
Finally, it should be acknowledged that there is
combine queries on different levels of analysis.
some overlap between the described system and
This allows, for example, to study the interaction
other existing systems. Well-established systems
between phonological and lexical development.
such as Sketch Engine (Kilgarriff et al., 2014)
provide powerful solutions for corpus
9 Discussion
lexicography, and the CLAN program
The described set of tools can help creating (MacWhinney, 2000) can be used for studying
resources for under-resourced languages. For lexical development.
example, it was used for creating a lexicon with Compared to these programs, CPA is currently
corpus frequency data for Hebrew (Gafni, 2019), limited in areas such as collocation analysis and
and it is currently being used in an ongoing POS tagging. On the other hand, CPA has some
longitudinal study of phonological development in advantages over these programs. Its unique built-in
twins. In addition to building a lexicon for each lexical development tool allows for more
participating child and assessing lexical comprehensive study of language development,
development, the system can assist in improving and its querying system allows for combined
the quality of the transcribed data. lexical and phonological corpus analysis. The user-
Given that the transcribed data can contain many friendly interface enhances user experience and
errors (typos, misperceptions), it is important to saves the need to learn complex query syntax, as
have it validated. Since the amount of transcribed used by the CLAN program. In addition, CPA is
data can be enormous (tens of thousands of tokens) distributed as an Excel file. This means that Excel
and the transcription task is very time-consuming, users can perform the various analysis tasks in the
it is impractical to have every token transcribed by natural environment of the data, without the need
multiple transcribers. One possibility to check data to install (or purchase) additional software.
quality is to have a random subset of the corpus To conclude, this paper views corpus
(e.g., 10% of the tokens) be transcribed by more lexicography in a wide context of linguistic
than one transcriber, and calculate inter-transcriber research. Accordingly, the described tools are
reliability. However, such an approach can help integrated in a single, user-friendly system
detecting problems in a limited part of the corpus. designed to support any task requiring corpus
The tools described in this paper offer a more analysis. Future improvements to the current
systematic approach to quality check of transcribed system will include the addition of standard
data. In this lexicon-based approach, one goes over lexicographic functions, such as collocation
the entries in the automatically generated corpus analysis and morphological analysis that does not
lexicon and looks for suspicious entries. For the require overt marking.
lexicon of target words, this mainly involves
358
References
Galit Adam and Outi Bat-El. 2009. When Do Universal
Preferences Emerge in Language Development? the
Acquisition of Hebrew Stress. Brill’s Journal of
Afroasiatic Languages and Linguistics, 1(1):255–
282.
Elizabeth Bates and Judith C. Goodman. 1997. On the
Inseparability of Grammar and the Lexicon:
Evidence from Acquisition, Aphasia and Real-time
Processing. Language and Cognitive Processes,
12(5–6):507–584.
Chen Gafni. 2015. Child Phonology Analyzer:
processing and analyzing transcribed speech. In The
Scottish Consortium for ICPhS 2015, editor,
Proceedings of the 18th International Congress of
Phonetic Sciences., pages 1–5, paper number 531,
Glasgow, UK: the University of Glasgow.
Chen Gafni. 2019. General Lexicons of Hebrew:
Resources for Linguistic and Psycholinguistic
Research (version 1.0).
Adam Kilgarriff, Vít Baisa, Jan Bušta, Miloš
Jakubíček, Vojtěch Kovář, Jan Michelfeit, Pavel
Rychlý, and Vít Suchomel. 2014. The Sketch
Engine: ten years on. Lexicography, 1(1):7–36.
Adam Kilgarriff, Vojtěch Kovář, and Pavel Rychlý.
2010. Tickbox lexicography. In Sylviane Granger
and Magli Paquot, editors, eLexicography in the
21st century: new challenges, new applications.
Proceedings of eLex 2009, volume 7, pages 411–
418, Louvain-La-Neuve. UCL Presses
Universitaires De Louvain.
Adam Kilgarriff and David Tugwell. 2002. Sketching
words. In Marie-Hélene Corréard, editor,
Lexicography and Natural Language Processing: A
Festschrift in Honour of B. T. S. Atkins, pages 125–
137. Euralex.
Brian MacWhinney. 2000. The CHILDES Project:
Tools for Analyzing Talk. Lawrence Erlbaum
Associates Inc, Mahwah, NJ, 3rd edition.
359
A Universal System for Automatic Text-to-Phonetics Conversion
Chen Gafni
Bar-Ilan University
chen.gafni@gmail.com
360
Proceedings of Recent Advances in Natural Language Processing, pages 360–366,
Varna, Bulgaria, Sep 2–4 2019.
preparation” macro, which segments the text into can add them to the phonetic dictionary for future
words. use.
Segmentation is performed on the basis of blank
spaces and additional word-dividing characters, 4 The Linguistic Model
which can be defined in CPA’s “Word dividers”
table (Figure 1). There are two types of word For languages with some degree of regular
dividers, which can be used for separating words mapping between spelling and sound, phonetic
even at the absence of a blank space: punctuation transcription of text can be generated automatically
marks (e.g., comma) are deleted during on the basis of deterministic rules. This section
segmentation, while final letters are not (final describes a multi-stage model of automatic
letters are special letter forms appearing only at phonetic transcription, guided by linguistic
word endings. See some examples from Hebrew in principles. Underlying this model is the assumption
Figure 1). Once the text is transformed into a vector that, for any regular orthographic system, phonetic
format, transcription can be performed. One of transcription rules can be defined in terms of a
CPA tools (“Reconstruct corpus”) can then be used small set of general operations. The general
to recombine the phonetic word forms according to operations themselves are hard-coded in the
the structure of the original text. software, but an unlimited number of language-
specific rules can be defined on the basis of these
operations. This flexible method allows the system
to produce automatic phonetic transcription for
every language that has, at least partly, regular
orthography.
The proposed model has four components,
which will be described in the following
subsections. The components operate
independently of one another, but they should be
Figure 1: Word dividing symbols applied in the order in which they are listed. The
first component alone produces sufficient results
for most purposes. If needed, the additional three
3 Phonetic Dictionaries components can be used, together, for fine-tuning.
For languages with irregular spelling or high 4.1 Pre-Prosody Transcription
proportion of homographs, such as English,
Hebrew, and Arabic, automatic phonetic This component takes as input the vector of words
transcription requires a source of ready-made to-be-transcribed, a table of pre-prosody
phonetic forms (i.e., a phonetic dictionary) for transcription rules (Table 1), and a table containing
irregular and ambiguous words. CPA has a built-in sets of symbols and strings, called “phono-
macro that can import such ready-made forms. The orthographic groups” (Table 2). 4 The pre-prosody
macro receives as input a vector of written words transcription 5 applies the transcription rules in
to-be-transcribed and a phonetic dictionary – a successive order to the list of words. Entities
table of written word forms and corresponding
phonetic forms. The macro matches phonetic forms
from the dictionary to written words in the vector.
For words that are not found in the phonetic
dictionary, transcription needs to be generated,
either manually or with the automatic linguistic
model (see next section). However, once the
additional phonetic word forms are supplied, CPA
Figure 2: Transcribed words
4 5 The
The tables of rules are generated manually, in principle. term ‘pre-prosody’ indicates that the transcription
CPA has a set of editable tables that contain proposed rules rules applied by this component disregard the prosodic
for several languages (see also 2.1). properties of the word, including syllable structure and
stress pattern.
361
Target Output Type Preceding Following
environment environment
1 ph f
2 c k
3 ay eɪ #? #
4 [cg] [sj] [Front_vowel]
5 [Consonant][Consonant] [Consonant] Degemination
6 ◌ّ Lengthening
7 [V_diac][C_diac] [C_diac][V_diac] Metathesis
Table 1: Pre-prosody transcription rules
defined in the table of phono-orthographic groups target, output or environment strings; an asterisk
may be called by transcription rules. The output of (*) stands for any number of successive characters;
the process is a vector of phonetic forms and, a hash sign (#) represents word boundaries.
corresponding to the written words (Figure 2). Wildcards can be used for defining phonological
The table of pre-prosody transcription rules has and morphological words patterns. For example,
five fields (Table 1): (1) Target: the input to the rule, rule 3 in Table 1 captures the pronunciation of ay at
i.e., the string to be converted. All rules must have the end of three-letter English words, such as bay
a value for the target string. The other fields are (/beɪ/). The question mark in this rule indicates that
optional. (2) Output: the string replacing the target; the rule applies to ay sequences preceded by a
if left empty, the target string will be deleted. (3) single character. The hash signs suggest that the
Type: the type of operation to be performed by the rule applies only when word boundaries are present
rule; if left empty, simple substitution will be at both edges.
performed (see below for other types of rules). (4) Rules can also be generalized by the inclusion of
Preceding environment, and (5) Following phono-orthographic groups, defined in a separate
environment: these fields are used for formulating table. Each entry in the table of phono-orthographic
context-sensitive rules. When either field is not groups has two fields (Table 2): the name of the
empty, the transcription rule will apply only to group, and its members. For example, the term
words in which the target string is preceded by the Front_vowel can be used for grouping e, i, and y.
‘preceding environment string’ and/or followed by Transcription rules can include phono-
the ‘following environment string’. orthographic groups by enclosing the name of the
Substitution rules can be used for simple group between brackets (e.g., [Front_vowel]).
grapheme-to-phoneme conversions, which can be When a group is embedded in a transcription rule,
either context-free or context-sensitive. For the algorithm converts the compact rule into a set
example, rule 1 in Table 1 is a context-free of simple rules, each applying to a different
substitution of ph by f in words such as phone member of the group. For example, rule 4 in Table
(/foʊn/). Rule 2 in Table 1 is an example for 1 uses groups to capture the pronunciation of c and
context-sensitive rule – deleting c before k in words g before front vowels (/s/ in cent /sent/ and /ʤ/ in
such as back (/bæk/). gene /ʤiːn/, respectively). This single, compact
Substitution rules may include wildcards to rule stands for six simple rules: c→s/_e, c→s/_i,
define more general entities. Three types of c→s/_y, g→ʤ/_e, g→ʤ/_i, g→ʤ/_y (where the
wildcards are defined in the software: a question formula A→B/_X is read: A becomes B before X).
mark (?) stands for any single character in the In addition to substitution rules, several types of
Group Members special operations can be used by specifying the
cg c,g name of the operation in the ‘Type’ field in the table
of rules. Three types of operations are defined in
sj s,ʤ
the software: degemination, lengthening, and
Front_vowel e,i,y
metathesis.
Consonant b,d,f,g,l,m,n,p,r,s,t,z Degemination is used for collapsing a sequence
C_diac ּ◌,ׂ◌,ׁ◌,' of two identical phones when pronounced as a
V_diac ְ◌, ֻ◌,ֹ◌, ָ◌, ַ◌, ֶ◌, ֵ◌, ִ◌ single, short sound. For example, rule 5 in Table 1
Table 2: Phono-orthographic groups
362
collapses sequences of identical consonants (e.g.,
mm is pronounced as a single m in hammer).
Lengthening realizes the function of diacritical
marks of lengthening/gemination. For example,
rule 6 in Table 1 realizes the function of the Arabic
Shaddah (e.g., the letter مis pronounced /m/ in its
plain form, but as /mm/ when modified by a
Shaddah, i.e.,ّ )م.
Finally, metathesis switches the order of Figure 3: Syllabified phonetic word forms
elements in sequences of two phono-orthographic
groups (i.e., if the target contains a member of In order to determine sites of sonority minimum,
group 1 followed by a member of group 2, they are the syllabification procedure converts the phonetic
switched in the output). For example, in pointed word forms into strings of sonority levels. Sonority
Hebrew scripts, diacritics are used for indicating levels are non-negative integers specified for each
vowels as well as for modifying the phonetic value phone in the phonetic table of CPA (Figure 4). For
of consonants. A single letter can host both example, if fricatives, nasals, liquids, and vowels
consonant and vowel diacritics (the C_diac and have sonority levels of 1, 2, 3, and 5, respectively,
V_diac groups in Table 2, respectively). Thus, the the sonority-level representation of fæməli will be
string ָבּ, pronounced /ba/, is composed of the letter 152535. In this representation, 2 and 3 are local
( בrepresenting the consonants /b/ and /v/), a sonority minima (Sonority minima at word edges
consonant diacritic ּ◌ (specifying the consonant are ignored).
/b/), and a vowel diacritic ָ◌ (representing the vowel
/a/). Although the order in which these diacritics
are attached to the letter does not affect the visual
form of the text, it is important for the purpose of
phonetic transcription – consonant diacritics must
be attached before vowel diacritics. For example,
the string ָבּcan be formed by combining the three
elements in two ways: ב+ ּ◌+ ָ◌, or ב+ ָ◌+ ּ◌. However,
only the first order reflects the phonological
structure of the string. A rule of metathesis can be
defined to switch the order in sequences of vowel
Figure 4: The phonetic table
diacritics + consonant diacritic to guarantee correct
ordering (rule 7 in Table 1).
After performing the pre-prosody transcription,
certain modifications might be needed due to
phonological processes related to prosodic
structure. This post-prosody transcription (see 4.4)
requires that the phonetic forms be parsed into
syllables and have stress markers assigned to them.
These components are described below.
4.2 Syllabification
This component takes as input a vector of phonetic
word forms, a list of binary parameters and
parameter weights, and a phonetic table. The output
is a vector of syllabified phonetic word forms
(Figure 3). The basic sites of syllable boundaries
are around local sonority minima (e.g., the
boldfaced consonants in fæməli ‘family’ →
fæ.mə.li). Figure 5: Syllabification parameters
363
Trigger Tier Position Process Result
1 [+STRID][+STRID] Features Coda Vowel epenthesis ə
2 Unstressed vowel CV Vowel reduction ə
3 Sonority decrease Sonority Onset Vowel epenthesis ə
Table 3: Post-prosody transcription rules
The basic sites of syllable boundaries can be
adjusted by a set of binary parameters, which
handle various cases, such as consonant sequences
(e.g., whether /dɪspleɪ/ ‘display’ should be parsed
as dɪs.pleɪ or dɪ.spleɪ). The system currently has 10
built-in parameters, which can be switched on and
off according to the properties of the examined
language (Figure 5). Some of the parameters
require phones to be recognized as vowels or
consonants. This information is also specified in
CPA’s phonetic table (Figure 4). For example, if the Figure 6: Phonetic word forms with stress
Complex onset > Coda parameter is switched on, frequent stress pattern. Manual corrections can be
consonant sequences will be parsed as complex made afterwards. If stress position depends on the
onsets (dɪ.spleɪ). If the parameter is switched off, number of syllables, it is possible to run stress
consonant sequences will be split between coda assignment multiple times, starting with the rule for
and onset positions (dɪs.pleɪ). the longer words. Checking the ‘Keep existing
When two parameters are potentially in conflict, stress markers’ option will prevent stress rules for
they can be ranked relative to each other by shorter words from applying to longer words, for
assigning different integer weights to them. For which stress has been assigned already.
example, if the Onset parameter is on, onsetless
syllables will be dispreferred (e.g., ˈmʌni → ˈmʌ.ni
‘money’). This can be overridden (e.g., ˈmʌn.i) by
switching on the Coda maximization in stressed
syllables parameter and giving it a higher weight
than the onset parameter (this requires that stress
would be marked on the word before running
syllabification).
364
structure triggering the required modification. A help studying the phonological properties of text
trigger can be a specific element or a sequence of corpora and it is also useful for creating resources
elements defined in the phonetic table (e.g., for under-resourced languages. For example, it was
[+STRID][+STRID] stands for a sequence of two used for creating a phonological dictionary for
stridents such as /s/ and /z/). In addition, there are Hebrew (Gafni, 2019). In addition, the linguistic
several types of special pre-defined triggers: The model of the software, by itself, can be used as a
Sonority decrease, Sonority increase, and Sonority research and educational tool. The pre-prosody
plateau triggers handle phone sequences in which transcription tool, in particular, can be used for
the sonority level decreases, increases, or remains exploring and demonstrating the effect of rule-
unchanged, respectively (see Figure 4). The No ordering – a common practice in theoretical
vowel trigger handles syllables with no vowels. The phonology. In fact, the studied language need not
Unstressed vowel trigger handles vowels in have a writing system at all; the input corpus can
unstressed syllables. be a list of hypothesized phonological underlying
(2) Tier: the phonological tier relevant to the representations, and the transcription rules can be
trigger. Features tier is used with phonological phonological rules transforming the underlying
features triggers (e.g., [+STRID]), while CV tier is representations to surface representations.
used with No vowel and Unstressed vowel triggers. In addition, the linguistic model can be used for
Sonority tier is used with all sonority-related calculating indices of linguistic complexity by
triggers. assessing the proportion of words that have regular
(3) Position: for triggers applying to consonants spelling in a given language and the number of
(sonority and feature triggers), this field indicates deterministic rules needed to capture the patterns of
the prosodic position (Onset or Coda) in which the orthographic regularity in a language. Such
trigger must be found in order to trigger the measures of complexity can be valuable for literacy
modification. education (e.g., Smythe et al., 2008).
(4) Process: the type of modification applied to It should be noted that the described system is
phonetic word forms in which the trigger is found. still under development. At its current state, the
Currently, the software can perform two types of transcription system can perform perfectly on
modifications: Vowel epenthesis inserts a vowel to completely regular orthographies with a fixed
correct ill-formed sequences, and Vowel reduction stress pattern. Several planned improvements will
replaces unstressed vowels with a default neutral allow the system to handle more complex cases.
vowel. For example, the stress assignment component
(5) Result: this field specifies inserted elements should handle secondary stress and stress rules that
(epenthetic vowels and neutral vowels). are sensitive to syllable weight. In addition, the
The following examples demonstrate the post-prosody transcription should include more
application of post-prosody rules. Rule 1 in Table 3 options, such as referring to pretonic syllables,
inserts an epenthetic ə to break sequences of two which are relevant sites for certain phonological
stridents in coda position (e.g., makss → maksəs processes like vowel reduction in Russian
‘Max's’, where makss is the output of the pre- (Asherov et al., 2016).
prosody transcription, which converted M to m and Furthermore, the generalizability of the system
x to ks, and deleted the apostrophe in Max's). Rule can greatly improve by adding machine learning
2 in Table 3 replaces vowels in unstressed syllables procedures, such as sequence-to-sequence models
by ə (e.g., ˈe.le.fant → ˈe.lə.fənt ‘elephant’, where with greedy decoding (Chae et al., 2018). This will
ˈe.le.fant is the result of pre-prosody transcription, allow the system to generate rules automatically
syllabification and assignment of antepenultimate based on examples. It will also be able to handle
stress to elephant). cases of homography (e.g., whether wind should be
transcribed /wɪnd/ (noun) or /waɪnd/ (verb)) by
5 Discussion analyzing token frequency and contextual effects
(syntax and semantics). Such improvements will
This paper describes a system of text-to-phonetics make the transcription system more powerful and
conversion. The system is incorporated in a reliable.
general-purpose linguistic software that includes
tools for building dictionaries, as well as corpus
analysis functions. Thus, the described system can
365
References
Daniel Asherov, Alon Fishman, and Evan-Gary Cohen.
2016. Vowel Reduction in Israeli Heritage Russian.
Heritage Language Journal, 2:113–133.
Timur Baytukalov. 2019. EasyPronunciation.com: All-
in-one solution to learn pronunciation online.
Moon Jung Chae, Kyubyong Park, Linhyun Bang,
Soobin Suh, Longhyuk Park, Namju Kimt, and
Longhun Park. 2018. Convolutional sequence to
sequence model with non-sequential greedy
decoding for grapheme to phoneme conversion. In
Proceedings of the 2018 IEEE International
Conference on Acoustics, Speech and Signal
Processing (ICASSP 2018), pages 2486–2490.
Chen Gafni. 2015. Child Phonology Analyzer:
processing and analyzing transcribed speech. In The
Scottish Consortium for ICPhS 2015, editor,
Proceedings of the 18th International Congress of
Phonetic Sciences., page 1–5, paper number 531,
Glasgow, UK: the University of Glasgow.
Chen Gafni. 2019. General Lexicons of Hebrew:
Resources for Linguistic and Psycholinguistic
Research (version 1.0).
Ian Smythe, John Everatt, Nasser Al-Menaye, Xianyou
He, Simone Capellini, Eva Gyarmathy, and Linda S.
Siegel. 2008. Predictors of word-level literacy
amongst Grade 3 children in five diverse languages.
Dyslexia, 14(3):170–187.
Eran Tomer. 2012. Automatic Hebrew Text
Vocalization. Ph.D. thesis, Ben-Gurion University
of the Negev.
366
Two Discourse Tree - Based Approaches to Indexing Answers
Boris Galitsky1 and Dmitry Ilvovsky2
1
Oracle Inc. Redwood Shores CA
2
National Research University Higher School of Economics
boris.galitsky@oracle.com; dilvovsky@hse.ru
367
Proceedings of Recent Advances in Natural
1 Language Processing, pages 367–372,
Varna, Bulgaria, Sep 2–4 2019.
Rhetoric relations are split into the classes of For example, the phrase “it is generally a good
coordinate and subordinate; these relations hold idea” adds little to the answer, whereas “consider
across two or more text spans and therefore im- not using the proceeds from the conversion” is
plement coherence. These text spans are called el- informative to the user who posed the original
ementary discourse units (EDUs). The leaves of a question. Each answer in the question-answer in-
discourse tree correspond to EDUs, the contigu- dex may provide additional insight in terms of
ous atomic text spans. Adjacent EDUs are con- additional questions that can be answered, which
nected by coherence relations (e.g., attribution, are in turn indexed, increasing the usefulness of
sequence), forming higher-level discourse units. the data. For example, “at what age do I pay a
The term "nuclearity" in RST refers to which penalty for using retirement funds?” could be an-
text segment, fragment, or span, is more central to swered by the text (e.g., “age 59 ½”). We can de-
an author's purpose. A “nucleus” refers to a span termine informative text from a body of text and
of text that is more central to an author’s purpose such additional questions that can be answered
than a “satellite”, which is less central to the topic. from the body of text.
More particularly, we use the determined EDUs
of a discourse tree for a body of text and the rela-
tions between the EDUs to determine which
EDUs should be indexed for search. Different rhe-
torical relations (e.g., elaboration, contrast, etc.)
can employ different rules.
In general, we hypothesize that a satellite may
express a detail of information that is unlikely to
be explicitly queried by a user (Galitsky, 2015;
Jasinskaja and Karagjosova, 2017).
368
2
for the “same-unit” and “joint” relations, both the with the query via finding a maximal common
nucleus and the satellite are indexed. Different sub-parse tree (Galitsky, 2017).
rhetoric relations can have different rules, as
shown in Table 1 below. 3.1 Rule-Based Indexing
We take question-answer pairs and create, for
3 The Methodology of Question An- each answer, a discourse tree using RST-parser
swering (Surdeanu et al., 2015; Joty et al., 2013). For
each non-terminal node in each answer, we then
The developed methodology of the DT-based identify a rhetorical relationship associated with
analysis of answers is going to be applied in the the non-terminal node and label each terminal
following way, given an index of Q/A pairs: node associated with the non-terminal node as ei-
1. Search a user query against an index of ther a nucleus or a satellite. Then we apply a set
available questions; of rules (see Table 1) associated with the rhetori-
2. If no or too few results, generate addi- cal relationships and select, based on the rule,
tional search queries from the answers in- one or more of the fragment associated with the
dexed by proposed approach; nucleus or the fragment associated with the satel-
lite. Finally, we create a searchable index of ad-
3. If still no or too few results, search
ditional questions which includes multiple en-
against original answers.
tries corresponding to one of the selected frag-
We now focus on 2) and consider two meth- ments for the answers.
ods for indexing answers: rule-based and
classification-based. 3.2 Classification-Based Indexing
369
3
train the model since this algorithm is capable to Dataset Ques- Total # gener- Avg #
learn directly on a parse tree structure. tion/An # ated AQ words
swer / # sent
4 Datasets and Evaluation Yahoo! Q 3700 5.5 12.3
Answers
A 3700 8.1 124.1
We used a few datasets to evaluate the contribu-
tion of our methodology to search quality. Fidelity Q 500 3.4 6.2
Yahoo! Answer (Webscope, 2017) subset of A 500 6.2 118.0
question-answer pairs with broad topics where
main question is a single sentence (possibly, Car Re- Q 10000 4.2 5.5
pair
compound) with ten-fifteen keywords. The da- A 10000 7.0 141.3
taset includes various domains, and domain
knowledge coverage is shallow. Table 2: Dataset statistics
Financial questions1 scraped from Fideli- The proposed method delivers about 13 % im-
ty.com. This dataset demonstrates how search provements in the recall have the precision al-
relevance improvement may occur in a vertical most unaffected, for the Nucleus/Satellite rules.
domain with reasonable coverage. There is a further 3% improvement by using the
Car repair conversations2 selected from automated classifier of EDUs. Since the deploy-
www.2carpros.com including car problem de- ment of such classifier in a domain-dependent
scriptions and recommendation on how to rectify manner is associated with substantial efforts, it is
them. These pairs were extracted from dialogues not necessarily recommended when this 3% im-
as first and second utterances. provement in search accuracy is not critical.
For each search session, we only consider the We also compare performance of the proposed
first results and reject the others. For all these da- search on the extended framework derived from
tasets we assume that there is only one correct SQuAD 2.0 dataset (Rajpurkar et al., 2018) and
answer (from the Q/A pair) and the rest of an- applied to why? and how-to? questions. Instead
swers are incorrect. of addressing a question to a single Wikipedia
Evaluation results for the proposed methodol- text as standard evaluations do, we run them
ogy are presented in Table 3. Recall of the base- against all text. We use our approach vs neural
line search is on average 78% including the im- extractive reading comprehension one and ex-
provement by 8% by using syntactic generaliza- ceed recall of BiDaf (Gardner et al., 2017) and
tion on top of Lucene search (not shown). The DeepPavlov (Burtsev et al., 2018) by at least 8%
relevance of this system is determined by many with the search engine trained on our corpus
factors and is therefore not very insightful, so we (Table 4).
focus at the change in recall (), from this search Da- Baseline Nucleus Classifica-
system to the one extended by the proposed ap- taset / /Satellite tion-based,
proach. Meth rules, im- improve-
od provement ment
R P R, P, R, P,
% % % %
Ya- 79 74 +12. +0.1 +14 -0.04
hoo! 5
An-
swers
Fidel- 77 80 +10 -0.1 +6 +0.1
ity
Car 79 81 +16 +0.0 +18 +0.0
Re-
pair
1
https://github.com/bgalitsky/relevance-based-on-parse-
trees/examples/Fidelity_FAQs_AnswerAnatomyDataset1.cs Table 3: Evaluation results for new datasets
v.zip
2
https://github.com/bgalitsky/relevance-based-on-parse-
trees/examples/CarRepairData_AnswerAnatomyDataset2.csv.
zip.
370
4
Galitsky, B. 2017. Matching parse thickets for open
Method Recall Precision domain question answering. Data & Knowledge
Engineering, Volume 107, January 2017, Pages 24-
50.
BiDaf (AllenNLP) 68 71
Gardner, M., Grus, J., Neumann, M., Tafjord, O.,
Dasigi, P., Liu, N.H., Peters, M., Schmitz, M., &
DeepPavlov 67 72 Zettlemoyer, L.S. A Deep Semantic Natural Lan-
guage Processing Platform. arXiv:1803.07640.
Jansen, P., M. Surdeanu, and P. Clark. 2014. Dis-
Rule-based 75 71
course Complements Lexical Semantics for Non-
factoid Answer Reranking. ACL.
Classification-based 76 71 Jasinskaja, K., Karagjosova, E. 2017. Rhetorical Re-
lations: The Companion to Semantics. Oxford:
Wiley.
Table 4: Evaluation results for SQuAD dataset
Joty, Shafiq R, Giuseppe Carenini, Raymond T Ng,
5 Conclusions and Yashar Mehdad. 2013. Combining intra-and
multi- sentential rhetorical parsing for document-
In the search engines and chat bot industry, whole level discourse analysis. In ACL (1), pages 486–
texts are usually indexed for search. Because of 496.
that, frequently irrelevant answers are delivered
Kim, S., Bracewell, R., Wallace, K. 2004. From Dis-
because their insignificant keywords (the ones course Analysis to Answering Design Questions.
providing auxiliary information and not central for International Workshop on the Application of Lan-
the document) were matched. To overcome this guage and Semantic Technologies to support
well-known problem, only questions from Q/A Knowledge Management Processes (EKAW 2004).
pairs are indexed, which dramatically decreases At: Whittlebury Hall, Northamptonshire, UK.
the search recall. To address this limitation of in-
M. Burtsev, A. Seliverstov, R. Airapetyan, M.
dexing, we proposed and evaluated our approach
Arkhipov, D. Baymurzina, N. Bushkov, O.
of indexing only those EDUs of text which are de- Gureenkova, T. Khakhulin, Y. Kuratov, D. Kuz-
termined to be important (and therefore form al- netsov, A. Litinsky, V. Logacheva, A. Lymar, V.
ternative questions). This substantially improves Malykh, M. Petrov, V. Polulyakh, L. Pugachev, A.
the recall in applications such as FAQ search Sorokin, M. Vikhreva, M. Zaynutdinov. 2018.
where only questions of Q/A pairs are indexed. DeepPavlov: Open-Source Library for Dialogue
Systems. ACL-System Demonstrations. p. 122–127.
Acknowledgements
M. Sun, J. Y. Chai. 2007. Discourse Processing for
Section 3 (algorithms of question answering) Context Question Answering Based on Linguistic
were written by Dmitry Ilvovsky supported by Knowledge. Knowledge-Based Systems 20(6)(6):
the Russian Science Foundation under grant 17- 511-526
11-01294 and performed at National Research Mann, William and Sandra Thompson. 1988. Rhetori-
University Higher School of Economics, Russia. cal structure theory: Towards a functional theory
Section 4 (experimental investigations) was pre- of text organization. Text - Interdisciplinary Journal
pared within the framework of the HSE Univer- for the Study of Discourse, 8(3):243–281.
sity Basic Research Program and funded by the Moschitti, A. 2006. Efficient Convolution Kernels for
Russian Academic Excellence Project '5-100'. Dependency and Constituent Syntactic Trees. In
The rest of the paper were written and performed Proceedings of the 17th European Conference on
at Oracle Corp. Machine Learning, Berlin, Germany.
371
5
Structure Theory Parsers. Proceedings of the Con-
ference of the North American Chapter of the As-
sociation for Computational Linguistics - Human
Language Technologies: Software Demonstrations
(NAACL HLT), 2015.
Webscope 2017. Yahoo! Answers Dataset.
https://webscope.sandbox.yahoo.com/catalog.php?
datatype=l
Zong, H., Yu, Z., Guo, J., Xian, Y., Li, J. 2011. An an-
swer extraction method based on discourse struc-
ture and rank learning. 7th International Confer-
ence on Natural Language Processing and
Knowledge Engineering (NLP-KE).
372
6
Discourse-Based Approach to Involvement of Background Knowledge
for Question Answering
Boris Galitsky1 and Dmitry Ilvovsky2
1
Oracle Inc. Redwood Shores CA
2
National Research University Higher School of Economics
boris.galitsky@oracle.com; dilvovsky@hse.ru
373
Proceedings of Recent Advances in Natural Language Processing, pages 373–381,
Varna, Bulgaria, Sep 2–4 2019.
communicated (addressed) in augmented DT- boosting the weight of the topical part of a
A. document improves its estimation of relevance.
Hence instead of relying on an ontology that Regretfully these relations are relatively rare.
would have definitions of entities which are Sun and Chai (2007) investigated the role of
missing in a candidate answer we mine for the discourse processing and its implication on
rhetorical relations between these entities query expansion for a sequence of questions in
online. This procedure allows us to avoid an scenario-based context Q/A. They consider a
offline building of bulky and costly ontologies. sequence of questions as a mini discourse. An
At the same time, the proposed approach can empirical examination of three discourse
be implemented on top of a conventional theoretic models indicates that their discourse-
search engine. based approach can significantly improve Q/A
The paper structure is as follows. In Section performance over a baseline of plain reference
2 we compare the related work with our resolution.
proposal. In Section 3 we introduce the In a different task (Wang et al, 2010) authors
concept of a virtual discourse tree and present parse Web user forum threads to determine the
a number of examples illustrating how they can discourse dependencies between posts in order
be used and constructed. In Section 4 we to improve information access over Web forum
propose Q/A filtering algorithm which is the archives.
core part of our approach. In Section 5 we Suwandaratna and Perera (2010) present a
describe and discuss evaluation for the re-ranking approach for Web search that uses
question answering task on a few datasets that discourse structure. They report a heuristic
were compiled for this research. algorithm for refining search results based on
their rhetorical relations. Their implementation
2 Related Work and evaluation is partly based on a series of ad-
hoc choices, making it hard to compare with
2.1 Discourse and IR other approaches. They report a positive user-
Typically, every part in most coherent text has based evaluation of their system for ten test
some plausible reason for its presence, some cases.
function that it performs to the overall semantics Since rhetoric parsers for English (Joty et al.,
of the text. Rhetorical relations, e.g. contrast, 2013, Surdeanu, 2015) have become more
cause, explanation, describe how the parts of a available and accurate, their application in
text are linked to each other. Rhetorical relations search engine indexing is becoming more
indicate the different ways in which the parts of feasible. Precision and recall of search systems
a text are linked to each other to form a coherent ignore discourse level information and users do
whole. not find products, services and information
Marir and Haouam (2004) introduced a they need. It was shown that discourse features
thematic relationship between parts of text are valuable for passage re-ranking (Jansen et
using RST based on cue phrases to determine al., 2014). DTs have been also found to assist
the set of rhetorical relations. Once these in answer indexing to make search more
structures are determined, they are put in an relevant: query keyword should occur in
index, which can then be searched not only by nucleus rather than a satellite of a rhetoric
keywords, as traditional information retrieval relation (Galitsky et al., 2015). In this study we
systems do, but also by rhetorical relations. go beyond leveraging discourse features and
It was observed (Teufel and Moens, 2002) construct DTs from actual candidate answers
that different rhetorical relations perform and also virtual DTs for necessary background
differently across evaluation measures and knowledge.
query sets. The four rhetorical relations that 2.2 Discourse Analysis and Entities
improve performance over the baseline
consistently for all evaluation measures and At any point in the discourse, some entities are
query sets are: background, cause-result, considered more salient than others (occurring
condition and topic-comment. Topic-comment in nucleus parts of DTs), and consequently are
is one of the overall best-performing rhetorical expected to exhibit different properties. In
relations, which in simple terms means that Centering Theory (Poesio et al., 2004), entity
374
importance determines how they are realized in The next step for an A to be good for Q is to
an utterance, including pronominalized relation follow the logical flow of Q. Since it is hard to
between them. establish relations between entities E, which
Barzilay and Lapata (2008) automatically are domain dependent, we try to approximate
abstracts a text into a set of entity transition these relations by using logical flow of Q and
sequences and records distributional, syntactic, A, expressible in domain-independent terms,
and referential information about discourse such as rhetorical relation. Hence we require a
entities. The authors formulated the coherence certain correspondence between DT-Q and
assessment as a learning task and show that DT-A, considering additional labels for DT
their entity-based representation is well-suited nodes by entities (we denote such DT as EDT):
for ranking-based generation and text
classification tasks. EDT-Q ~ EDT-A. (2)
Nguyen and Joty (2017) presented a local
However a common case is that some
coherence model based on a convolutional
entities E are not explicitly mentioned in Q but
neural network that operates over the
instead are assumed. Moreover, some entities
distributed representation of entity transitions
in A used to answer Q do not occur in A but
in the grid representation of a text, can model
instead are substituted by more specific or
sufficiently long entity transitions and can
general entities do. How would we know that
incorporate entity-specific features without
these more specific entities are indeed
losing generalization power.
addressing issues from Q? We need some
Kuyten et al., (2015) developed a search
external, additional source which we call
engine that leverages the discourse structure in
virtual EDT-A to establish these relationships.
documents to overcome the limitations
This source contains the information on
associated with the bag-of-words document
inter-relationships between E which is omitted
representations in information retrieval. This
in Q and/or A but is assumed to be known by
system does not address the problem of
the interlocutor. For an automated Q/A system,
rhetoric coordination between Q and A, but
we want to obtain this knowledge at the
given a Q, this search engine can retrieve both
discourse level:
relevant A and individual statements from A
that describe some rhetorical relations to the EDT-Q ~ EDT-A + virtual EDT-A. (3)
query.
Our approach is to discover ontological 3.2 Discourse Trees for Answer and
relations between entities on the fly, finding Question
document fragments where a rhetorical relation We start with a simple example:
links these entities. Once all such text Q: What is an advantage of electric car?
fragments are found, we add the respective DT A: No need for gas.
fragments as virtual DTs to our main answer How can search engine figure out that A is a
DT. good one for Q? We have an abstract general-
sense entity advantage and a regular noun
3 Answering Questions via Discourse entity car. We need to link explicit entities in A
Trees {need, gas}. Fragments of a possible virtual
EDT-A are shown below:
3.1 Virtual Discourse Tree
Q: [When driving the cruise control][the
The baseline requirement for an A to be relevant engine will turn off][when I want to accelerate
to Q is that entities (E) of A cover the entities of ,][although the check engine light was off .] [I
Q: have turned on the ignition][and listen for the
engine pump running][to see][if it is building
E-Q E-A. (1) up vacuum .] [Could there be a problem with
the brake sensor under the dash ?] [Looks like
Naturally, some E-A (entities in an answer) there could be a little play in the plug.]
are not explicitly mentioned in Q but are
needed to provide a recommendation yielded
by Q (recipe-type A).
375
Figure 1: DTs of Q, A and imaginary DT-Aimg1 and DT-A img2
Figure 2: How Virtual DTs would enable Google search to explain missing keywords
376
Confidential Review Copy. DO NOT DISTRIBUTE.
A: [A faulty brake switch can effect the cruise in a domain independent manner: no offline
control .] [If it is,][there should be a ontology construction is required.
code][stored in the engine control module .] Search relevance is then measured as the
[Since it is not an emissions fault ,][the check inverse number of unaddressed E0 –Q once DT-A
engine light will not illuminate .] [First of all, is augmented with virtual DT-Aimg . This
watch the tachometer][to see][if engine speed relevance is then added to a default one.
increases 200 rpm][when this happens .] [If it Fig. 2 shows an example how Virtual DT
does ,][the torque converter is unlocking component would improve a web search.
transmission .] Currently, search engines show certain keywords
We do not need to know the details they do not identify in a given search result.
concerning how this Enablement occurs, we just However, it is possible to indicate how these
need evidence that these rhetorical links exist. keywords are relevant to the search result by
We could have used semantic linked between finding documents where these unidentified
entities but for that we would need a domain- keywords are rhetorically connected with the
specific ontology. ones occurring in the query. This feature would
Let us explain how a match between a Q and naturally improve the answer relevance on one
an A is facilitated by DTs (Fig. 1). A explains a hand and provide an “explainability” for the user
situation and also offer some interpretation, as
well as recommends a certain course of action. A Algorithm 1 Filtering Algorithm
introduces extra entities which are not in Q, and
needs to involve background knowledge to Input: Question
communicate how they are related to E-Q. We Parameter: Background knowledge B
do it by setting a correspondence between E-Q
and E-A, shown by the horizontal curly (red) Output: Most relevant Answer
arcs. 1: Build EDT-Q.
Notice that some entities E0 in Q are
unaddressed: they are not mentioned in A. E0-Q 2: Obtain E-Q
includes {Engine pump, Brake sensor and 3: Form a query for E-A
Vacuum}. It means that either A is not fully
relevant to Q omitting some of its entities E0 or it 4: Obtain a set of candidate answers As
uses some other entities instead. Are E0-Q 5: for each Ac in As do
ignored in A? To verify the latter possibility, we
6: Build discourse tree for the
need to apply some form of background answer DT-Ac.
knowledge finding entities Eimg which are linked
to both E0-Q and E-A. 7: Establish mapping E-Q E-Ac
It is unclear how E-A = Torque Convertor is 8: Identify E0 -Q.
connected to Q. To verify this connection, we
obtain a fragment of text from Wikipedia (or 9: Form queries from E0 –Q and E0 – Ac
another source) about Torque Convertor, build (entities which are not in E0 –Q)
DT-Aimg1 (shown on the left-bottom of Fig. 1) 10: Obtain search results from B for queries
and observe that it is connected with Engine via
rhetoric relation of elaboration. Hence we 11: Build imaginary DTs-Ac.
confirm that E-A = Torque Convertor is indeed 12: Calculate the score = |E0|
relevant for Q (a vertical blue arc).
13:end for
It is also unclear how E-Q pump is addressed
in Q. We find a document on the web about 14:Select A with the best score
Engine Pump and Vacuum and attempt to
15:return A
connect them to E-A. It turns out that DT-Aimg2
connects Vacuum and Engine via elaboration.
Hence the combined DT-A includes real DT-A on how her keywords are addressed in the
plus DT-Aimg1 and DT-Aimg2 . Both real and virtual answer. In the default search, munro is missing.
DTs are necessary to demonstrate that an answer However, by trying to rhetorically connect
is relevant by employing background knowledge munro with the entities in the question, the
Virtual DT approach finds out that Munro is an
377
5
Confidential Review Copy. DO NOT DISTRIBUTE.
2
https://github.com/bgalitsky/relevance-based-on-parse-
1
trees/examples/CarRepairData_AnswerAnatomyDataset2.csv
https://webscope.sandbox.yahoo.com/catalog.php?datatype=l .zip.
378
6
Confidential Review Copy. DO NOT DISTRIBUTE.
matched entities delivers more than 13% SQuAD and evaluated on the unseen questions.
performance boost. For the baseline we used Whereas a deep learning system gets 86% F1 on
standard implementation of Lucene search SQuAD 1.1, it achieves only 66% on SQuAD
engine based on matching keywords. 2.0 where some questions should not be
The bottom three rows show the Q/A quality answered.
when discourse analysis is applied. Assuring a Approach F1 Reference
rule-based correspondence between DT-A and
DT-Q gives 13% increase over the baseline, and BiDaf (Allen NLP) 64.8 Our experiments
(Gardner et al., 2017)
using virtual DT gives further 10%. Finally,
proceeding from rule-based to machine learned DeepPavlov 61.0 Our experiments
Q/A correspondence (SVM TK) gives the (Burtsev et al., 2018)
performance gain of about 7%. Microsoft Asia As reported by the
74.2
The difference between the best performing (Hu et al., 2018) authors (full dataset)
SVM TK for <EDT-Q EDT-A+EDT-Aimgi>
SVM TK for
row and the above row is only the machine <EDT-Q, EDT-A 73.3 Current study
learning algorithm: representation is the same. +EDT-Aimgi>
The bottom row shows the human evaluation
of Q/A on a reduced dataset of 200 questions for Table 2: Evaluation results on the SQuAD 2.0 dataset
each domain. We used human evaluation to
make sure the way we form the training dataset We applied the model trained on
reflects the Q/A relevance as perceived by Yahoo!Answers and Car Repair to our subset of
humans. This is important to confirm, in questions from SQuAD 2.0. Because most
particular, that the negative dataset includes unanswerable questions contain entities or entity
unsatisfactory answers. For a 1/3 fraction of this types which do not occur in text, virtual DT is a
dataset we measured Krippendorff’s alpha good means to handle such cases. At the same
measure for the inter-annotator agreement (two time, by the nature of neural learning, it is hard
annotators) which exceeds 80%. to learn to refuse to answer. The best
To summarize this experiment, the tree kernel performance on SQuAD 2.0 for the totality of
learning of virtual discourse trees turned out to questions, including much simpler ones than our
be a preferred approach. The contribution of formed 460 questions dataset, is achieved by (Hu
virtual DTs might be insignificant for simpler, et al., 2018) and exceeds our model by less than
shorter, factoid questions when traditional 1%.
measures of similarity between Q and A work The model of Hu et al. is specific to
well. However, we demonstrate that involvement Wikipedia pages and the way questions are
of background knowledge via virtual DTs for formulated, whereas our model learns once and
complex convergent questions requiring for all which discourse structures is correlated
entailment is significant. with which forms of background knowledge.
We believe this performance, achieved by
5.2 Experiments on a Standard Q/A training and testing on the same kind of Q/A
Dataset dataset is comparable with the results of the
We also compare the performance of virtual general model of the current study with the focus
DT Q/A with neural extractive reading on convergent why/how to questions.
comprehension approaches (Table 2). We made
this comparison on the why? and how-to? 6 Conclusions
questions from SQuAD 2.0 (Rajpurkar et al.,
Answering questions in the domain of this
2018), the dataset with unanswerable questions
study is a significantly more complex task than
which look similar to answerable ones. In total
factoid Q/A such as Stanford Q/A dataset, where
460 questions were selected.
it is just necessary to involve one or two entities
Deep learning systems can often locate the
and their parameters. To answer a “how to solve
correct answer to a question in a short text, but
a problem” question, one needs to maintain the
experience difficulties on questions for which
logical flow connecting the entities in the
the correct answer is not stated in the context.
questions. Since some entities from Q are
Rajpurkar et al. (2018) trained their system on
379
7
Confidential Review Copy. DO NOT DISTRIBUTE.
380
8
Confidential Review Copy. DO NOT DISTRIBUTE.
Christina Lioma, Birger Larsen, and Wei Lu. 2012. Pranav Rajpurkar, Zhang, Jian; Lopyrev, Konstantin;
Rhetorical relations for information retrieval. Liang, Percy. 2016. SQuAD: 100,000+ Questions
SIGIR, Portland, OR, pp. 931-940. for Machine Comprehension of Text. in EMNLP
2016.
D. T. Nguyen and S. Joty. A Neural Local Coherence
Model. 2017. ACL. pp. 1320-1330. Regina Barzilay and Mirella Lapata. Modeling local
coherence: An entity-based approach. Comput.
Gardner, M., Grus, J., Neumann, M., Tafjord, O.,
Linguist. 34, 1 (March 2008), 1-34.
Dasigi, P., Liu, N.H., Peters, M., Schmitz, M., &
Zettlemoyer, L.S. A Deep Semantic Natural S. Teufel and M. Moens. Summarizing scientific
Language Processing Platform. arXiv:1803.07640. articles: Experiments with relevance and rhetorical
status. Computational Linguistics, 28(4):409–445.
Jansen, P., M. Surdeanu, and Clark P. 2014. Discourse
2002.
Complements Lexical Semantics for Nonfactoid
Answer Reranking. ACL. Surdeanu, Mihai, Thomas Hicks, and Marco A.
Valenzuela-Escarcega. 2015. Two Practical
Joty, Shafiq R and A. Moschitti. Discriminative
Rhetorical Structure Theory Parsers. NAACL HLT.
Reranking of Discourse Parses Using Tree
Kernels. EMNLP 2014. Suwandaratna, N. and U. Perera. 2010. Discourse
marker based topic identification and search results
Joty, Shafiq R, Giuseppe Carenini, Raymond T Ng,
refining. In Information and Automation for
and Yashar Mehdad. 2013. Combining intra-and
Sustainability (ICIAFs), 2010 5th International
multi- sentential rhetorical parsing for document-
Conference on, pages 119–125.
level discourse analysis. In ACL (1), pages 486–
496. Wang, W., Su, J., Tan, C.L. 2010. Kernel Based
Discourse Rela-tion Recognition with Temporal
M. Burtsev, A. Seliverstov, R. Airapetyan, M.
Ordering Information. ACL.
Arkhipov, D. Baymurzina, N. Bushkov, O.
Gureenkova, T. Khakhulin, Y. Kuratov, D. William Mann and Sandra Thompson. 1988.
Kuznetsov, A. Litinsky, V. Logacheva, A. Lymar, Rhetorical Structure Theory: Toward a functional
V. Malykh, M. Petrov, V. Polulyakh, L. Pugachev, theory of text organization. Text 8(3):243–281.
A. Sorokin, M. Vikhreva, M. Zaynutdinov. 2018.
Yangfeng Ji and Noah Smith. 2017. A Neural
DeepPavlov: Open-Source Library for Dialogue
Discourse Structure for Text Categorization. ACL
Systems. ACL-System Demonstrations, p. 122–
2017.
127. 2018.
M. Sun and J. Y. Chai. 2017. Discourse processing for
context question answering based on linguistic
knowledge. Know.-Based Syst., 20:511–526,
August 2007.
Marir F. and K. Haouam. 2004. Rhetorical structure
theory for content-based indexing and retrieval of
Web documents, ITRE 2004, pp. 160-164.
Minghao Hu, Furu Wei, Yuxing Peng, Zhen Huang,
Nan Yang, Dongsheng Li. 2018. Read + Verify:
Machine Reading Comprehension with
Unanswerable Questions. arXiv:1808.05759
P. Kuyten, D. Bollegala, B. Hollerit, H. Prendinger
and K. Aizawa. 2015. A Discourse Search Engine
Based on Rhetorical Structure Theory. Advances in
Information Retrieva (ECIR). pp 80--91.
Poesio, M., R. Stevenson, B. Di Eugenio, and J.
Hitzeman. 2004. Centering: A parametric theory
and its instantiations. Computational Linguistics,
30(3):309–363.
Pranav Rajpurkar, Robin Jia, Percy Liang. Know
What You Don't Know: Unanswerable Questions
for SQuAD. arxiv.org/abs/1806.03822
381
9
On a Chatbot Providing Virtual Dialogues
1
382
Proceedings of Recent Advances in Natural Language Processing, pages 382–387,
Varna, Bulgaria, Sep 2–4 2019.
Conversion a text into a dialogue is different into a sequence of answers A = [A1…An] to form a
from the dialogue generation problem; the former dialogue
is a training set–based foundation for the latter. [A1, <Q1, A2>, …,<Qn-1, An>],
Response generation for dialogue can be viewed where Ai answers Qi-1 and possibly previous
as a source-to-target transduction problem. question, and Ai = T. Qi-1 needs to be derived
(Sordoni et al., 2015) rescores the outputs of a from the whole or a part of Ai by linguistic means
phrasal machine translation-based conversation and generalization; also some inventiveness may
system with a neural model incorporating prior be required to make these questions sound natural.
context. Recent progress in sequence-to-sequence To achieve it, we try to find a semantically similar
models has been leveraged to build end-to-end phrase on the web and merge it with the candidate
dialogue systems that firstly applies an utterance question.
message to a distributed vector representation The main foundation of our dialogue
using an encoder, then generates a response from construction algorithm is Rhetorical Structure
this representation. Theory (RST, Mann and Thompson, 1988). RST
(Li et al., 2016) simulate dialogues between represents the flow of entities in text via
two virtual agents, using policy gradient methods Discourse Tree – a hierarchical structure that sets
to reward sequences that display three useful inter-relations between text fragments: what
conversational properties: “informativity”, elaborates on what, what explains what, what is
coherence, and ease of answering. attributed to what, what is contradicting what, etc.
We measured comparable dialogue Such text fragments are called elementary
effectiveness, properties such as the speed of discourse units (EDUs). Most rhetorical relations
arrival to a search result, a decision and domain are binary anti-symmetric, specifying which EDU
coverage, in the current study. has more important (nucleus) compared to less
Dialogue acts are an important source which important (satellite).
differentiates between a plain text and a dialogue. A dialogue is formed from text by the
Proposed algorithm of virtual dialogues can assist following rule: once nucleus EDU is finished, and
with building domain-specific chatbot training before satellite EDU starts, questions against this
datasets. Recently released dataset, DailyDialog satellite EDU is inserted. In terms of dialogue
(Li et al., 2017), is the only dataset that has flow between a text author and a person asking
utterances annotated with dialogue acts and is question, the latter “interrupts” the author to ask
large enough for learning conversation models. his question such that the satellite EDU and
possibly consecutive text would be an answer to
3 Demo Description this question. The question is supposed to be
about the entity from the nucleus, but this nucleus
3.1 Dialogue Construction from Plain Text does not contain an answer to this question. The
To form a dialogue from text sharing person asking questions only interrupts the text
information or explaining how to do things, we author when his question sounds suitable; it does
need to split it into parts which will serve as not have to be asked for any nucleus-satellite
answers. Then for each answer a question needs to transition.
be formed. The cohesiveness of the resultant Once we split a text into EDUs, we know
dialogue should be assured by the integrity of the which text fragments will serve as answer to
original text; the questions are designed to questions: satellites of all relations. Elaboration
“interrupt” the speaker similar to how journalists rhetorical relation is default and What-question to
do interviews. a verb phrase is formed. Background relation
We employ a general mechanism of conversion yields another What-question for the satellite
of a paragraph of text of various styles and genres ‘…as <predicate>-<subject>’. Finally,
into a dialogue form. The paragraph is split into Attribution relation is a basis of “What/who is
text fragments serving as a set of answers, and source” question.
questions are automatically formed from some of A trivial approach to question generation would
these text fragments. The problem of building be to just convert satellite EDU into a question.
dialogue from text T is formulated as splitting it But it would make it too specific and unnatural,
such as ‘the linchpin of its strategy handled just a
2
383
small fraction of the tests then sold to whom?’. 3.2 System Architecture
Instead, a natural dialogue should be formed with System Architecture for building a dialogue from
more general questions like ‘What does its text is shown in Fig. 2. Each paragraph of a
strategy handle?’. document is converted into a dialogue via
An example of converting a text into a virtual building a communicative discourse tree for it and
dialogue is shown in Fig. 1. Answers are obtained then building questions from its Satellite
by splitting text into EDUs, and questions are Elementary Discourse Units. Current chatbot is a
inserted in text before satellite EDUs. Questions development of the previously built tool that
are shown in angle brackets and bolded. Each conducted task-oriented conventional dialogues
rhetorical relation in this example such as contrast (Galitsky et al., 2017).
ranges over a nucleus and a satellite. Each leave
of this discourse tree starts with ‘TEXT’ (Fig. 1).
Paragraph [A1, A2, .., An]
The reader of a virtual dialogue might feel that
the interviewer is guessing what the speaker is
Build CDT Form a list of Satellite EDUs
going to answer for each question. A discourse –
tree based approach does not deliver most natural
dialogues however it is a systematic method of Convert Satellite EDU into a generic question form
building ones without distorting the logical flow
of answers. Select the question focus: entity / attribute
3
384
and this information such as reviews should be 5 Sample Chatbot Session
convincing enough for making such decision;
We also measure how many entities (in We present an exploratory session that
linguistic sense) were explored during a session combines information delivery in the form of
with the chatbot. We are interested in how traditional textual answers (conventional
thorough and comprehensive the chatbot session dialogue) and also in a form of a virtual dialogue.
is, how much a user actually learns from it. This
assessment is sometimes opposite to the above Bot>ask a new question
two measures but is nevertheless important for User>advantages and new features of 5G
understanding the overall usability of various Bot>exploring….
conversational modes. I believe these are the main topics of your
query: is that what you meant? Please select
Table 1. Evaluation of comparative effectiveness of demonstrating the benefits of the technology[1].
conventional and virtual dialogues wide range of people from student'[2]. 5g new
Conventional radio nr specification'[3]. … next generation
Virtual dialogues
dialogues
mobile networks alliance[4]. the next stage in
mobile technology[5].
Coverюexploration #
# iterations till
# of iterations
# of iterations
Coverage of
till found
decision
decision
faster speed. …
Are you OK with this answer?
Conventional 4.6 6.3 10.8 - - - yes/more/no/specify [different topic] / reduce
only search to web domain / virtual dialogue
Virtual only - - - 4.1 6.0 13.7
User > virtual dialogue
Conventional 4.0 5.7 7.6 6.1 11.3 15.1
followed by Bot> This is what other people are chatting about
virtual your topic:
Virtual 5.6 7.1 12.3 3.7 7.0 11.5 User1>What Are the Benefits of Technology?
followed by Bot1> If you hopped over to Ericsson,
conventional you'd find a ''5G avenue'' with multiple exhibits
demonstrating the benefits of the technology…
We do not compare precision and recall of
User1>How are these benefits leveraged by
search sessions with either dialogue mode since
new tablets?
the same information is delivered, but in distinct
Bot1>Samsung kicked off its press conference
modes.
not with its new tablets, but 5G is the next
In the first and second rows, we assess the
wireless network in telecom industry …
stand-alone systems. One can observe that virtual
User2>Are the features right for them?
dialogues take less iteration on average for
Bot1>It is important for people to start learning
information access and about the same number of
about 5G technology now so that they can
iterations for decisions as conventional dialogues
decide if the features are right for them...
do.
User2 >Which are the dual 4G VoLTE sim
Notice that the bottom row, the chat scenario
smartphones in India?
proceeds from right to left. In the bottom two
Bot1> The current 4G VoLTE smartphones
rows, we observe the usability of the hybrid
with dual SIM slots that are available in the
system. When a conventional dialogue is followed
Indian markets may support 4G on single or
by a virtual one, a lower portion of users is
both SIM slots.
satisfied by the first step in comparison to the
Fig. 3. Chatbot session with conventional and virtual
inverse architecture, where virtual is followed by dialogue
conventional. The dialogue starts from the user question,
‘advantages and new features of 5G’. The chatbot
consults the sources (in this case, public URLs)
and extracts the content from each page (or
4
385
documents) expected to be relevant for the query. 6 Conclusions
In this example, seven URLs were processed,
from domain-specific to general knowledge We proposed a novel mode of chatbot
portals like Quora.com. The chatbot forms the list interaction via virtual dialogue. It addresses
of topics extracted from these search results so sparseness of dialogue data on one hand and
that the user might select one of his interest. convincingness, perceived authenticity of
Once the chatbot forms the topics for information presented via dialogues on the other
clarification of the user search intent, it shows hand. We quantitatively evaluated improvement of
them as a list. user satisfaction with virtual dialogue in
The user selects his topic of interest and comparison to regular chatbot replies and
requests a specific answer via the topic number of confirmed the strong points of the former. We
the topic expression. Once the answer is read, conclude that virtual dialogue is an important
there are multiple options: feature related to social search to be leveraged by
navigate to the next answer from the a chatbot.
chatbot list; Chatbot demo videos (please, check 10 min
video) and instructions on how to use it are
navigate to a specific answer from the available at
chatbot list; https://github.com/bgalitsky/relevance-based-on-
parse-trees in the “What is new?” section.
reject this answer and attempt to
reformulate the query; 7 Acknowledgements
reduce search to a specified web domain The article was prepared within the framework
(such as quota.com, for example); of the HSE University Basic Research Program
and funded by the Russian Academic Excellence
proceed in the same direction to more Project '5-100'.
search results in the form of a virtual
dialogue; References
Mann, William and Sandra Thompson. 1988.
accept the answer and conclude the session.
Rhetorical structure theory: Towards a functional
The user selects the last option and the chatbot theory of text organization. Text - Interdisciplinary
Journal for the Study of Discourse, 8(3):243–281.
builds a virtual dialogue. It is a conversation
between an imaginary people but the topic stays Joty, Shafiq R, Giuseppe Carenini, Raymond T Ng,
the same, matching the original query. Virtual and Yashar Mehdad. 2013. Combining intra-and
dialogues are shown in frames. As long as an multi- sentential rhetorical parsing for document-
level discourse analysis. In ACL (1), pages 486–
imaginary chatbot responds to the same person, 496.
the dialog is intended to stay cohesive;
coreferences in the follow-up questions are Galitsky, B, Ilvovsky, D. and Kuznetsov SO. 2015.
Rhetoric Map of an Answer to Compound Queries.
maintained. The main dialogue can be viewed as
ACL-2, 681–686.
a one in the meta-level, and the object-level
dialogue is naturally embedded into the meta-level Kipper, K. Korhonen, A., Ryant, N. and Palmer, M.
one. 2008. A large-scale classification of English verbs.
Language Resources and Evaluation Journal, 42,
Now the user can either browse the built virtual
pp. 21-40.
dialogue or search it to find a fragment of
conversation which is relevant to the user current Cox, R J. McKendree, R. Tobin, J. Lee, and T. Mayes.
exploration intent. The user now types the query Vicarious learning from dialogue and discourse: A
controlled comparison. Instructional Science,
‘Are the features right for me?’ and gets directed
27:431– 458, 1999.
to the virtual dialogue fragment where some other
users are discussing if the technology is ‘right for Alessandro Sordoni, Michel Galley, Michael Auli,
them’. The search matches the query either against Chris Brockett, Yangfeng Ji, Meg Mitchell, Jian-
Yun Nie, Jianfeng Gao, and Bill Dolan. 2015. A
the fragments of an original text, generated
neural network approach to context-sensitive
questions or both. generation of conversational responses. InProc. of
NAACL-HLT, May–June.
5
386
Craig, S, B. Gholson, M. Ventura, A. Graesser, and
the Tutoring Research Group. Overhearing
dialogues and monologues in virtual tutoring
sessions: Effects on questioning and vicarious
learning. International Journal of Artificial
Intelligence in Education, 11:242–253, 2000.
Li Y, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao,
and Shuzi Niu. 2017. Dailydialog: A manually
labelled multi-turn dialogue dataset. arXiv preprint
arXiv:1710.03957.
Piwek, Paul; Hernault, Hugo; Prendinger, Helmut and
Ishizuka, Mitsuru (2007). T2D: Generating
Dialogues Between Virtual Agents Automatically
from Text. Lecture Notes in Artificial Intelligence,
Springer, Berlin Heidelberg, pp. 161–174.
Mitkov R, L. A. Ha, and N. Karamanis. A computer-
aided environment for generating multiple-choice
test items. Natural Language Engineering: Special
Issue on using NLP for Educational Applications,
12(2):177–194, 2006.
Boris Galitsky and Dmitry Ilvovsky. 2017. Chatbot
with a discourse structure-driven dialogue
management. EACL System Demonstrations.
6
387
Assessing Socioeconomic Status of Twitter Users: A Survey
Dhouha Ghazouani1 , Luigi Lancieri2 , Habib Ounelli1 and Chaker Jebari3
1
Faculty of Sciences of Tunis, University Campus, Tunis 1060, Tunisia
(dhouha.ghazouani,habib.ounelli)@fst.utm.tn
2
Univ.Lille, Research Center in Signal and Automatic Computing of Lille, F-59000 Lille, France
luigi.lancieri@univ-lille.fr
3
Information Technology Department, Colleges of Applied Sciences, Ibri, Oman
jebarichaker@yahoo.fr
388
Proceedings of Recent Advances in Natural Language Processing, pages 388–398,
Varna, Bulgaria, Sep 2–4 2019.
important predictors of a person’s morbidity and work) and the material capital (Jones et al., 2007).
mortality experience (Kitagawa and Hauser, 1973; Similarly, the primary metrics of the SES are ed-
Marmot et al., 1987). The significant impact of ucation, occupation, income levels and wealth or
SES on public health renders its definition and lifestyle (Van Berkel-Van Schaik and Tax, 1990;
measurement of critical importance. When the White, 1982).
SES is low, it does not only involves poverty and People are usually divided into groups according
poor health, but it also affects the educational to these metrics, from the least advantaged to the
achievements hence the whole society. Thus, the most advantaged, medium, or high SES. Educa-
research of (Morgan et al., 2009) finds that chil- tion is one of the widely used indicator and it is
dren from low-SES households develop a slow considered by many to be the canonical element
academic behavior than children belonging to of SES because of its influence on later income
higher SES groups. For these reasons, marketing and occupation (Krieger et al., 1997). This in-
campaigns, as well as economic and sociological dex can be defined by two dimensions: the field
studies, have found it interesting to determine the of education and the level at which the education
socioeconomic status of particular persons. was followed. Income reflects spending power,
This article is going to be divided into six sections. housing, diet, and medical care. Occupation mea-
After the introduction, section 2 will discuss the sures like prestige, responsibility, physical activ-
metrics of socioeconomic status used with Twit- ity, as well as work exposures. The occupational
ter data. Section 3 will discuss SES indicators and status influences the social capital of individuals
its features. Section 4 will examine the different and it strengthen the connection with more pro-
techniques employed in SES inference and section fessional people enjoying wealth and power. Sim-
5 will present the data collection and analysis pro- ilarly, education indicates skills requisite for ac-
cess. Section 6 is going to be the conclusion for quiring a positive social, psychological, and eco-
this article opening the horizons for further dis- nomic resources (Antonovsky, 1967). Likewise,
cussions. social classes are measurements that, like SES,
aim to locate ones position in the social hierarchy.
2 Evaluation of Socioeconomic Status
classes are social categories sharing subjectively-
Socioeconomic status (SES) can be defined as salient attributes used by people to rank those cat-
one’s possession of social, financial, cultural, and egories within a system of economic stratification”
human capital resources. Parental and neighbor- (Wright and Ritzer, 2003). By refering to the def-
hood properties are considered as additional com- inition presented by (Wright and Ritzer, 2003),
ponents (Cowan et al., 2012). We can also note classes refer to how people are objectively located
that SES is a complex unit of measurement of a in distributions of material inequality”.
person’s economic and sociological standing, like Before the rise of social networks, different studies
for instance, his prestige, power or else his eco- have looked into other data sources from various
nomic well-being (Hoff et al., 2002; Oakes and domains, like internet browsing behaviors, written
Rossi, 2003). Consequently, one can conclude that texts, telephone conversations, real-world mobile
the SES is a complex measure of evaluation that network and communication records.
differs from a research to another because it takes
• (French, 1959): introduced the relationship
into account the work experience, the economic
between different measures of 232 under-
position, or the social status.
graduate students and their future jobs. This
The concept of the SES detection in literature goes
work concluded that occupational member-
back to the beginning of the 20th century (Chap-
ship could be predicted with the use of vari-
man and Sims, 1925). In the 20th century the eval-
ables such as the ability of persons in using
uation of SES was based on questions like: How
mathematical and verbal symbols, the social
many years did your father go to school?”, Do
class of family and the personality compo-
you have a telephone?” or Do you work out of
nents.
school hours?”. Currently, there is an agreement
that SES is influenced by three significant factors: • (Schmidt and Strauss, 1975): have also de-
the cultural (comprised of skills, capacities, and signed the relationship between the types of
knowledge), the social (social network combined occupation and the particular demographic
with the status and power of the people in that net- attributes such as gender, race, experience,
389
education, and location. Their study iden- They collect Twitter accounts which users are
tified biases and the types of discrimination located in Los Angeles. Concerning the sen-
that can possibly exist in various types of oc- timent analysis, they used SentiStrength 3 .
cupations. The study shows that people with higher in-
comes are associated with weaker social ties.
The recent excessive use of online social media
and the user-generated content in microblogging • (Quercia et al., 2012): treat the relationship
platforms such as google+, Facebook, or Sine between sentiment expressed in tweets and
Weibo 2 has allowed the study of author profiling the community socioeconomic well-being. In
on an unprecedented scale. their research, they collect Twitter accounts
which users are located in London. Concern-
• (Li et al., 2014): proposed a framework for
ing the sentiment analysis, they used word
assessing the user’s features on Twitter us-
count technique and the maximum entropy
ing Google+ API. They constructed a pub-
classifier. Socio-demographic data obtained
licly available dataset using distant supervi-
from Index of Multiple Deprivation scores
sion. They submitted their model on three
(composite score based on income, employ-
user profile attributes, i.e., Job, Spouse and
ment, education, health, crime, housing, and
Education.
the environmental quality for each commu-
• (Zhong et al., 2015): investigated the pre- nity) of each of the 78 census areas in Lon-
dictive power of location check-in, extracted don.
from points of interest of Sina Weibo. In or-
der to determine the demographic attributes 3 SES Features and Indicators
of the users such as education background The quality of features influences the value of
(university and non-university), marital status a machine learning pattern from which it origi-
(single, courtship, in love or married) using nates. Microblogging platforms offer a different
human mobility as an informative and funda- number of potential features. Different traditional
mental user behavior. They developed a com- text-based corpora features are used to explore the
prehensive location to profile (L2P) frame- relationship between these characteristics. Differ-
work to detect temporality, spatiality, and lo- ent types of indicators can help infer the SES of
cation knowledge at the same time. Twitter users used over years. This idea is going
to be developed later in the article.
• (Sullivan et al., 2018) has recently reported
that Facebook has patented technology that 3.1 Message Content
utilizes a sample decision tree to determine Twitter message text represents the backbone of
its users’ social class. Decision tree uses as most research works within the field of SES in-
an input information about a user’s demo- ference as this helps to understand the context of
graphic information, device ownership, inter- messages themselves. The messages of the social
net usage, household data, etc. The output media platform include abbreviations and non-
provides a probability that the user belongs to standard formulation as there is no precise rule of
a given socioeconomic class: working class, writing since most of the tweets are sent via mo-
middle class or upper class. bile phones.
In his thesis, (Mentink, 2016) used Bag-of-Words
Recent studies tackled the inference of socioeco-
to analyze the discussed topics of users. (Preoţiuc-
nomic characteristics of Twitter users.
Pietro et al., 2015b) used clustering algorithms to
• (Lerman et al., 2016): analyze a large cor- build a list of most frequent unigrams and then
pus of geo-referenced tweets posted by so- they reached their vector representations, conse-
cial media users from US metropolitan ar- quently using Word2Vec model to compute dense
eas. They measure emotions expressed in the word vectors (grouping words into clusters or top-
tweets posted from a particular area with the ics). While (Lampos et al., 2016) applied spec-
inference of socioeconomic characteristics. tral clustering to derive clusters of 1-gram that
2 3
http://www.weibo.com/signup/signup.php http://sentistrength.wlv.ac.uk/
390
capture some potential topics and linguistic ex- friends and the ratios of tweets to retweets are
pressions, (Preoţiuc-Pietro et al., 2015a) used Nor- considered as statistical features. (Lampos et al.,
malized Pointwise Mutual Information (NPMI) to 2016) use these features to compile a set of la-
compute word to word similarity, then applied sin- tent topics that Twitter users were communicating.
gular value decomposition (SVD) to obtain an em- (Culotta et al., 2016) use the Twitter REST API
bedding of words into a low-dimensional space. A and followers/ids request to sample many follow-
good approach of content analysis would take into ers for each account, and the results are ordered
consideration all possible instances of SES indica- with the most recent following first. And with the
tors being expressed within the message. For most same methods, they use friends/ids API request to
studies, the use of message content aims at infer- collect a list of friends. The example of (Barberá,
ring morphological characteristics and language 2016) best illustrates this idea it enables to over-
use. come the collection of information about the en-
(Barberá, 2016) used the emoji characters as fea- tire network of a particular user that is costly and
tures (bag-of-emoji) and the author used word requiring multiple API calls, focuses on verified
counts as another features (bag-of-words) with accounts. (Ikeda et al., 2013) use a community-
the application of TF-IDF transformation. In or- based method with the extraction of the commu-
der to obtain a robust result, the most successful nity from follower/followee relations followed by
techniques used employ message content initially, estimation of the demographics of the extracted
alongside other features. communities. The demographic category of each
community group is estimated using text-based
3.2 User Profiles method and the use of Fast Modularity Commu-
Although it must be admitted that in creating nity Structure Inference Algorithm. Some stud-
a new Twitter account, personal information are ies assume that people within a given social class
limited, however, they can give beneficial insights tend to have similar lifestyles using their income
for the SES of particular users. Users’ profiles levels and common experience. Their interaction
contain a different number of metadata such as the is called homophily. In the same context (Ale-
user’s biography, followers, name, and location. tras and Chamberlain, 2018) use the information
The expectation is that a user’s biography offers an extracted from the extended networks of Twitter
important source of demographic data. However, users in order to predict their occupational class
Twitter users’ biography is left empty for 48% of and their income. They demonstrated that user’s
users, and others do not supply good-quality in- social network and their language use are comple-
formation (Culotta et al., 2016). (Preoţiuc-Pietro mentary.
et al., 2015a) use the profile information of the
account to capture users with self-disclosed oc- 3.4 Spatial Information
cupations by annotating the user description field. The majority of smartphones are now equipped
(Lampos et al., 2016) use also profile description with Global Positioning System (GPS) functions
field of UK Twitter users to search for occupa- and they work with geo-satellites which accurately
tion mentions. In order to infer the user’s socioe- infer the user’s location with latitudes and longi-
conomic status, most studies use description field tudes coordinates. This would be an optional field
and attempt to search for related information given for a particular user to enable due to their privacy
by a particular user. These data are also useful in choice. This indicator is very helpful when the
order to validate other SES features inferred from person is mobile and usually updates their loca-
tweet messages. tion profile. (Bokányi et al., 2017) obtained 63
million of Twitter geolocated messages from the
3.3 Social Network Relations area of the United States and assigned a county
The followers of a user represent a good indica- to each tweet. Once aggregated, daily tweeting
tor of their SES. Following reciprocal relationship activity allows to measure human activities and
can provide evidence of strong user connection. constitutes an important socioeconomic indicator
Some indicators can group regular exchanges of whether a particular user is employed or not. In
messages or frequent mention to names in mes- order to build a social class dataset, some stud-
sages. The number of tweets, mentions, links, ies attempt to show that the wealthier the place,
hashtags and retweets, the number of followers, the richer the users who usually visit it. (Mi-
391
randa Filho et al., 2014) used the lifestyle and the ranging from different areas of study involving
wealth of neighborhood people typically visit to machine learning, statistics, natural language pro-
label Brazilian users into various social classes. cessing to regression models. Various methods
Then, they utilized Foursquare to label places ac- achieved different levels of success. The effec-
cording to the wealth of the neighborhood. They tiveness and granularity levels produced by these
selected users who had at least one Foursquare in- methods continue to be improved.
teraction (Foursquare interactions include check- Most recent researchers use a three-step method-
in (the user told a friend he/she was at a given ology to infer the SES. First, they collect available
place), tips (the user posts tips and opinion about information about a number of Twitter users. Sec-
a given place) and mayorship (title given to the ondly, they develop the classification method us-
most frequent user in a given location in the past ing additional data (number of followers, the con-
60 days). (Zhong et al., 2015) investigate the pre- tent of tweets). And finally, they classify users
dictive power of location and the mobility to in- who do not provide any concrete information ac-
fer users’ demographics with the use of location to cording to SES. (Preoţiuc-Pietro et al., 2015a)
profile (L2P) framework. The data crawling mod- for example, extracts occupation information from
ule accumulates user profiles and location check- Twitter user profiles and uses text analysis to cate-
in with corresponding information on Sine Weibo. gorize users into occupational classes.
In general, a common approach to demographic
3.5 Temporal Information inference is supervised classification, from a train-
Twitter enables researchers to analyze human ac- ing set of labeled users, a model is fit to predict
tivities during the 24 hours of the day because they user features from the content of their writings.
are biologically bound to exhibit daily periodic be- In other words, inferring user characteristics is
havior. In this context (Bokányi et al., 2017) ag- framed as a predictive task validated on held-out
gregate monday to friday relative tweeting activi- data. This is done by establishing regression or
ties for each hour in each US County to form an classification methods.
average workday activity pattern, assuming that
the activity patterns form a linear subspace of the 4.1 Regression Methods
24-hour ”time-space”. This study shows that this Various techniques for the inference of SES of
measure correlates with county employment and Twitter users have been adopted from data min-
unemployment rates in relation to lifestyles con- ing and machine learning techniques. Some stud-
nected to regular working hours. The relationship ies used the linear regression method, others used
between daily activity patterns and employment non-linear regression method and a third party
data can be captured using Twitter data. used a hybrid approach that combines both lin-
ear and non-linear methods. A standard non-linear
3.6 Demographic Attributes method does not inform which features are the
Some researchers attempted to include demo- most important in the predictive task. Then, the in-
graphics as features. Age, for example, has a vi- terpretability of linear methods allows performing
tal role in income prediction. Old people earn an extensive qualitative and quantitative analysis
significantly more than young ones. Higher age of the input features. (Flekova et al., 2016) used
leads to, on average, more work experience and both linear with Elastic Net regularization meth-
education, which is translated into higher income. ods and non-linear with Support Vector regression
(Flekova et al., 2016) explored the relationship be- together with an RBF kernel method. The authors
tween stylistic and syntactic features, authors’ age, found that machine learning regression methods
and income, to conclude that the hypothesis of nu- can be used to predict and analyze user’s income.
merous feature type writing style and age use is (Lampos et al., 2016) used a non-linear genera-
predictive of income. tive learning approach, which consists of Gaus-
4 Inference Methods for SES Evaluation sian Process (GP) and Kernel, to classify Twitter
on Twitter users according to SES as having upper, middle
Different techniques have been used in the past or lower level. Further, in (Preoţiuc-Pietro et al.,
and are being employed now to improve the ac- 2015b), the authors used similar methods to study
curacy of SES inference methodologies and algo- the user behavior and its power to predict income.
rithms. This burgeoning field lends techniques It is important to note that GPs is a Bayesian non-
392
parametric statistical framework that formulates sors within the population.
priority functions. (Hasanuzzaman et al., 2017)
used linear and non-linear methods. The linear 5.1 Data Corpuses and Ressources
method is a logistic regression with Elastic Net The corpus size of tweets grouped have varied
regularization. In order to capture the non-linear from relatively small datasets to as large as three
relationship between a user’s temporal orientation billion tweets (Mentink, 2016). The time span of
and their income, the authors used GP for regres- the data collected was usually in the range of a
sion. (Culotta et al., 2016) used a regression model few weeks to a couple of months, and sometimes
for the prediction in order to understand the demo- a year. Table 1 shows some datasets and their sizes
graphics of users. Due to the high dimensionality over the past years. First, the REST API is help-
of features, the authors used elastic net regulariza- ful for gathering particular user tweets, allowing
tion. Since each output variable consists of subject the backtracking of their timeline. For example,
categories of demographic characteristics. They to collect their most recent 3.200 tweets. Second,
used a multitask variant of the elastic net to ensure the streaming API that manages the tweets as they
the same features as selected for each category. are being broadcast would only be able to receive
4.2 Classification Methods 1% of the Firehose. Twitter data partners furnish a
premium service that supplies messages covering
(Mentink, 2016) employed two different ap- a longer duration as well as 100% access to the
proaches to classify the users in the dataset. The Firehose. (Preoţiuc-Pietro et al., 2015a) created
first is named the individual approach, it deter- a publicly available data-set 4 of users, including
mines the performing classifier per feature group their profile information and historical text content
and consequently combines them via a soft-voting as well as a label to their occupational class from
ensemble method. The second is named the the ”Standard Occupational Classification” taxon-
combined approach, it calculates the performance omy.
scores for all possible combinations of classi- This public available dataset used by several re-
fiers and their respective ensemble (also via soft- searchers containing a group of 5, 191 users in
voting). The author used Logistic Regression, total. However, the extraction of social network
Support Vector Machines, Naive Bayes and Ran- information of some accounts are not allowed.
dom Forest algorithms. The author runs the al- These accounts may have been annulled or be-
gorithm to determine what occupation and what come private. For example (Aletras and Cham-
education-level label should be given to a particu- berlain, 2018) reported results of 4, 625 users,
lar user, to overcome the data imbalance, noise and from the original subset, that are still publicly
bias, (Chen and Pei, 2014) used a typical imbal- available. Various studies (Preoţiuc-Pietro et al.,
ance classification approach which uses multiple 2015b; Lampos et al., 2016; Preoţiuc-Pietro et al.,
classifier systems (MCS) and a sampling method 2015a; Flekova et al., 2016) in the dataset cre-
which is a class-based random sampling method ation mapped Twitter users to their income or their
an extension of random under-sampling. The ob- job title using standardized job classification tax-
jective is to classify users according to their occu- onomy. The Standard Occupational Classification
pation. (Miranda Filho et al., 2014) evaluated a (SOC) is a UK-governmental system developed by
large number of classifiers using their WEKA ver- the Office of National Statistics (ONS) for listing
sion to generate classification models, including and grouping occupations. Jobs are organized hi-
multinomial Naive Bayes (MNB), Support Vector erarchically based on skill requirements and con-
Machine (SVM), and Random Forest. As MNB tent.
is more efficient than other algorithms, the authors (Culotta et al., 2016) mapped Twitter users accord-
used this method to infer social class for each par- ing to their educational level (No College, Col-
ticular user. lege, Grad School) and other traits using Quant-
5 Tweet Gathering and Analysis cast.com, an audience measurement society that
Messages on Twitter are publicly accessible in tracks the demographics of users of millions of
the online domain and can be gathered for study websites. The estimated demographics of a large
purposes. This availability makes Twitter an effi- number of sites are publicly accessible through the
cient tool in retrieving and analyzing public mes-
sages by allowing its users to become social sen- 4
https://sites.sas.upenn.edu/danielpr/data
393
References Corpus size Period Covered Corpus Origin
(Miranda Filho et al., 2014) 15.435 Users Sep’13-Oct’13 Brazilian
(Preoţiuc-Pietro et al., 2015b) 10.796.836 Aug’14 US
(Barberá, 2016) 1.000.000.000 Jul’13-May’14 US
(Lampos et al., 2016) 2.082.651 Feb’14-Mar’15 US
(Mentink, 2016) 3.000.000.000 Nov’14-Oct’15 Dutch
(Hu et al., 2016) 9.800 Users US
(Bokányi et al., 2017) 63.000.000 Jan’14 and Oct’14 US
(van Dalen et al., 2017) 2.700.000 Sep’16 Dutch
(Abitbol et al., 2018) 170.000.000 Jul’14-May’17 French
(Levy Abitbol et al., 2019) 90.369.215 Aug’14-Jul’15 French
use of searchable web interface. For each variable, as user profile features, users psycho-demographic
Quantcast gives the expected percentage of visi- features, or user emotion features, accuracy has
tors to a website with a given demographic. improved with the recent studies of (Mentink,
2016; Lampos et al., 2016) achieving a 75% ac-
5.2 Results and Metrics curacy.
The conclusions reached by different studies have
6 Conclusion and Future Prospectives
been significantly improved over time with re-
gards to increased accuracy and other measure- The study of socioeconomic status inference is
ments. Table 2 shows some techniques and their one of the most active field of information re-
results over the past years. This has been driven trieval. Such works are positioned at a crossroads
by improvements in algorithms and inclusion of of multiple disciplines.
more useful features. It is important to note that Different studies that introduced the inference of
the effectiveness and the reliability of occupation hidden user characteristics (Al Zamal et al., 2012;
representativeness increase when estimating pro- Miranda Filho et al., 2014; Volkova et al., 2015)
fession, using non-standard and out-of-vocabulary are salient in the field. The results of these works
(OOV) occupation names. In this context, to over- are not only of interest to statistics agencies but
come the limitation of the work of (Sloan et al., also necessary for studies in the social science
2015) and (Mac Kim et al., 2016), (Kim et al., (targeted advertising, personalized recommenda-
2016) builded a machine learning model attempts tions of user posts and the possibility of extracting
to capture linguistically noisy or open-ended oc- authoritative users (Pennacchiotti and Popescu,
cupations in Twitter. This induces in more reliable 2011)). It is important to introduce the role of
occupation representativeness. SES in politics such as the works of (Barberá and
Different approaches have been introduced to Rivero, 2015), (Burckhardt et al., 2016), (Kalsnes
compare the performance and results of the meth- et al., 2017), (Vargo and Hopp, 2017) and (Brown-
ods. They include accuracy and use of two other Iannuzzi et al., 2017). Twitter is increasingly con-
standard metrics: Pearson’s correlation coefficient sidered as politically transformative communica-
and Mean Absolute Error (MAE). To validate the tion technology that allows citizens and politi-
effectiveness of the approaches against different cians to connect, communicate, and interact easily
baselines, the k-fold cross validation has been well (Chadwick, 2006). The flaw in previous studies of
utilized for precision, recall and F-measure: the political behavior using Twitter data is the lack of
standard metrics for classification methods. information about the sociodemographic charac-
Over time, accuracy levels of results have contin- teristics of individual users. Policy makers have
ued to be improved starting from 2013 when the recently suggested introducing well-being com-
inference of users’ occupation was used. taking munity which will help governments do a better
into consideration other information such as Twit- job at directing public policy towards promoting
ter links, friends, user tweets, profiles, and other quality of life.
metadata associated with the message. Further- The inference of SES is an ambitious problem as
more, with the adoption of various features such it may belong to a combination of environmen-
394
References Technique Accuracy (%) Class
(Ikeda et al., 2013) Hybrid Method 71.60 Occupation
(Siswanto and Khodra, 2013) Machine Learning 77.00 Occupation
(Miranda Filho et al., 2014) Machine Learning 73.00 Social Class
(Preoţiuc-Pietro et al., 2015a) Gaussian Process 52.70 Occupation
(Mentink, 2016) Hybrid Method 75.00 SES
(Lampos et al., 2016) Gaussian Process 75.00 SES
(Poulston et al., 2016) SVM Classifier 50.47 SES
(van Dalen et al., 2017) Logistic Regression 72.00 Income
(Hasanuzzaman et al., 2017) Supervise Learning 74.40 Income
(Aletras and Chamberlain, 2018) Gaussian Process Regression (GPR) 50.44 Occupation
tal variables and individual characteristics. Some are important issues to address in future Twitter
of these characteristics can be easier determined socio-demographic inference studies. First, there
like gender or age while others, are sometimes is a need to look at the relationship between a
complicated with privacy issues and relying to user’s actual demographic characteristics and how
some degree on self-definition, are harder to deter- demographic categorization tools classify that user
mine like occupation, ethnicity, education level or as a function of how profile information is pre-
home location. Neverthless, there are many chal- sented and a virtual identity constructed. To con-
lenges in the inference of SES for Twitter users. clude, there is a need to link Twitter profiles and
Manual classification and data sampling are time- survey data. Researchers can start theorizing bet-
consuming, hard process and not scalable. Models ter working machine learning models to improve
are learned by referring to a datasets which were accuracy and scalability. In addition, the method-
manually labeled using Amazon Mechanical Turk ologies used in different research projects can be
at a high monetary cost. Another issue is that peo- coupled to increase efficiency. Another purpose
ple often misrepresent themselves on various on- for future research projects is to construct a less
line social platforms. This can lead to false data human effort, low computational cost and focus
interpretations which as a result can affect the ac- on the construction of a stronger evaluation frame-
curacy of the research. Automated detection tools work.
are based on the supposition that users will in-
troduce information on their demographic back- References
ground through profile information or metadata. Jacob Levy Abitbol, Márton Karsai, Jean-Philippe
While it is not possible to expect that all users do Magué, Jean-Pierre Chevrot, and Eric Fleury. 2018.
Socioeconomic dependencies of linguistic patterns
this, those who did were a random group of the in twitter: A multivariate analysis. In Proceedings of
Twitter population, then we would not expect to the 2018 World Wide Web Conference. International
discover conflicts in prevalence rates for sociode- World Wide Web Conferences Steering Committee,
mographic characteristics (Sloan et al., 2015). An- pages 1125–1134.
other problem is that Twitter data cannot represent Faiyaz Al Zamal, Wendy Liu, and Derek Ruths. 2012.
all the populace as discussed previously. (Sloan, Homophily and latent attribute inference: Infer-
2017) treated the issue of using human validation ring latent attributes of twitter users from neighbors.
ICWSM 270:2012.
to find the accuracy of methods applying profile
data to assign users to occupational groups and, Nikolaos Aletras and Benjamin Paul Chamberlain.
he deduced that this process could provide mis- 2018. Predicting twitter user socioeconomic at-
classifications due to users reporting their hobbies tributes with network and language information.
arXiv preprint arXiv:1804.04095 .
and interests rather than their actual occupations
(e.g, writer, artist). Another limit is that deriving Aaron Antonovsky. 1967. Social class, life expectancy
income statistics from job labels is not a suitable and overall mortality. The Milbank Memorial Fund
Quarterly 45(2):31–73.
method.
Given the findings presented above, the following Pablo Barberá. 2016. Less is more? how demographic
sample weights can improve public opinion esti-
395
mates based on twitter data. Technical report, Work- Temporal orientation of tweets for predicting in-
ing Paper. come of users. In Proceedings of the 55th Annual
Meeting of the Association for Computational Lin-
Pablo Barberá and Gonzalo Rivero. 2015. Understand- guistics (Volume 2: Short Papers). volume 2, pages
ing the political representativeness of twitter users. 659–665.
Social Science Computer Review 33(6):712–729.
Erika Hoff, Brett Laursen, Twila Tardif, et al. 2002.
Eszter Bokányi, Zoltán Lábszki, and Gábor Vattay. Socioeconomic status and parenting. Handbook of
2017. Prediction of employment and unemployment parenting Volume 2: Biology and ecology of parent-
rates from twitter daily rhythms in the us. EPJ Data ing 8(2):231–252.
Science 6(1):14.
Tianran Hu, Haoyuan Xiao, Jiebo Luo, and Thuy-
Jazmin L Brown-Iannuzzi, Kristjen B Lundberg, and vy Thi Nguyen. 2016. What the language you tweet
Stephanie McKee. 2017. The politics of socioeco- says about your occupation. In Tenth International
nomic status: how socioeconomic status may influ- AAAI Conference on Web and Social Media.
ence political attitudes and engagement. Current
Kazushi Ikeda, Gen Hattori, Chihiro Ono, Hideki
opinion in psychology 18:11–14.
Asoh, and Teruo Higashino. 2013. Twitter user pro-
Philipp Burckhardt, Raymond Duch, and Akitaka Mat- filing based on text and community mining for mar-
suo. 2016. Tweet as a tool for election forecast: Uk ket analysis. Knowledge-Based Systems 51:35–47.
2015. general election as an example. En: Third Rosie Jones, Ravi Kumar, Bo Pang, and Andrew
annual meefing of the Asian Polifical Methodology Tomkins. 2007. I know what you did last summer:
Society in Beijing . query logs and user privacy. In Proceedings of the
sixteenth ACM conference on Conference on infor-
Andrew Chadwick. 2006. Internet politics: States, cit- mation and knowledge management. ACM, pages
izens, and new communication technologies. New 909–914.
York, NY .
Bente Kalsnes, Anders Olof Larsson, and Gunn Sara
J Crosby Chapman and Verner Martin Sims. 1925. Enli. 2017. The social media logic of political inter-
The quantitative measurement of certain aspects of action: Exploring citizens and politicians relation-
socio-economic status. Journal of Educational Psy- ship on facebook and twitter. First Monday 22(2).
chology 16(6):380.
Sunghwan Mac Kim, Stephen Wan, and Cécile Paris.
Ying Chen and Bei Pei. 2014. Weakly-supervised oc- 2016. Occupational representativeness in twitter.
cupation detection for micro-blogging users. In Nat- In Proceedings of the 21st Australasian Document
ural Language Processing and Chinese Computing, Computing Symposium. ACM, pages 57–64.
Springer, pages 299–310.
Evelyn M Kitagawa and Philip M Hauser. 1973. Dif-
Charles D Cowan, Robert M Hauser, R Kominski, ferential mortality in the united states: A study in
Henry M Levin, S Lucas, S Morgan, and C Chap- socioeconomic epidemiology .
man. 2012. Improving the measurement of socioe-
conomic status for the national assessment of ed- Nancy Krieger, David R Williams, and Nancy E Moss.
ucational progress: A theoretical foundation. Na- 1997. Measuring social class in us public health
tional Center for Education Statistics. Retrieved research: concepts, methodologies, and guidelines.
from http://files. eric. ed. gov/fulltext/ED542101. pdf Annual review of public health 18(1):341–378.
. Vasileios Lampos, Nikolaos Aletras, Jens K Geyti, Bin
Zou, and Ingemar J Cox. 2016. Inferring the socioe-
Aron Culotta, Nirmal Kumar Ravi, and Jennifer Cutler.
conomic status of social media users based on be-
2016. Predicting twitter user demographics using
haviour and language. In European Conference on
distant supervision from website traffic data. Jour-
Information Retrieval. Springer, pages 689–695.
nal of Artificial Intelligence Research 55:389–408.
Kristina Lerman, Megha Arora, Luciano Gallegos,
Lucie Flekova, Daniel Preoţiuc-Pietro, and Lyle Ungar. Ponnurangam Kumaraguru, and David Garcia.
2016. Exploring stylistic variation with age and in- 2016. Emotions, demographics and sociability in
come on twitter. In Proceedings of the 54th Annual twitter interactions. In ICWSM. pages 201–210.
Meeting of the Association for Computational Lin-
guistics (Volume 2: Short Papers). volume 2, pages Jacob Levy Abitbol, Eric Fleury, and Márton Karsai.
313–319. 2019. Optimal proxy selection for socioeconomic
status inference on twitter. Complexity 2019.
Wendell L French. 1959. Can a man’s occupation
be predicted? Journal of Counseling Psychology Jiwei Li, Alan Ritter, and Eduard Hovy. 2014. Weakly
6(2):95. supervised user profile extraction from twitter. In
Proceedings of the 52nd Annual Meeting of the As-
Mohammed Hasanuzzaman, Sabyasachi Kamila, Man- sociation for Computational Linguistics (Volume 1:
deep Kaur, Sriparna Saha, and Asif Ekbal. 2017. Long Papers). volume 1, pages 165–174.
396
Yong Li, Mengjiong Qian, Depeng Jin, Pan Hui, and Daniele Quercia, Jonathan Ellis, Licia Capra, and Jon
Athanasios V Vasilakos. 2015. Revealing the effi- Crowcroft. 2012. Tracking gross community hap-
ciency of information diffusion in online social net- piness from tweets. In Proceedings of the ACM
works of microblog. Information Sciences 293:383– 2012 conference on computer supported cooperative
389. work. ACM, pages 965–968.
Sunghwan Mac Kim, Stephen Wan, Cécile Paris, Jin Alan Ritter, Sam Clark, Oren Etzioni, et al. 2011.
Brian, and Bella Robinson. 2016. The effects of data Named entity recognition in tweets: an experimental
collection methods in twitter. In Proceedings of the study. In Proceedings of the conference on empiri-
First Workshop on NLP and Computational Social cal methods in natural language processing. Asso-
Science. pages 86–91. ciation for Computational Linguistics, pages 1524–
1534.
Michael G Marmot, Manolis Kogevinas, and
Maryann A Elston. 1987. Social/economic Peter Schmidt and Robert P Strauss. 1975. The pre-
status and disease. Annual review of public health diction of occupation using multiple logit models.
8(1):111–135. International Economic Review pages 471–486.
Fons Mentink. 2016. Machine driven predictions of Elisafina Siswanto and Masayu Leylia Khodra. 2013.
the socio-economic status of Twitter users. Master’s Predicting latent attributes of twitter user by em-
thesis, University of Twente. ploying lexical features. In Information Technology
and Electrical Engineering (ICITEE), 2013 Interna-
Renato Miranda Filho, Guilherme R Borges, Jussara M tional Conference on. IEEE, pages 176–180.
Almeida, and Gisele L Pappa. 2014. Inferring user
social class in online social networks. In SNAKDD. Luke Sloan. 2017. Who tweets in the united kingdom?
pages 10–1. profiling the twitter population using the british so-
cial attitudes survey 2015. Social Media+ Society
Paul L Morgan, George Farkas, Marianne M Hille- 3(1):2056305117698981.
meier, and Steven Maczuga. 2009. Risk factors for
learning-related behavior problems at 24 months of Luke Sloan, Jeffrey Morgan, Pete Burnap, and
age: Population-based estimates. Journal of abnor- Matthew Williams. 2015. Who tweets? deriving the
mal child psychology 37(3):401. demographic characteristics of age, occupation and
social class from twitter user meta-data. PloS one
J Michael Oakes and Peter H Rossi. 2003. The mea- 10(3):e0115545.
surement of ses in health research: current practice
Aaron Smith and Joanna Brenner. 2012. Twitter use
and steps toward a new approach. Social science &
2012. Pew Internet & American Life Project 4.
medicine 56(4):769–784.
Brendan M Sullivan, Gopikrishna Karthikeyan, Zuli
Marco Pennacchiotti and Ana-Maria Popescu. 2011. Liu, Wouter Lode Paul Massa, and Mahima Gupta.
Democrats, republicans and starbucks afficionados: 2018. Socioeconomic group classification based on
user classification in twitter. In Proceedings of user features. US Patent App. 15/221,587.
the 17th ACM SIGKDD international conference on
Knowledge discovery and data mining. ACM, pages AB Van Berkel-Van Schaik and B Tax. 1990. Towards
430–438. a standard operationalisation of socioeconomic sta-
tus for epidemiological and socio-medical research.
Adam Poulston, Mark Stevenson, and Kalina Rijswijk: ministerie van WVC .
Bontcheva. 2016. User profiling with geo-located
posts and demographic data. In Proceedings of the Reinder Gerard van Dalen, Léon Redmar Melein, and
First Workshop on NLP and Computational Social Barbara Plank. 2017. Profiling dutch authors on
Science. pages 43–48. twitter: Discovering political preference and income
level. Computational Linguistics in the Netherlands
Daniel Preoţiuc-Pietro, Vasileios Lampos, and Niko- Journal 7:79–92.
laos Aletras. 2015a. An analysis of the user occupa-
tional class through twitter content. In Proceedings Chris J Vargo and Toby Hopp. 2017. Socioeconomic
of the 53rd Annual Meeting of the Association for status, social capital, and partisan polarity as predic-
Computational Linguistics and the 7th International tors of political incivility on twitter: a congressional
Joint Conference on Natural Language Processing district-level analysis. Social Science Computer Re-
(Volume 1: Long Papers). volume 1, pages 1754– view 35(1):10–32.
1764.
Svitlana Volkova and Yoram Bachrach. 2016. Inferring
Daniel Preoţiuc-Pietro, Svitlana Volkova, Vasileios perceived demographics from user emotional tone
Lampos, Yoram Bachrach, and Nikolaos Aletras. and user-environment emotional contrast. In Pro-
2015b. Studying user income through language, ceedings of the 54th Annual Meeting of the Associa-
behaviour and affect in social media. PloS one tion for Computational Linguistics (Volume 1: Long
10(9):e0138717. Papers). volume 1, pages 1567–1578.
397
Svitlana Volkova, Yoram Bachrach, Michael Arm-
strong, and Vijay Sharma. 2015. Inferring latent
user properties from texts published in social media.
In AAAI. pages 4296–4297.
Karl R White. 1982. The relation between socioeco-
nomic status and academic achievement. Psycho-
logical bulletin 91(3):461.
Erik Olin Wright and G Ritzer. 2003. Encyclopedia of
social theory.
398
Divide and Extract –
Disentangling Clause Splitting and Proposition Extraction
reserved for the role of the subject, in this case 2 Related Work
“The waitress”, while “at her friend” and “now”
are arguments, without further sub-specification. Proposition are relational tupels extracted from
Propositions are used in language understanding sentences in the form of predicate-argument struc-
tasks such as relation extraction (Riedel et al., 1
https://github.com/MeDarina/review_
2013; Petroni et al., 2015), information retrieval propositions
399
Proceedings of Recent Advances in Natural Language Processing, pages 399–408,
Varna, Bulgaria, Sep 2–4 2019.
tures (Marcus et al., 1994). There are proposition Sentence The waitress smiled at her friend now
Systems Subject Predicate Other Elements
models that further distinguish between the type
Allen The waitress smiled at her friend | now
of arguments. They do not only identify the sub- ClausIE The waitress smiled at her friend now
The waitress smiled now
ject, but more complex roles such as temporal and her has friend
locational objects or causal clauses. ReVerb The waitress now smiled at her friend
Stanford waitress smiled at her friend
Besides the theory and formalization of propo- waitress now smiled at her friend
sition, proposition extraction systems have perfor- OLLIE
OpenIE
The waitress
The waitress
now smiled at
smiled
her friend
now | at her friend
mance issues on real data. BL1 The waitress smiled at her friend now
BL2 The waitress smiled at her friend now
2.1 Comparison of Proposition Systems Us The waitress smiled at her friend | now
Although there have been comparative studies of Table 1: Output of Proposition Extraction Systems
proposition extraction systems, there has been no and Our Two Baselines for the Sentence The waitress
extensive study on the impact of sentence com- smiled at her friend now
plexity on proposition extraction system perfor-
mance.
Comparative Studies Niklaus et al. (2018) pre- Gemulla, 2013), and Stanford OIE (Angeli et al.,
sented an overview of proposition extraction sys- 2015) against their own system.
tems and classified them into the classic categories Stanovsky et al. (2018) evaluates ClausIE,
of learning-based, rule-based, and clause-based PropS (Stanovsky et al., 2016), and Open IE-4
approaches, as well as approaches capturing inter- against their new system, that we will call Allen
propositional relationships. They described the (Stanovsky et al., 2018) herein, using precision-
specific problems each system tackles as well as recall, area under the curve, and F1-score. They
gaps on the overall evolution of proposition ex- compare the individual proposition elements. For
traction systems. a proposition to be judged as correct, the predicate
Schneider et al. (2017) present a benchmark for and the syntactic heads of the arguments need to
analyzing errors in proposition extraction systems. be the same as the gold standard.
Their classes are wrong boundaries, redundant ex- Saha et al. (2018) evaluate ClausIE, OpenIE-4,
traction, wrong extraction, uninformative extrac- and CALMIE (a part of OpenIE) using precision.
tion, missing extraction, and out of scope. Their With the findings of this comparison, they intro-
pre-defined classes do not map directly to sentence duce a new version of their system, OpenIE-53 ,
complexity, although wrong boundaries and out of In all described comparisons, the system of the
scope would also be of some interest in an even respective authors is the best, which makes sense
more detailed error analysis. as it addresses the issue shown by the authors.
Furthermore, according to Stanovsky and Da-
gan (2016) and Niklaus et al. (2018) there are no 2.2 Propositions from Simple Sentences
common guidelines and followingly no gold stan- According to Saha et al. (2018) conjunctive sen-
dard defining a valid extraction. tences are one of the issues in proposition extrac-
Systems Table 1 shows the outputs from differ- tion, as conjunctions are a challenge to depen-
ent systems, our baselines, and our gold standard. dency parsers (Ficler and Goldberg, 2016) which
In their study, Gashteovski et al. (2017) aim at proposition extraction systems are mostly built
finding a system with minimal attributes, meaning upon. Hence, Saha et al. (2018) built a system that
that hedging2 and attributes expressed e.g. through automatically creates simple sentences from sen-
relative clauses or adjectives, can be optionally re- tences with several conjunctions that are used for
moved. Thus, they use recall and two kinds of proposition extraction. For the proposition extrac-
precision in the evaluation in order to account for tion of the simple sentences they used ClausIE and
the feature of minimality. To explain this in more OpenIE. They evaluated their data using three dif-
detail does not lie within the scope of this pa- ferent proposition datasets. The correctness of the
per. Gashteovski et al. (2017) evaluates OLLIE extracted proposition from the original sentence
(Mausam et al., 2012), ClausIE (Del Corro and were evaluated manually. In their study, simple
2
In pragmatics, hedging is a textual construction that sentences were sentences without conjunctions.
lessens the impact of an utterance. It is often expressed
3
through modal verbs, adjectives, or adverbs. http://knowitall.github.io/openie/
400
Quirk (1985) defines a simple sentence as a sen- propositions, even for humans, we included a pre-
tence consisting of exactly one independent clause liminary step of creating reduced sentences. A
that does not contain any further clause as one of reduced sentence is a sentence that contains only
its elements. Hence, a complex sentence consists a portion of the original sentence, e.g. the origi-
of more than one clause. This is also the definition nal sentence “The server was cool and served food
that we use in our study. and drinks” could be reduced to “The server was
cool” or “The server served food”. The inten-
2.3 Crowdsourcing Gold Standard tion behind this step was to create sentences with
Propositions one proposition only. Hence, the guidelines con-
Recent work used crowdsourcing for creating and tained rules such as decomposing conjunctive sen-
evaluating proposition extraction (Michael et al., tences or creating independent sentences from rel-
2018; FitzGerald et al., 2018) in the setting of ative clauses.5 We perform this preliminary step
question answering. In short, they asked their via crowdsourcing and evaluate it qualitatively.
crowdworkers to produce questions and answers Definition of Reduced Sentences We in-
in a way that resulted in the extraction of their structed our workers to produce reduced sentences
predicates and arguments, without directly asking from the original sentence. To prevent nested
for predicate-argument structures. structures, a reduced sentence was not allowed
to be split in further reduced sentences, at least
3 Corpus Creation within the output of one worker.6 Ideally, the
crowdworkers could have created sentences that
We create a corpus to evaluate the performance of contain exactly one proposition. However, this
proposition extraction systems entangled with and might even be a difficult task for experts, as there
disentangled from the task of clause splitting. are non-trivial sentence constructions that would
Our source corpus is the portion of the Aspect need long guidelines to create sentences with ex-
Based Sentiment Analysis (ABSA) task (Pontiki actly one proposition. However, our guidelines in-
et al., 2014) concerned with restaurant reviews sured that sentences were reduced in comparison
within one aspect – service. We use all 423 sen- to the original version, if possible. In this way,
tences that were annotated with this aspect. In we are able to create a sufficiently big set of both
a preliminary step, we produce a corpus of re- simple and more complex sentences, as shown in
duced sentences. To examine the influence of Table 2.
sentence complexity, we classify the reduced sen-
Crowdsourcing We used Amazon Turk for
tences as either 1) simple sentences, meaning sen-
crowdsourcing our data. Michael et al. (2018)
tences with potentially just one proposition, and
crowdsourced gold data for evaluating proposi-
2) complex sentences, meaning sentences with po-
tions. The sentence reduction performed here and
tentially multiple propositions. Then, we produce
also in Saha et al. (2018) is very similar to syn-
propositions from the reduced sentences using ex-
tactic sentence simplification as performed by Lee
pert annotation and evaluate it by calculating the
and Don (2017). We paid 0.04 $ per HIT and
inter-annotator agreement.
0.01 $ for each further reduced sentence. Each
Our corpus contains 2,181 sentences (class dis- sentence was reduced by 3 workers. In this pro-
tribution in Table 2) and 2,526 propositions. cess, 2181 unique reduced sentences, which are
all used in the following corpus creation process,
3.1 Preliminary Step: Creating Reduced
were created from 423 original sentences.
Sentences
Evaluation of Reduced Sentences To mea-
As a preliminary step, we created a gold corpus sure the quality of the crowdsourced reduced sen-
of reduced sentences formed from originally more tences, we chose 100 random reduced sentences
complex sentences. together with their original sentence and evalu-
To do so, we use 423 sentences from review
texts4 . As these are quite difficult for producing 5
However, this step turned out to be more difficult than
expected, as some sentences contained several factors that
4 could be reduced. However, this did not influence our goal
Online users’ restaurant reviews are a fruitful domain
for proposition extraction, as propositions extracted from re- of determining the influence of sentence complexity.
6
views would be useful for several user-centered tasks, as they The annotation instructions are also available on our
would allow to display only information pieces of interest. Github page.
401
Complexity Class # of Occurrences it. In our notation, the first position of the proposi-
tion is the subject, the second is the predicate and
No Verb 101
the order of the other elements is irrelevant.7
Simple 1,648
The arguments may also contain further propo-
Complex 432
sitions, e.g. here, the sentence “I think their food
All 2,181
is great” is split in two propositions – “I | think |
their food is great” and “their food | is | great ”.
Table 2: Distribution of Sentence Complexity Classes
in Our Reduced Sentence Set This definition is restrictive in that it asks for ex-
actly two propositions in the given example. Addi-
tionally, it is not bound to a clearly defined theory
ated their correctness using the following non- (as there is no clearly defined theory on proposi-
exclusive categories: O RIGINAL S IMPLE, R E - tions). However, it is the representation that is
DUCED , S IMPLE , G RAMMAR , and I NFERENCE needed to extract information from reviews, as it
(see Table 3b). would help to reduce redundancies, e.g. by clus-
In Table 3a, we provide an exemplary sentence for tering sentences such as “Their food is great” and
each category, except for O RIGINAL S IMPLE, as “I think their food is great”. Furthermore, we are
it means that the original is already a simple sen- not interested in inferred information, e.g. “They
tence, containing only one proposition which can- | have | food” from the previously discussed sen-
not be further reduced. 20 sentences in the random tence. This choice will also be reflected in the per-
sample were categorized as being O RIGINAL S IM - formance of systems that do not adhere to our un-
PLE . However, some workers still tried to reduce derstanding of propositions. However, this does
some of these sentences – 2 of them were gram- not necessarily cloud the performance compari-
matically incorrect (G RAMMAR) and 3 fell into son of simple and complex sentences, as we will
the class I NFERENCE. This means that their con- still measure the influence of sentence complexity.
tent was not explicitly mentioned in the original Each sentence is processed by two annotators and
sentence, but was lexically inferred. the disagreements are curated in a subsequent step.
There were 66 R EDUCED sentences, meaning that Creation As the creation of propositions is not
the sentences have been successfully reduced. 60 a trivial task, due to many different cases that need
of the R EDUCED resulted in S IMPLE sentences, to be explained in the guidelines8 , this task should
which means that they contained only one propo- be performed by people who were trained longer
sition after the reduction, and 6 were simpler than than a crowdsourcing platform allows for. Thus,
the original sentence, but contained more than one we produced proposition annotations in a double-
proposition. annotation process by three graduate students9 .
We believe that the results are usable as is, as The disagreements were curated by the first author
the error rate is quite low – only 17 of the re- of the paper. The result of the curation builds the
duced sentences in the random sample were in- gold standard. The gold standard, all annotations,
correct (G RAMMAR and I NFERENCE), as many of and the guidelines are available.
the G RAMMAR errors stem from the original sen-
tence. Furthermore, we show that our reduction 3.3 Evaluation of Proposition Creation
step was necessary to produce enough simple sen- To evaluate our dataset, we report inter-annotator
tences for our experiment, as 80% of the random agreement as well as agreement with the curator
sample were originally complex.
7
We are not interested in different types of objects and
3.2 Creating Propositions from Simple modifiers, similar to Stanford, OpenIE, and AllenNLP, and
Sentences thus we do not discuss this information. For a better
overview, we asked the annotators to present the other ele-
To evaluate the performance of proposition extrac- ments in their order of occurrence.
8
tion systems, we created a gold standard corpus for The guidelines include explanations of what predicates,
arguments, and nested propositions are. This in itself is not
propositions from the reduced sentences. difficult. However, such instructions consume more time and
In this paper, we follow the most simple possi- need more training, as simple mistakes are made by untrained
ble annotation, similar to Stanovsky et al. (2018). annotators. We saw this in a training set for this task, that is
not included or discussed here due to space restrictions.
We want to extract English propositions with 9
The result is shown in Table 4. A1 annotated the whole
one main verb and all arguments that are linked to set, while A2 and A3 annotated parts.
402
The server was cool and served food Sentence Class #
Original Sentence
and drinks.
O RIGINAL S IMPLE 20
R EDUCED The server was cool and served food. R EDUCED 66
S IMPLE The server was cool. S IMPLE 87
G RAMMAR The server was. G RAMMAR 5
I NFERENCE The server is good. I NFERENCE 12
(a) Classification Examples (b) Distribution of Classification
on both the proposition (see Table 4a) and propo- proposition element level is .67 and .7 on complex
sition element level (see Table 4b). sentences and .83 and .85 for simple sentences -
Evaluation Metric In order to see differences nearly double of the whole proposition agreement.
in the annotation, we performed inter-annotator
agreement using %-agreement (accuracy). We use Simple Complex All
the same measure for system performance, which A1 Gold A1 Gold A1 Gold
enables a direct comparison. Although we are A1 - .80 - .66 - .76
aware that agreement is ignorant of chance agree- A2 .71 .79 .53 .63 .66 .74
ment, we believe that it is the best measure for A3 .61 .66 .39 .48 .57 .62
this problem, as chance agreement is quite low
in the case of this complex annotation problem. (a) Inter-Annotator Agreement on Propositions
Furthermore, it is difficult to interpret these results Simple Complex All
in comparison to other works. As previously de- A1 Gold A1 Gold A1 Gold
scribed, there are no clear guidelines for propo- A1 - .90 - .77 - .86
sitions and also no manual gold datasets created A2 .85 .79 .70 .63 .81 .74
explicitly for this purpose. We could compare the A3 .83 .83 .67 .70 .80 .80
results of our inter-annotator agreement to simi-
lar tasks, where sentences are split into compo- (b) Inter-Annotator Agreement on Proposition Elements
nents, as e.g. answers prepared for question an-
Table 4: Inter-Annotator Agreement in Accuracy
swering, paraphrase alignment, translation align-
ment etc. However, they also have different setups
and evaluation metrics and it is out of the scope of 4 Evaluation of Proposition Extraction
this work to discuss these differences. Systems
Levels of Evaluation We perform the evalua-
Similar to Saha et al. (2018); Schneider et al.
tion on two levels - proposition level and propo-
(2017) and Niklaus et al. (2018), we evaluate
sition element level. On the proposition level,
proposition system performance. They do not,
we calculate the agreement of whole propositions.
however, regard the task of proposition extraction
On the proposition element level we calculate the
disentangled from the intrinsic subtask of clause
agreement of individual elements of the proposi-
splitting. By showing the performance of both
tions whilst taking their label (subject, predicate,
simple and complex sentences, we are furthermore
or other element) into account.
able to show the impact of clause splitting.
Inter-annotator Agreement Table 4a shows
that the inter-annotator agreement on the propo- 4.1 Setup
sition level is .39 and .53 on complex sentences To identify the system that performs best when
and .61 and .71 on simple sentences. These agree- disentangled from the task of clause splitting,
ment differences show that clause splitting is also we use the herein produced corpus to analyze
difficult for humans. and evaluate the performance of various proposi-
Agreement with Curator The agreement with tion extraction systems as used in evaluations by
the curator is .05 to .19 higher than the inter- Stanovsky and Dagan (2016), Gashteovski et al.
annotator agreement. The agreement on the (2017), Saha et al. (2018), and Stanovsky et al.
403
(2018). Hence, we will analyze proposition ex- is much better than on proposition level. Further-
traction performance using AllenNLP, ClausIE, more, the table shows that for all systems but Re-
ReVerb, Stanford Open Information Extraction, Verb, the performance is much better on the simple
OLLIE, and OpenIE-5.10 Furthermore, we pro- sentences, which was expected.
vide two baseline systems. It is also interesting that although the perfor-
We use %-agreement to measure the perfor- mance of both baselines on whole propositions
mance of systems. We want full agreement, is 0, the performance of the second baseline on
not just matching phrase heads, as performed by proposition elements is competitive. This shows,
Stanovsky et al. (2018). Furthermore, we evaluate that the task of proposition extraction can, to a
only agreement, as in our setup the argument or big part, be solved by correct verb extraction. It
the predicate matching is what we are interested outperforms ReVerb, Stanford, and on simple and
in, meaning we do not need precision and recall complex sentences also OLLIE. The second base-
in our setting. In this way, our evaluation setup is line performs a little worse on all sentences, as
similar to Saha et al. (2018), who also identified these also include sentences without a verb and
specific issues in proposition extraction systems. this baseline is verb-based. This shows that either
As in inter-annotator agreement, we calculate the automatic systems have problems with the ex-
agreement on two levels: proposition and proposi- traction of verbs or they have deeper issues, e.g.
tion element level. The results of the performance they do not extract from a lot of sentences, as is
comparison is shown in Table 5. discussed in Section 4.4.1. The second baseline
performs almost equally on both simple and com-
4.2 Baselines plex sentences. This may show correct verb ex-
We provide two baselines in order to better com- traction alone solves only a particular portion of
pare the systems. Both baselines create proposi- proposition extraction.
tions with three elements at most: subject, pred- Other systems, especially the two best ones,
icate, and one other element. The first baseline perform about two times better on the simple sen-
(BL1) takes the first word as subject, the second tences but then have a much bigger drop on the
word as predicate and the rest as one other ele- complex sentences. This may show that clause
ment. The second baseline (BL2) is a little more splitting has a bigger impact on better or proba-
engineered and uses POS-tags. It makes a propo- bly more intelligent systems than on more simple
sition for each verb. All words before the verb systems.
are the subject and all words after the verb are On both levels, OpenIE is the best system, very
one other element. Examples for the baselines are closely followed by Allen, whereas the other sys-
shown in the Table 1. The baselines are kept sim- tems are well-beaten.
ple on purpose to show how simple algorithms can
solve the given problem. A baseline that appears 4.4 Analysis of System Performance
intuitive is using a dependency parser and filtering Identifying further problems except clause split-
for the root and its dependants. However, decid- ting could improve current proposition extraction
ing which parts are its dependents and especially systems. On the one hand there are sub-issues in
the span of arguments is ambiguous. This would clause splitting. On the other hand, there are issues
not be a baseline, it would be a rule-based system besides clause splitting.
that is not out of the box. Hence, we decided not In the case of ClausIE and ReVerb, many further
to do it. clauses and also arguments are cut, as these consist
of a maximum of three elements, which makes the
4.3 System Performance comparison difficult.
Table 5a shows that performance of proposition
4.4.1 General Issues
extraction on whole propositions is equally bad
for both simple and complex sentences. Table 5b We first manually examined some potential is-
shows that performance on proposition elements sues in the proposition extraction from simple sen-
tences. After the manual analysis of potential is-
10
We will not use MinIE (Gashteovski et al., 2017), as it sues, we calculated the system performance if the
is an extension of ClausIE providing additional information
such as modality and whether an argument is necessary or issue would be eliminated. One big issue we found
unnecessary, which is disregarded in this work. is missing propositions, meaning that systems do
404
not always extract propositions. Except for the
missing propositions, there was no big difference
Systems Simple Complex All
in the system performance with or without the is-
Allen .08 .09 .08 sue. Also, some systems have different models
ClausIE .06 .09 .07 of propositions, which may also affect their per-
ReVerb .02 .02 .02 formance. On the one hand, there are issues with
Stanford .01 .01 .01 previous steps, e.g. negations or quantifiers are
OLLIE .03 .04 .03 ignored. On the other hand, there are issues with
OpenIE .09 .12 .09 formatting, e.g. a different treatment of preposi-
BL1 .00 .00 .00 tions or conditionals.
BL2 .00 .00 .00 Missing Propositions One big issue is that propo-
(a) System Performance on Propositions sition extraction systems often do not produce any
extraction from a sentence. Unsurprisingly, this
Systems Simple Complex All issue is bigger among the systems that do not per-
Allen .50 .40 .46 form well - namely ReVerb (58% of sentences
ClausIE .37 .36 .36 do not have an extraction), Stanford (39%), and
ReVerb .15 .14 .14 OLLIE (33%), whereas the better performing sys-
Stanford .20 .09 .17 tems have much lower rates - Allen (3%), ClausIE
OLLIE .24 .19 .22 (4%), and OpenIE (10%). In ReVerb, Stanford,
OpenIE .51 .42 .47 and OLLIE we could not find a clear reason why
BL1 .05 .04 .05 there are no extractions. In the case of Allen,
BL2 .26 .24 .21 there are only no extractions from sentences with-
out verbs.11 ClausIE and OpenIE have no extrac-
(b) System Performance on Proposition Elements
tions from sentences that are missing a verb or a
Table 5: System Performance Measured in Accuracy subject. Additionally, OpenIE has no extractions
from existential clauses.
In Table 6a, where we show the performances
of systems on full propositions without the dis-
Systems Missing Conditional Temporal cussed issues, it is shown that systems perform
slightly better when eliminating missing proposi-
Allen .08 .13 .19 tions from simple sentences. However, the im-
ClausIE .06 .11 .13 provement is clearer in Table 6b on the element
ReVerb .03 .00 .03 level. Especially for the systems that had more
Stanford .02 .00 .00 missing propositions, namely Stanford, ReVerb,
OLLIE .04 .06 .02 and OLLIE, the change is between .06 - .17.
OpenIE .10 .19 .17 Conjunctions As already stated by Saha et al.
(a) System Performance on Propositions Excluding Specific Is- (2018), conjunctive sentences pose an issue to
sues proposition extraction systems. In our case, we
Systems Missing Conditional Temporal wanted to separate all conjunctive sentences in in-
dividual propositions, e.g. the sentence “The wait-
Allen .50 .57 .55
ress smiled at her friend and at me.” contains the
ClausIE .38 .40 .38
propositions “The waitress | smiled | at her friend”
Stanford .26 .03 .14
and “The waitress | smiled | at me.”. OpenIE
ReVerb .32 .00 .21
and Stanford have the same guidelines on conjunc-
OLLIE .31 .00 .20
tions, whereas Allen, ClausIE, and ReVerb keep
OpenIE .54 .53 .50
the conjuncted elements together – from the pre-
(b) System Performance on Proposition Elements Excluding vious sentence they would create one proposition
Specific Issues – “The waitress | smiled | at her friend and me.”.
Table 6: System Performance Excluding Specific Is- Negations Stanford does not extract from negated
sues
11
These sentences are classified as neither simple nor com-
plex, but are included in all.
405
sentences and Allen has problems with negated On the one hand, we showed that sentence
sentences missing a verb. The rest can deal with complexity has a measurable impact on proposi-
negations. These specific problems are difficult to tion extraction performance of both humans and
show in numbers, as they are rare – only about 7% machines. Hence, one step towards improving
of the sentences contained negations. the performance of such systems, is the improve-
Prepositions OLLIE, ReVerb, and Stanford place ment of clause splitting. Furthermore, we be-
the prepositions with the predicate, whereas all lieve that the performance of the original complex
other systems as well as our gold standard place sentences, without the preliminary reduction step,
it with the associated argument, as is shown in the would pose an even bigger problem to proposition
example in Table 1. For these cases we would need systems, which implies that using these systems
adjusted evaluations that ignore this difference. on real data could be problematic.
Quantifiers Stanford ignores “every” in proposi- On the other hand, our study also showed that
tions. the ranking of systems is similar among simple
and complex sentences. This means, that the best
4.4.2 Issues with Complex Sentences
performing systems among simple sentences that
We looked at issues within complex clauses, are disentangled from the task of clause splitting,
namely conditional and temporal clauses. are also the best in complex sentences, where
Conditional Clauses In some cases, Allen, clause splitting also needs to be performed. This
ClausIE, OLLIE, and OpenIE extract the if-clause may mean that to find the overall best system, one
for the argument, but delete the “if”, which leads does not need to classify between simple and com-
to disagreements on both full proposition and plex sentences. However, it is necessary to find
proposition element level. Comparing the perfor- that sentence complexity is one problem of propo-
mance on all complex clauses as shown in Table 5a sition extraction.
to complex clauses without conditional clauses, as Also, our intelligent baseline system, that was
shown in Table 6a, all systems, except for Re- able to extract verbs, outperformed three of the
Verb and Stanford, clearly perform better. Allen systems. However, the better systems did not only
is better by .04 and OpenIE by .05, which shows perform much better, but they were also more af-
that they have the biggest issues with conditional fected by sentence complexity.
clauses. On proposition element level this be- Additionally, we looked into further problems
comes even clearer. Here, the three better systems, of proposition extraction systems. The main issues
ClausIE, Allen, and OpenIE perform .04 - .17 bet- in complex sentences that we could identify were
ter without conditional clauses. conditional and temporal clauses.
Temporal Clauses Conceptually, Allen, OLLIE,
and OpenIE extract temporal clauses correctly, 6 Future Work
but have some problems if the sentence is too
long. Stanford cuts out the “when”. For temporal In future work, we plan to enlarge the corpus in
clauses, the performance is similar to conditional order to use it for studies on user-specific recom-
clauses. The three better systems perform .06 -.11 mendations. We plan to display proposition-like
better on full proposition level, and .02-.09 better information to the user to provide more specific
on proposition element level. Stanford and OLLIE information than is given by a long sentence. This
perform worse without the temporal clauses. work may help in clause splitting, as we not only
provide a gold standard for it, but also describe a
5 Summary method on how to create it. Furthermore, we plan
to built a proposition extraction system based on
In this work, we described a method on how to cre-
the findings from this paper.
ate a dataset of reduced sentences from originally
complex ones. We created an English dataset ac-
Acknowledgement
cording to this method and further classified this
dataset as simple and complex. It can be used for We would like to thank Venelin Kovatchev and
further evaluation of proposition extraction sys- Marie Bexte for their annotations. This work
tems. The dataset enabled us to research the per- has been funded by Deutsche Forschungsgemein-
formance of proposition extraction detached from schaft within the project ASSURE.
the task of clause splitting.
406
References pages 114–119. Association for Computational
Linguistics.
Gabor Angeli, Melvin Jose Johnson Premkumar, and
Christopher D Manning. 2015. Leveraging Linguis- Mausam, Michael Schmitz, Robert Bart, Stephen
tic Structure For Open Domain Information Extrac- Soderland, and Oren Etzioni. 2012. Open language
tion. In Proceedings of the 53rd Annual Meeting of learning for information extraction. In Proceedings
the Association for Computational Linguistics and of the 2012 Joint Conference on Empirical Methods
the 7th International Joint Conference on Natural in Natural Language Processing and Computational
Language Processing, volume 1, pages 344–354. Natural Language Learning, pages 523–534. Asso-
ciation for Computational Linguistics.
Luciano Del Corro and Rainer Gemulla. 2013.
ClausIE: Clause-Based Open Information Extrac- Gabor Melli, Zhongmin Shi, Yang Wang, Yudong Liu,
tion. In Proceedings of the 22nd international con- Anoop Sarkar, and Fred Popowich. 2006. Descrip-
ference on World Wide Web, pages 355–366. ACM. tion of SQUASH, the SFU Question Answering
Summary Handler for the DUC-2006 Summariza-
Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko tion Task. In Proceedings of the 6th Document Un-
Horn, Ni Lao, Kevin Murphy, Thomas Strohmann, derstanding Conference (DUC 2006).
Shaohua Sun, and Wei Zhang. 2014. Knowledge
vault: A web-scale approach to probabilistic knowl- Julian Michael, Gabriel Stanovsky, Luheng He, Ido
edge fusion. In Proceedings of the 20th ACM Dagan, and Luke Zettlemoyer. 2018. Crowdsourc-
SIGKDD international conference on Knowledge ing Question-Answer Meaning Representations. In
discovery and data mining, pages 601–610. ACM. Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computa-
Jessica Ficler and Yoav Goldberg. 2016. A Neural Net- tional Linguistics: Human Language Technologies,
work for Coordination Boundary Prediction. In Pro- volume 2, pages 560–568.
ceedings of the 2016 Conference on Empirical Meth-
ods in Natural Language Processing, pages 23–32. Christina Niklaus, Matthias Cetto, André Freitas, and
Siegfried Handschuh. 2018. A survey on open infor-
Nicholas FitzGerald, Julian Michael, Luheng He, and mation extraction. In Proceedings of the 27th Inter-
Luke Zettlemoyer. 2018. Large-Scale QA-SRL national Conference on Computational Linguistics,
Parsing. arXiv preprint arXiv:1805.05377. pages 3866–3878.
Kiril Gashteovski, Rainer Gemulla, and Luciano Fabio Petroni, Luciano Del Corro, and Rainer Gemulla.
Del Corro. 2017. Minie: minimizing facts in open 2015. CORE: Context-Aware Open Relation Ex-
information extraction. In Proceedings of the 2017 traction with Factorization Machines. In Proceed-
Conference on Empirical Methods in Natural Lan- ings of the 2015 Conference on Empirical Methods
guage Processing, pages 2630–2640. in Natural Language Processing, pages 1763–1773.
Rachayita Giri, Yosha Porwal, Vaibhavi Shukla, Palak Maria Pontiki, Dimitris Galanis, Haris Papageorgiou,
Chadha, and Rishabh Kaushal. 2017. Approaches Ion Androutsopoulos, and Suresh Manandhar. 2014.
for information retrieval in legal documents. In Con- Semeval-2014 task 4: Aspect based sentiment anal-
temporary Computing (IC3), 2017 Tenth Interna- ysis. In Proceedings of the 10th international work-
tional Conference on, pages 1–6. IEEE. shop on semantic evaluation (SemEval-2014), pages
27–35.
Tushar Khot, Ashish Sabharwal, and Peter Clark. 2017. Randolph Quirk. 1985. A grammar of contemporary
Answering Complex Questions Using Open Infor- English, 11. impression edition. Longman, London.
mation Extraction. In Proceedings of the 55th An-
nual Meeting of the Association for Computational Sebastian Riedel, Limin Yao, Andrew McCallum, and
Linguistics, volume 2, pages 311–316. Benjamin M Marlin. 2013. Relation extraction with
matrix factorization and universal schemas. In Pro-
John Lee and J Buddhika K Pathirage Don. 2017. ceedings of the 2013 Conference of the North Amer-
Splitting Complex English Sentences. In Proceed- ican Chapter of the Association for Computational
ings of the 15th International Conference on Parsing Linguistics: Human Language Technologies, pages
Technologies, pages 50–55. 74–84.
Alexander Löser, Sebastian Arnold, and Tillmann Swarnadeep Saha et al. 2018. Open Information Ex-
Fiehn. 2011. The GoOlap Fact Retrieval Frame- traction from Conjunctive Sentences. In Proceed-
work. In European Business Intelligence Summer ings of the 27th International Conference on Com-
School, pages 84–97. Springer. putational Linguistics, pages 2288–2299.
Mitchell Marcus, Grace Kim, Mary Ann Rudolf Schneider, Tom Oberhauser, Tobias Klatt, Fe-
Marcinkiewicz, Robert MacIntyre, Ann Bies, lix A Gers, and Alexander Löser. 2017. Analysing
Mark Ferguson, Karen Katz, and Britta Schas- errors of open information extraction systems. In
berger. 1994. The Penn Treebank: annotating Proceedings of the First Workshop on Building Lin-
predicate argument structure. In Proceedings of guistically Generalizable NLP Systems, pages 11–
the workshop on Human Language Technology, 18.
407
Gabriel Stanovsky and Ido Dagan. 2016. Creating a
Large Benchmark for Ipen Information Extraction.
In Proceedings of the 2016 Conference on Empiri-
cal Methods in Natural Language Processing, pages
2300–2305.
Gabriel Stanovsky, Ido Dagan, et al. 2015. Open IE as
an intermediate structure for semantic tasks. In Pro-
ceedings of the 53rd Annual Meeting of the Associ-
ation for Computational Linguistics and the 7th In-
ternational Joint Conference on Natural Language
Processing, volume 2, pages 303–308.
Gabriel Stanovsky, Jessica Ficler, Ido Dagan, and Yoav
Goldberg. 2016. Getting More Out Of Syntax with
PROPS. arXiv preprint arXiv:1603.01648.
Gabriel Stanovsky, Julian Michael, Luke Zettlemoyer,
and Ido Dagan. 2018. Supervised open information
extraction. In Proceedings of the 2018 Conference
of the North American Chapter of the Association
for Computational Linguistics: Human Language
Technologies, volume 1, pages 885–895.
408
Sparse Coding in Authorship Attribution for Polish Tweets
409
Proceedings of Recent Advances in Natural Language Processing, pages 409–417,
Varna, Bulgaria, Sep 2–4 2019.
(Sapkota et al., 2015). There were a couple of cluding 95 435 564 characters in total) ob-
attempts to classify tweets based on their char- tained through Twitter API1 by applying
acter n-gram representations (Schwartz et al., Polish language filter (automatic Twitter
2013) (Sari et al., 2017) (Zhang et al., 2015) filter that can produce some noise) and in
with different approaches to extracting a sub- addition tracking the top 100 words from
set of meaningful characters (Plakias and Sta- the frequency list generated for the Pol-
matatos, 2008) (Sapkota et al., 2015). In terms ish language (Kazojć, 2009) in the search
of a classifier used for such a task, CNNs have query,
been recently widely explored and proved suc-
cessful for text analysis (Sierra et al., 2017) Influencer Set – 138 486 tweets written by
(Ruder et al., 2016). 28 Polish influencers who were chosen
A combination of the two approaches ap- manually.
pears in (Shrestha et al., 2017) where a text The Tweet Corpus consists mostly of tweets
is classified based on a sequence of input char- in Polish. We decided not to filter out tweets
acters. The method uses straightforward se- including fragments in foreign languages as it
quences of characters and their n-grams which is a common practice for the Twitter users to
are then embedded and processed in a CNN write their content with injections of other lan-
classifier. Another approach which is using a guages. Data has been gathered using twint
byte pair encoding algorithm for text encod- Python module (Zacharias and Poldi, 2018).
ing together with the application of the same For the Influencer Set we selected the au-
CNN classifier, proves that the method is com- thors mainly due to their activity and the num-
parable to the state-of-the-art despite another ber of followers. In addition we pre-selected
level of text compression (Wang, 2018). The four categories of authors, namely: journalists
latter, however, uses a limited range of char- (8 authors), politicians (6 authors), publicists
acters with at least punctuation and white (4 authors) and computer gamers (2 authors).
space characters ignored. Additionally, in both The topic-related groups are extrated as one
works the embedding layer is trained simulta- of the goals of the experiments was to verify
neously with the classifier, which makes it im- whether the subject on which particular users
possible to transfer the acquired knowledge to tweet is a strong distinguishing factor which
external data sets. affects obtained results. As the activity of dif-
3 Contribution ferent authors is diversified, the Influencer Set
is highly unbalanced with the number of tweets
In our work we extend the solution presented per author deviating from 314 to 18 204 tweets
in (Shrestha et al., 2017) by using Senten- per account. This problem was mediated by
cePiece with Byte-Pair Encoding algorithms subsampling during the experiments.
(Kudo and Richardson, 2018), without exclu-
sion of any characters. Moreover, instead of 5 Text Representation
using a data-specific text embedding trained Tweets are very short texts and do not include
simultaneously with the classifier, we teach a many repeated occurrences of typical stylom-
separate distributional language model on a etry markers like functional words. We can
corpus of Polish tweets. We also test robust- observe prevalence of the information content
ness of our solution by alternating topics in over typical stylistic markers. Thus, we need to
the training and testing corpora, as well as us- search for semantic elements reoccurring for a
ing disjunctive time windows for texts in both single author, as well as characteristic idiosyn-
corpora. crasies of his language, e.g. words or expres-
4 Data sions.
Tweet mostly include many abbreviations,
For the needs of the development and testing typos, or language errors that result in rela-
of our solution we collected two separate data tively high variability of the language and com-
sets: plexity of the statistical picture. Thus, we need
Tweet Corpus – 1 020 234 Polish tweets (in- 1
https://twitter.com
410
Figure 1: High-level structure of the processing pipeline used for obtaining text representation
in our solution. From left to right: encoding model training, data encoding, embedding training.
to transform the original texts into a represen- brary (Mikolov et al., 2013). As a result, every
tation reducing this complexity. Very often, code receives its vector representation. The
e.g. (Shrestha et al., 2017), tweets are repre- whole process is presented in Figure 1.
sented by n-grams instead of words. However, As codes represent different character se-
the number of different n-grams is still very quences: from bi-grams, through sub-words up
high. Instead of using n-grams, we processed to even sub-expressions, we obtain a distribu-
the tweets with the help of the Byte Pair En- tional semantics model which describes text
coding algorithm from SentencePiece library. units of varied granularity that reflect to some
Byte Pair Encoding (henceforth BPE) is a text extent the statistical granularity of a corpus
compression method that constructs a tree-like used for building the particular BPE model.
structure of codes: recursively, two adjacent A side effect is that vectors are also built for
characters or symbols are encoded with a code codes representing single letters that are rather
represented by a single character that does not meaningless, but this causes no harm to the
occur in the whole text. All punctuation and overall properties of the model, as it will be
white characters are treated in the same way as visible.
other symbols. So there is no need for tokeni- The vector size in Skip gram was set to 300
sation. The algorithm starts from the most fre- elements. As the maximum length of a single
quent pairs of characters (and next codes) and tweet is 140 characters, so we made the rep-
ends when all sequences are covered or it has resentation of a single tweet to be a matrix
reached the vocabulary size. Due to its recur- of 140 code vectors2 . In case BPE encoding
sive work the codes mostly represent not only for a tweet uses less than 140 codes (that of-
bigrams but also higher-level n-grams, e.g. fre- ten happens) the rest of the matrix is padded
quent words or even expressions. In all our ex- with a special null vector, i.e. not produced by
periments we used the BPE vocabulary size of: word2vec.
4000 codes. We tested vocabulary sizes from In order to visualise the internal character
1000 to 8000 codes and we did not notice im- of a BPE-encoded text, we present the his-
provement in performance for vocabulary size togram of initial tweet lengths in Figure 2, and
larger than 4000 codes. contrast it with the histogram of the BPE-
BPE is often used for text compression, but encoded tweet lengths in Figure 3. It can
our purpose was to make it a basis for a sub- be noticed that the tweet lengths has become
word distributional semantics model. Texts shorter, more evenly distributed with a slight
encoded by BPE (i.e. tweets) were transformed dominance of the shorter representations. So,
into sequences of BPE codes separated by 2
In the worst case of the weakest BPE mode codes
white spaces and delivered in such a form to correspond to single characters (symbols) in text, i.e.
the Skip gram algorithm from the word2vec li- no more than 140 codes per a single tweet.
411
BPE-encoding results in a more packed text
representation based sometimes on a few codes
(representing sequences of letters).
Figure 4 presents the lengths of letter se-
quences represented by BPE codes obtained
for the domain of politicians. The histograms
are based on a subset of Influencer Set, con-
taining tweets of 6 authors. When we compare
it with the histogram of the whole domain in
Figure 5 it can be observed that the code dic-
tionary which is specific for the given domain Figure 4: Histogram of BPE code lengths
contains slightly larger number of longer spe- obtained by training the model on tweets of
cialised codes. If we take a look into this do- politicians (28 635 tweets).
main specific dictionary we can find codes rep-
resenting sometimes whole words that seem to
be quite accidentally included into the dictio-
nary. Thus, a dictionary built on the basis of
a large corpus can be better suited to analysis
of the authors’ style that will be visible in Sec-
tion 7 and especially in the results presented
in Tables 4 and 3.
6 Classification Model
412
removing elements that seemed to be not rele-
vant for the authors’ styles and which could
bias the classification process towards topic
recognition. Thus, we removed all hashtags,
mentions and URLs before constructing BPE
encoding. The BPE models were generated
from the preprocessed Tweet Corpus and se-
lected subcorpora. Nonetheless, the very fact
of using such extra-linguistic tweet elements,
their placement and frequency might be impor-
tant and relevant to authors’ tweeting styles.
Thus, we replaced all occurrences of extra-
linguistic elements with symbols representing
their types:
413
Tweets Acc - 2 Acc - 4 Acc - 8 Acc - 12 Acc - 17 Acc - 23 Acc - 26 Acc - 28
16000 97.82 % - - - - - - -
8000 97.41 % 90.89 % - - - - - -
4000 95.12 % 87.42 % 84.08 % - - - - -
2000 93.25 % 84.11 % 81.19 % 72.44 % 71.34 % - - -
1000 92.53 % 82.37 % 73.18 % 66.13 % 62.38 % 57.81 % - -
500 87.50 % 81.25 % 71.46 % 57.25 % 49.29 % 49.61 % 51.77 % -
250 88.03 % 69.52 % 63.00 % 44.67 % 48.47 % 38.43 % 40.23 % 42.14 %
100 82.51 % 63.25 % 58.62 % 42.08 % 39.41 % 31.11 % 32.48 % 30.53 %
Table 1: Classification results for a varying number of authors and number of tweets. First
column indicates the number of tweets per user used in experiment and the first row presents
how many authors were classified in it. For instance, Acc-8, means the column under this cell
presents accuracy results for 8-class classification (8 authors).
Table 2: Accuracy results for (Schwartz et al., 2013) data set, compared with results obtained
in (Shrestha et al., 2017) for 50 authors.
414
Table 3: Accuracy score for authors of different domains (rows) with usage of encoding and
embedding trained on different data (columns) with embedding layer active during training.
Encoding and embedding training data
Author
Sport Social Politics Publicists Gamers Corpus
Domain
Sport 0.89 0.85 0.79 - - 0.94
Classified
ing the encoding mode (i.e. this is the the rich morphology and a weekly constrained
baseline case). word order. However the latter was of minor
importance as the model works mostly on the
For domains in which the number of authors subword level and the distributional semantics
is not the same, we subsample from the larger model that does not depend on the word order
pool and average the results over possible com- (i.e. it is not sequential). The subword embed-
binations (e.g. Sport: 3 authors vs Gamers: 2 dings seem to be useful in decomposing Polish
authors – we subsampled 3 times). morphological forms into their natural compo-
In addition, we run the experiments in two nents.
setups:
In addition, Polish tweets often include frag-
• with the embedding layer active during ments written in English (of varied correct-
training a classifier ness) that also adds to the complexity of the
problem and models. It is worth mentioning
• and with embedding layer remaining that in our experiments we used a significantly
frozen. smaller amount of data than it was done in
The results from the first scenario are shown (Shrestha et al., 2017).
in Table 3, while the effects of the second We initially suspected that the model would
setting are presented in Table 4. The la- be biased by the authors’ domains bias (i.e.
bel “Corpus” means that the encoding model topics of tweets or characteristic elements of
and the embeddings were constructed on the the domain jargon). However, the result in
large Tweet Corpus, while the classifiers were the cross-domain model scenario, see Table 3
trained on the given domain. and Table 4, show a different picture. Firstly,
The lack of a value means that for the given in all cases the BPE-based embedding model
pair we did not have enough data to build built on the large corpus appeared to be supe-
a balanced training-testing subset by subsam- rior to the domain-based model. Definitely, the
pling to the size of the smaller domain. difference in size of the corpora mattered in all
cases in favour of Tweet Corpus. However, the
8 Results size of the code dictionary was constant and
During the experiments on the whole set of equal to 4 000, i.e. it was quite small, and we
authors, as expected, the results significantly can expect that the generalisation in the case
decrease with the increasing number of au- of the big corpus was substantial. When the
thors taken into account and consequently the embedding layer got frozen in the second cross-
decreasing number of tweets per an author domain experiment, all the results decreased,
(due to the subsampling). While the results but only slightly. So, in the case of the limited
fall behind those presented in (Shrestha et al., code dictionary – providing a sparse and gen-
2017), they still surpass initial expectations. eralised picture of texts – the domain-focused
It is worth to take into consideration that tuning of the embedding layer appeared to not
the Polish language is more complex than En- be very important.
glish from the statistical point of view due to The comparison with the method of
415
Table 4: Accuracy score for authors of different domains (rows) with usage of encoding and
embedding trained on different data (columns) with embedding layer frozen during training.
Encoding and embedding training data
Author
Sport Social Politics Publicists Gamers Corpus
Domain
Sport 0.90 0.84 0.82 - - 0.95
Classified
(Shrestha et al., 2017) in Table 2 showed that • comparing the proposed approach with
this reference method performs slightly better, other embedding learning methods,
however, the top results come from an ensem-
ble of the two different models while our results • and optimise the classifier architecture.
are achieved by the single method. Moreover,
While language filtering for using tweets in
our model is targeted on languages with rich
the Polish language exclusively seems to be a
inflection.
reasonable thing to do, it might not be per-
9 Conclusions and Further fectly adequate for the content such as tweets.
Research We can observe a growing trend for users to
post content with both languages present and
The proposed method for the authorship attri- the usage of such a practice may also be consid-
bution for Polish tweets expressed good perfor- ered as a characteristic feature of their writing
mance for a large group of authors and large style. Language switch recognition in such a
data set. It is based on the idea of an ap- short text like tweets might be erroneous, but
plication of the Convolutional Neural Network its tracking can improve the representation.
proposed by (Shrestha et al., 2017), but it ex- The question to what extent the method
pands this approach with a distributional se- recognises an author’s style, and to what ex-
mantics model based on the prior application tent this is only due to the correlation with
of BPE text encoding (i.e. embedding vectors some topics can be answered by the analysis of
are built for BPE codes, not words). As a re- the confusion matrices between authors, dis-
sult the proposed methods seems to be better tinctive features and stability of the recogni-
suited for processing short texts in a highly in- tion in time. Such an analysis could be com-
flectional language. It is worth to emphasise bined with a strict filtering of tweets concern-
that the proposed approach is language inde- ing the same events or topics and then verify-
pendent, as the BPE model is driven by the ing the classification performance. This is es-
statistical patterns in the training corpus. pecially important due to the fact that Twitter
The idea of an adaptive, coarse-grained dis- data gets quickly irrelevant as topics change
tributional text representation for the needs of very fast.
non-semantic classification seems to be attrac-
tive and opens several questions. There are Acknowledgments
various points in which the whole process could
be improved, including but not being limited This work is partially financed as part of the
to: investment in the CLARIN-PL research infras-
tructure funded by the Polish Ministry of Sci-
• collecting a larger tweet corpus, ence and Higher Education.
• improving language-based filtering of ob-
tained tweets,
References
• investigating more closely the influence of Georgia Frantzeskou, Efstathios Stamatatos,
the code dictionary size, Stefanos Gritzalis, Carole E. Chaski, and
416
Blake Stephen Howald. 2007. Identifying Conference of the North American Chapter of
authorship by byte-level n-grams: The source the Association for Computational Linguistics:
code author profile (scap) method. IJDE 6(1). Human Language Technologies. Association for
Computational Linguistics, Denver, Colorado,
Tim Gollub, Martin Potthast, Anna Beyer, pages 93–102. https://doi.org/10.3115/v1/N15-
Matthias Busse, Francisco Rangel, Paolo Rosso, 1010.
Efstathios Stamatatos, and Benno Stein. 2013.
Recent trends in digital text forensics and its Yunita Sari, Andreas Vlachos, and Mark Steven-
evaluation. In Pamela Forner, Henning Müller, son. 2017. Continuous n-gram representations
Roberto Paredes, Paolo Rosso, and Benno Stein, for authorship attribution. In Proceedings of the
editors, Information Access Evaluation. Mul- 15th Conference of the European Chapter of the
tilinguality, Multimodality, and Visualization. Association for Computational Linguistics: Vol-
Springer Berlin Heidelberg, Berlin, Heidelberg, ume 2, Short Papers. Association for Computa-
pages 282–302. tional Linguistics, Valencia, Spain, pages 267–
273. https://www.aclweb.org/anthology/E17-
Jerzy Kazojć. 2009. Otwarty słownik 2043.
frekwencyjny leksemów v.06.2009.
https://web.archive.org/web/20091116122442/ Roy Schwartz, Oren Tsur, Ari Rap-
http://www.open- poport, and Moshe Koppel. 2013. Au-
dictionaries.com/slownikfrleks.pdf. thorship attribution of micro-messages.
Moshe Koppel, Jonathan Schler, and Shlomo In EMNLP . ACL, pages 1880–1891.
Argamon. 2011. Authorship attribution in http://aclweb.org/anthology//D/D13/D13-
the wild. Lang. Resour. Eval. 45(1):83–94. 1193.pdf.
https://doi.org/10.1007/s10579-009-9111-2. Prasha Shrestha, Sebastian Sierra, Fabio Gon-
Moshe Koppel and Yaron Winter. 2014. De- zalez, Manuel Montes, Paolo Rosso, and
termining if two documents are written by Thamar Solorio. 2017. Convolutional neural
the same author. Journal of the Association networks for authorship attribution of short
for Information Science and Technology 65. texts. In Proceedings of the 15th Confer-
https://doi.org/10.1002/asi.22954. ence of the European Chapter of the Associa-
tion for Computational Linguistics: Volume 2,
Taku Kudo and John Richardson. 2018. Sen- Short Papers. Association for Computational
tencepiece: A simple and language indepen- Linguistics, Valencia, Spain, pages 669–674.
dent subword tokenizer and detokenizer for neu- https://www.aclweb.org/anthology/E17-2106.
ral text processing. CoRR abs/1808.06226.
http://arxiv.org/abs/1808.06226. Sebastián Sierra, Manuel Montes y Gómez,
Thamar Solorio, and Fabio A. González. 2017.
Nicci Macleod and Tim Grant. 2011. Whose Convolutional neural networks for author profil-
tweet?: authorship analysis of micro-blogs and ing in pan 2017. In CLEF .
other short form messages.
Zhenduo Wang. 2018. Text embedding methods on
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg different levels fortweets authorship attribution.
Corrado, and Jeffrey Dean. 2013. Distributed http://www.d.umn.edu/ tpederse/Pubs/zhenduo-
representations of words and phrases and report.pdf.
their compositionality. CoRR abs/1310.4546.
http://arxiv.org/abs/1310.4546. Cody Zacharias and Francesco Poldi. 2018. Twint
project. https://github.com/twintproject/
Spyridon Plakias and Efstathios Stamatatos. 2008. twint.
Tensor space models for authorship identifica-
tion. In John Darzentas, George A. Vouros, Spy- Xiang Zhang, Junbo Jake Zhao, and Yann LeCun.
ros Vosinakis, and Argyris Arnellos, editors, Ar- 2015. Character-level convolutional networks
tificial Intelligence: Theories, Models and Appli- for text classification. CoRR abs/1509.01626.
cations. Springer Berlin Heidelberg, Berlin, Hei- http://arxiv.org/abs/1509.01626.
delberg, pages 239–249.
Sebastian Ruder, Parsa Ghaffari, and
John G. Breslin. 2016. Character-
level and multi-channel convolutional
neural networks for large-scale author-
ship attribution. CoRR abs/1609.06686.
http://arxiv.org/abs/1609.06686.
Upendra Sapkota, Steven Bethard, Manuel
Montes, and Thamar Solorio. 2015. Not all char-
acter n-grams are created equal: A study in au-
thorship attribution. In Proceedings of the 2015
417
Automatic Question Answering for Medical MCQs: Can It Go Further
than Information Retrieval?
418
Proceedings of Recent Advances in Natural Language Processing, pages 418–422,
Varna, Bulgaria, Sep 2–4 2019.
Contributions We introduce and compare two et al., 2016), and BiDAF (Seo et al., 2016)), as
approaches for automatic question answering that well as IR models and test them on a total of 7787
do not require training data in the form of MCQs, science questions. The questions are divided into
using Information Retrieval (IR) techniques and two sets, challenging and easy, and are targeted at
standard neural network models. Unlike previ- students between the ages of 8 and 13. It is im-
ous work, our neural approach is independent of portant to note that the authors define a question
the performance of the IR system, as it does not as being challenging or easy not on the basis of
build upon it. Thus, it is possible to achieve im- human performance or the age of the students it
provements over both systems by combining them, is targeted at, but based on whether it has been
as each system has an individual contributions to- answered incorrectly by at least two of the base-
wards solving the problem. The best combination line solvers. The results indicated that none of
results in 18% improvement over a random guess the algorithms performed significantly higher than
baseline. The neural models achieve a statistically the random guess baseline of 25% on the chal-
significicant improvement over the random base- lenging set, while the performance on the easy set
line on the challenging sets. The code used in this was within the range of 36% and 62%. Accord-
study, as well as the public data3 are made avail- ing to the authors, a possible explanation for the
able at: https://bit.ly/2jNW2ym. low accuracy is that nearly all models use some
form of information retrieval to obtain relevant
2 Related Work sentences, and the retrieval bias in these systems
is towards sentences that are very similar to the
Most of the recent work in the field focuses question, as opposed to sentences that individu-
on answering reading comprehension questions ally differ but together explain the correct answer
from benchmark datasets such as SQuAD (Ra- (Clark et al., 2018). Notably, the neural solvers
jpurkar et al., 2016), the release of which ignited performed poorly on the easy set, while the best
a rapid progress in the field. For example, Wang result was achieved by an IR-only system.
et al. (2017) use gated self-matching networks
and report accuracy as high as 75.9% over a ran- 3 Data
dom guess baseline of around 4% and a logis-
tic regression baseline of around 51%. Among In the USMLE data each test item is a single-
the most successful approaches in other studies best-answer MCQ consisting of a stem (question)
are ones that use neural models such as match- followed by several response options (distractors),
LSTM to build question-aware passage represen- one of which is the correct answer (key). An ex-
tation (Wang and Jiang, 2015), bi-directional at- ample of such an item is provided in Table 1. We
tention flow networks to model question-passage divide our data into two sets: private and public
pairs (Seo et al., 2016), or dynamic co-attention (Table 2). The private data set consists of a to-
networks (Xiong et al., 2016). tal of 2,720 MCQs and they are not available to
As mentioned in the previous section, auto- the public due to test security reasons. The public
matic question answering for science exams is a data set consists of 454 items from USMLE 2015
lot more challenging than for crowd-sourced read- Step 1, USMLE 2016 Step 1, USMLE 2014 Step
ing comprehension questions. When applied to 2, and USMLE 2017 Step 2 sample leaflets. These
science questions, IR techniques: i) still perform are available at the USMLE website4 and in our
somewhat close to the state-of-the-art and ii) fail repository. For the purpose of this study, we have
on tasks where the correct answer is not specifi- selected only those items that fulfill the following
cally contained in relevant sentences. Clark et al. criteria: i) whose correct answer contains at least
(2018) implement five of the best models from one heading from the Medical Subject Headings
the studies on the reading comprehension data sets (MeSH5 ) database that is at most three words, and
(TableILP (Khashabi et al., 2016), TupleInference ii) have exactly 5 options that have at least one
(Khot et al., 2017), Neural entailment models (De- MeSH heading that is at most three words. The
compAttn, DGEM, and DGEM -OpenIE) (Parikh 4
The items can be accessed at the USMLE web
site at http://www.usmle.org/, for exam-
3 ple:
See Section 3. The Public data set used in this study http://www.usmle.org/pdfs/step-1/
consists of questions released as training materials by the samples_step1.pdf
5
USMLE. https://www.nlm.nih.gov/mesh/
419
A 56-year-old man comes to the emergency department because of a 4-day history of colicky right flank pain that radiates
to the groin and hematuria. Ultrasound examination of the kidneys shows right-sided hydronephrosis and a dilated ureter.
Which of the following is most likely to be found on urinalysis?
(A) Erythrocyte casts
(B) Glucose
(C) Leukocyte casts
(D) Oval fat bodies
(E)* Uric acid crystals
Table 1: An example of an item from the USMLE exam (question 128, USMLE 2015 step 1 sample test questions)
420
Accuracy Public set Private set
Baselines
Random guess baseline 0.2 (.16-.26) 0.2 (.175-0.225)
IR-All baseline 0.25 (.18-.32) 0.332 (.302-.364)
IR
IR-NVA 0.32* (.24-.39) 0.362* (.332-.395)
IR-Nouns 0.33* (.26-.41) 0.311* (.282-.342)
Neural
LSTM 0.29 (.22-.37) 0.29* (.26-.32)
Conv1d attention 0.31* (.23-.37) 0.32* (.292-.353)
Ensemble(Conv1d+LSTM) 0.30* (.24-.39) 0.311* (.282-.342)
Neural +IR
log(IR NVA)+log(conv1d) 0.32* (.25-.40) 0.340* (.310-.373)
Neural as tie breaker 0.37** (.3-.45) 0.396** (.365-.429)
Table 3: Accuracy of the different systems. The values marked with * signify statistically significant difference
over the random guess baseline and ** signifies statistically significant improvement over both baselines.
421
A drawback of the neural approach proposed choice questions in a high-stakesmedical exam. In
in this paper is that it relies on the availability of Proceedings of the Fourteenth Workshop on Innova-
tive Use of NLP for Building Educational Applica-
a manually indexed database such as MEDLINE.
tions.
This limits the applicability of the approach to
other domains, however, this may change when Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke
more resources become available in the future. It Zettlemoyer. 2017. Triviaqa: A large scale distantly
supervised challenge dataset for reading comprehen-
is important to note that in this restricted setting sion. arXiv preprint arXiv:1705.03551.
the method solves a very difficult problem bet-
ter than any other approach so far. In the future, Daniel Khashabi, Tushar Khot, Ashish Sabhar-
wal, Peter Clark, Oren Etzioni, and Dan Roth.
instead of using the adhoc neural network archi-
2016. Question answering via integer programming
tectures presented in this paper, we plan to utilise over semi-structured knowledge. arXiv preprint
state-of-the-art architectures such as Elmo (Peters arXiv:1604.06076.
et al., 2018) or BERT (Devlin et al., 2018), while
Tushar Khot, Ashish Sabharwal, and Peter Clark. 2017.
using the prediction of MESH headings as an ad- Answering complex questions using open informa-
ditional learning objective. tion extraction. arXiv preprint arXiv:1704.05572.
Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Adam Trischler, Tong Wang, Xingdi Yuan, Justin Har-
Ashish Sabharwal, Carissa Schoenick, and Oyvind ris, Alessandro Sordoni, Philip Bachman, and Ka-
Tafjord. 2018. Think you have solved question an- heer Suleman. 2016. Newsqa: A machine compre-
swering? try arc, the ai2 reasoning challenge. arXiv hension dataset. arXiv preprint arXiv:1611.09830.
preprint arXiv:1803.05457.
Shuohang Wang and Jing Jiang. 2015. Learning nat-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and ural language inference with lstm. arXiv preprint
Kristina Toutanova. 2018. Bert: Pre-training of deep arXiv:1512.08849.
bidirectional transformers for language understand-
ing. arXiv preprint arXiv:1810.04805. Wenhui Wang, Nan Yang, Furu Wei, Baobao Chang,
and Ming Zhou. 2017. Gated self-matching net-
Le An Ha and Victoria Yaneva. 2018. Automatic works for reading comprehension and question an-
distractor suggestion for multiple-choice tests using swering. In Proceedings of the 55th Annual Meet-
concept embeddings and information retrieval. In ing of the Association for Computational Linguistics
Proceedings of the Thirteenth Workshop on Innova- (Volume 1: Long Papers), volume 1, pages 189–198.
tive Use of NLP for Building Educational Applica-
tions, pages 389–398. Caiming Xiong, Victor Zhong, and Richard Socher.
2016. Dynamic coattention networks for question
Le An Ha, Victoria Yaneva, Peter Baldwin, and Janet answering. arXiv preprint arXiv:1611.01604.
Mee. 2019. Predicting the difficulty of multiple
422
Self-Knowledge Distillation in Natural Language Processing
423
Proceedings of Recent Advances in Natural Language Processing, pages 423–430,
Varna, Bulgaria, Sep 2–4 2019.
2003; Mikolov et al., 2010). Recently, several ap- 2 Background
proaches have been proposed to make more effi-
In this section, we briefly review the cross-entropy
cient word embedding matrices, usually based on
and knowledge distillation. Also, since our pro-
contextual information (Søgaard et al., 2017; Choi
posed method is based on word embedding, the
et al., 2017).
layer right before the softmax operation, word em-
bedding process is summarized.
On the other hand, knowledge distillation was
proposed by (Hinton et al., 2015) to train new and 2.1 Cross Entropy
usually shallow networks using hidden knowledge For classification with C classes, neural networks
in the probabilities produced by the pretrained net- produce class probabilities pi , i ∈ {0, 1, ...C} by
works. It shows that there is knowledge not only using a softmax output layer which calculates class
in the target probability corresponding to the target probabilities from the logit, zi considering the other
class but also in the other class probabilities in the logits as follows.
estimation of the trained model. In other words,
the other class probabilities can contain additional exp (zi )
pi = P . (1)
information describing the input data samples dif- k exp (zk )
ferently even when the samples are in the same
In most classification problems, the objective
class. Also, samples from different classes could
function for a single sample is defined by the cross-
produce similar distributions to each other.
entropy as follows.
X
In this paper, we propose a new knowledge dis- J(θ) = − yk log pk , (2)
tillation method, self-knowledge distillation (SKD) k
based on the word embedding of the training model
where yk and pk are the target and predicted proba-
itself. That is, self-knowledge is distilled from the
bilities. The cross-entropy can be simply calculated
predicted probabilities produced by the training
by
model, expecting the model has more information
as it is more trained. In the conventional knowledge J(θ) = − log pt , (3)
distillation, the knowledge is distilled from the esti-
mated probabilities of pretrained (or teacher) mod- when the target probability y is a one-hot vector
els. Contrary, in the proposed SKD, knowledge defined as
is distilled from the current model in the training
process, and the knowledge is hidden in the word 1, if k = t(target class)
yk = . (4)
embedding. During the training process, the word 0, otherwise
embedding reflects the relationship between words Note that the cross-entropy objective function
in the vector space. A word close to the target says only how likely input samples belong to the
word in the vector space is expected to have similar corresponding target class, and it does not provide
distribution after softmax, and such information any other information about the input samples.
can be used to approximate the soft target proba-
bility as in knowledge distillation. We apply our 2.2 Knowledge Distillation
proposed method to two popular NLP tasks: LM A well trained deep network model contains mean-
and NMT. The experiment results show that our ingful information (or knowledge) extracted from
proposed method improves the performance of the training datasets for a specific task. Once a deep
tasks. Moreover, SKD reduces overfitting prob- model is trained for a task, the trained model can
lems which we believe is because SKD uses more be used to train new smaller (shallower or thinner)
information. networks as shown in (Hinton et al., 2015; Romero
et al., 2014). This approach is referred to as knowl-
The paper is organized as follows. Background edge distillation.
is reviewed in Section 2. In Section 3, we describe Basically, knowledge distillation provides more
our proposed method, SKD. Experiment results are information to new models for training and im-
presented and analyzed in Section 4, followed by proves the new model’s performance. Thus, when
Section 5 with conclusion. a new model which is usually smaller is trained
424
with the distilled knowledge from the trained deep distilled knowledge from the teacher model. Their
model, it can achieve a similar (or sometimes even experiment results show that the student models
better) performance compared to the pretrained outperform the teacher model. Also, even though
deep model. when the teacher model has a less powerful archi-
In the pretrained model, knowledge lies in the tecture, the knowledge from the trained teacher
class probabilities produced by softmax of the model can boost student models which have more
model as in Eq. (1). All probability values includ- powerful (or bigger) architectures. It means that
ing the target class probability describe relevant even the knowledge is distilled from a relatively
information about the input data. Thus, instead weak model, it can be useful to train a bigger
of one-hot representation of the target label where model.
only the target class is considered in cross-entropy,
2.3 Word Embedding
all probabilities over the whole classes from the
pretrained model can provide more information Word embedding is to convert symbolic represen-
about the input data in cross-entropy, and can teach tation of words to vector representation with se-
new models more efficiently. All probabilities from mantic and syntactic meanings, which reflects the
the pretrained model are considered as soft target relations between words. Including CBOW, Skip-
probabilities. gram (Mikolov et al., 2013), and GloVe (Penning-
In a photo tagging task, depending on the other ton et al., 2014), various word embedding methods
class probabilities, we understand the input image have been proposed to learn a word embedding
better than just target class. When a class ‘mouse’ matrix. The trained embedding matrix can be trans-
has the highest probability, if ‘mascot’ has a rel- ferred to other models like LM or NMT (Ahn et al.,
atively high probability, then the image would be 2016).
probably ‘mickey mouse’. If ‘button’ or ‘pad’ has CBOW predicts a word given its neighbor words,
a high probability, the image would be a mouse as a and Skip-gram predicts the neighbor words given
computer device. The other class probabilities have a word. They use feedforward layers, and the last
some extra information and such knowledge in the layer of CBOW includes the word embedding ma-
pretrained model can be transferred to a new model trix, W , as follows.
by using a soft target distribution of the training z = W h + b, (6)
set.
When the target labels are available, the objec- where b is a bias, h is hidden layer, and z is logits
tive function is a weighted sum of the conventional for the softmax operation.
cross-entropy with the correct labels and the cross- Words in the embedding space have semantic
entropy with the soft target distribution, given by and syntactic similarities, such that two similar
words are close in the space. Thus, when the classi-
X fication is not correct, the error can be interpreted
J(θ) = −(1 − λ) log pt − λ qk log pk , (5) differently depending on the similarity between the
k predicted word and the target word. For example,
where pk is probability for class k produced by when the target word is ‘learning’, if the predicted
current model with parameter θ, and qk is the soft word is ‘training’, then it is less wrong than other
target probability from the pretrained model. λ words like ‘flower’ or ‘internet’. In this paper, we
controls the amount of knowledge from the trained utilize such hidden information (or knowledge) in
model. Note that the conventional knowledge distil- the word embedding space, while training. Fig. 1
lation extracts knowledge from a pretrained model, shows where the word embedding is located in LM
and in this paper, we propose to extract knowledge and NMT, respectively.
from the current model itself without any pretrained
3 Self-Knowledge Distillation
model.
Furthermore, in a recently proposed paper by We propose a new learning method self-knowledge
(Furlanello et al., 2018), they proved that knowl- distillation (SKD) which distills knowledge from
edge distillation can be useful to train a new model a currently training model, following the conven-
which has the same size and the same architecture tional knowledge distillation. In this section, we
as the pretrained model. They trained a teacher describe an algorithm for SKD and its application
model first, then they trained a student model with to language model and neural machine translation.
425
meaning that the class n cannot be more correct
than the real target t, so Eq. (7) becomes
qn = min{exp{−σkwt − wn k2 }, 0.5},
qt = 1 − qn , (8)
426
tion (NMT). Although LM and NMT are actually
sentence generation rather than classification, they
have classification steps to generate words for the
target sentence. Also, the sum of cross-entropies
over the words in the sentence is adapted as an
objective function for them.
In addition, to check if SKD is robust against
errors in the word embedding space, we also evalu-
ate SKD when we add Gaussian noise in the word
embedding space for target words in the decoder.
Figure 2: Given a target class t, a soft target prob-
abilities are obtained based on the distance in the 4 Experiments
word embedding space. However, only the target
To evaluate self-knowledge distillation, we com-
class and the predicted class have soft target proba-
pare it to the baseline models for language model-
bilities in SKD.
ing and neural machine translation.
transits to Eq. (10). To implement the transition, 4.1 Dataset
another parameter α is introduced to Eq. (10), lead-
For language modeling, we use two different
ing to the final objective function as follows.
datasets: Penn TreeBank (PTB) and WiKi-2. PTB
J(θ) = −(1 − αλqn ) log pt − αλqn log pn , (11) was made by (Marcus et al., 1993), and we use the
pre-processed version by (Mikolov et al., 2010).
α starts from 0 with which Eq. (11) becomes the
In the PTB dataset, the train, valid and test sets
conventional cross-entropy. After K iterations, α
have about 887K, 70K, and 78K tokens, respec-
increases by η per iteration and eventually goes up
tively, where the vocabulary size is 10K. The WiKi-
to 1 with which Eq. (11) becomes the same as Eq.
2 dataset introduced by (Merity et al., 2016) con-
(9). In our experiments, we used a simplified equa-
sists of sentences that are extracted from Wikipedia.
tion as in Eq. (12) without λ so that the objective
It has about 2M, 217K, and 245K tokens for train,
function relies gradually more on the soft target
valid, and test sets. Its vocabulary size is about 33K.
probabilities as training goes.
We did not apply additional pre-processing for the
J(θ) = −(1 − αqn ) log pt − αqn log pn . (12) PTB dataset. The WiKi-2 dataset is pre-tokenized
data, therefore we only added an end-of-sentence
Table 1 summarizes the proposed SKD algorithm.
token (<EOS>) to every sentence.
For machine translation, we evaluated models on
Table 1: Self-Knowledge Distillation Algorithm three different translation tasks (En-Fi, Fi-En, and
Algorithm 1: SKD Algorithm En-De) with the available corpora from WMT’15
1 . The dictionary size is 10K for En-fi and Fi-En
Initialize the model parameters θ
translation task, and 30K for the En-De translation
Initialize α = 0 and σ
task.
(See the experiments for σ values.)
Repeat K times: 4.2 Language Modeling
Train the network based on the
Language modeling (LM) has been used in many
cross-entropy in Eq. (3)
different NLP tasks like automatic speech recogni-
Repeat until convergence:
tion (ASR), and machine translation (MT) to cap-
Train the network based on
ture syntactic and semantic structure of a natural
the SKD objective function in Eq. (12)
language. The neural network-based language mod-
Update α with α + η
els (NNLM) and recurrent neural network language
(See the experiments for η values.)
model (RNNLM) catch the syntactic and seman-
tic regularities of an input language (Bengio et al.,
3.3 NLP Tasks 2003; Mikolov et al., 2013). RNNLM is our base-
line, which consists of a single LSTM layer and
SKD is applied to two different NLP tasks: lan-
guage modeling (LM), and neural machine transla- 1
http://www.statmt.org/wmt15/translation-task.html
427
single feed forward layer with ReLU (Le et al., the objective function, we believe that the improve-
2015). ment by SKD is regardless of model architectures.
We evaluate four models: Baseline, Noise (with Table 3 shows that our proposed method im-
Gaussian noise on the word embedding), SKD, proves NMT performance by around 1 BLEU score.
and Noise+SKD. To show that the information by For qualitative comparison, some translation results
SKD is more knowledgeable than random noise, are presented below. The overall quality of transla-
we tested a noise injected model which injects only tion of the SKD model looks better than baseline
Gaussian noise to the word embedding space. The model’s. In other words, when the BLEU scores
word dimension is set to 500 and the number of hid- are similar, the sentences translated by the SKD
den nodes is 400 for all models. We set the σ and model look better.
η in the SKD algorithm in Table 1 0.1 (both PTB
• (src) Hallituslähteen mukaan tämä on yksi monista
and WiKi-2 dataset) and 0.0002 (PTB), 0.00011 ehdotuksista, joita harkitaan.
(WiKi-2), respectively. We applied the SKD object (trg) A governmental source says this is one of the
many proposals to be considered.
function after 500 batches for PTB and 900 batches (baseline) According to government revenue, this is one
for WiKi-2. Note that Wiki-2 data is larger than of the many proposals that are being considered to be
PTB. considered.
(SKD) According to the government, this is one of the
The evaluation metric is the negative log- many proposals that are being considered.
likelihood (NLL) for each sentence (the lower is
the better). Table 2 presents NLLs for the test data
• (src) Meillä on hyvä tunne tuotantoketjunvahvuudesta.
of two datasets with different models. It shows that (trg) We feel very good about supply chain capability.
our proposed methods (both noise injection and (baseline) We have good knowledge of the strength of
the production chain.
self-distillation knowledge) improve the results in (SKD) We have a good sense of the strength of the
the LM task. Note that SKD provides more knowl- production chain.
edgeable information than Gaussian noise.
• (src) En ole oikein tajunnut, että olen näin vanha.
Table 2: NLLs for LM with different models on (trg) I haven’t really realized that I’m this old.
(baseline) I have not been right to realise that I am so
PTB and Wiki-2. old.
Model PTB Wiki-2 (SKD) I am not quite aware that I am so old.
Baseline 101.40 119.49
+Noise 101.28 118.70 • (src) Ne vaikuttavat vasta tulevaisuudessa.
+SKD 99.38 116.85 (trg) They’ll have an impact in the future only.
(baseline) They will only be affected in the future.
+Noise+SKD 97.41 116.60 (SKD) They will only affect in the future.
4.3 Neural Machine Translation Fig. 3 shows a trajectory of the qn values and
NMT has been widely used in machine translation scheduling of the α value during training the En-Fi
research, because of its powerful performance and NMT model described in Eq. (12), respectively.
end-to-end training (Sutskever et al., 2014; Bah- As expected, the qn value becomes larger than 0.5
danau et al., 2015; Johnson et al., 2017). Attention- which means that wn (the predicted word vector)
based NMT models consist of an encoder, a de- is close enough to the wt (the target word vector).
coder, and the attention mechanism (Bahdanau Fig. 3(b) shows the scheduled value of α in Eq.
et al., 2015), which is our baseline in this paper (12). The α value starts from 0 and increases up to
except for replacing GRU with LSTM and using 1 while training. The model is trained with only the
BPE (Sennrich et al., 2016). The encoder takes the cross-entropy for K iterations, and then when the
sequence of source words in the word embedding model captures enough knowledge to be distilled,
form. The decoder works in a similar way to LM, α increases to utilize knowledge from the model.
except the attention mechanism. See (Bahdanau Also, as in Fig. 4, the SKD models are not (or
et al., 2015) for NMT and the attention mechanism more slowly) overfitted to the training data. We be-
in detail. lieve that SKD provides more information distilled
In the experiments, we check how much SKD by the training model itself to prevent overfitting.
can improve model’s performance using the simple Note that there is no significant difference in the
baseline architecture. Since SKD modifies only improvements by SKD and Noise, but Noise+SKD
428
(a) qn value during NMT model training
429
References Tomas Mikolov, Martin Karafiat, Lukas Burget, Jan
Cernocky, and Sanjeev Khudanpur. 2010. Recurrent
Sungjin Ahn, Heeyoul Choi, Tanel Pärnamaa, and neural network based language model. In INTER-
Yoshua Bengio. 2016. A neural knowledge language SPEECH 2010, 11th Annual Conference of the Inter-
model. CoRR abs/1608.00318:1–10. national Speech Communication Association. pages
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- 1045–1048.
gio. 2015. Neural Machine Translation by Jointly
Learning to Align and Translate. In Proc. Int’l Conf. Jeffrey Pennington, Richard Socher, and Christopher D.
on Learning Representations (ICLR). Manning. 2014. Glove: Global vectors for word rep-
resentation. In Empirical Methods in Natural Lan-
Yoshua Bengio, Aaron C. Courville, and Pascal Vin- guage Processing (EMNLP). pages 1532–1543.
cent. 2013. Representation learning: A review and
new perspectives. IEEE Trans. Pattern Anal. Mach. Joseph Redmon and Ali Farhadi. 2017. YOLO9000:
Intell. 35(8):1798–1828. better, faster, stronger. In 2017 IEEE Conference
on Computer Vision and Pattern Recognition, CVPR
Yoshua Bengio, Réjean Ducharme, and Pascal Vin- 2017, Honolulu, HI, USA, July 21-26, 2017. pages
cent. 2003. A Neural Probabilistic Language Model. 6517–6525.
The Journal of Machine Learning Research 3:1137–
1155. Adriana Romero, Nicolas Ballas, Samira Ebrahimi Ka-
hou, Antoine Chassang, Carlo Gatta, and Yoshua
Heeyoul Choi, Kyunghyun Cho, and Yoshua Bengio. Bengio. 2014. Fitnets: Hints for thin deep nets.
2017. Context-dependent word representation for CoRR abs/1412.6550.
neural machine translation. Computer Speech and
Language 45:149–160. Rico Sennrich, Barry Haddow, and Alexandra Birch.
2016. Neural machine translation of rare words with
Tommaso Furlanello, Zachary Chase Lipton, Michael subword units. In 54th Annual Meeting of the Asso-
Tschannen, Laurent Itti, and Anima Anandkumar. ciation for Computational Linguistics. pages 1715–
2018. Born-again neural networks. In Proceedings 1725.
of the 35th International Conference on Machine
Learning, ICML 2018, Stockholmsmässan, Stock- Anders Søgaard, Yoav Goldberg, and Omer Levy. 2017.
holm, Sweden, July 10-15, 2018. pages 1602–1611. A strong baseline for learning cross-lingual word
embeddings from sentence alignments. In Proceed-
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian ings of the 15th Conference of the European Chap-
Sun. 2016. Deep residual learning for image recog- ter of the Association for Computational Linguistics,
nition. In 2016 IEEE Conference on Computer Vi- EACL 2017, Valencia, Spain, April 3-7, 2017, Vol-
sion and Pattern Recognition, CVPR 2016, Las Ve- ume 1: Long Papers. pages 765–774.
gas, NV, USA, June 27-30, 2016. pages 770–778.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean.
Sequence to Sequence Learning with Neural Net-
2015. Distilling the knowledge in a neural network.
works. In Advances in Neural Information Process-
CoRR abs/1503.02531.
ing Systems (NIPS).
Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim
Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Tho- Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
rat, Fernanda B. Viégas, Martin Wattenberg, Greg Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
Corrado, Macduff Hughes, and Jeffrey Dean. 2017. Kaiser, and Illia Polosukhin. 2017. Attention is all
Google’s multilingual neural machine translation you need. In Advances in Neural Information Pro-
system: Enabling zero-shot translation. TACL cessing Systems 30: Annual Conference on Neural
5:339–351. Information Processing Systems 2017, 4-9 Decem-
ber 2017, Long Beach, CA, USA. pages 6000–6010.
Quoc V. Le, Navdeep Jaitly, and Geoffrey E. Hinton.
2015. A simple way to initialize recurrent networks
of rectified linear units. CoRR abs/1504.00941.
Mitchell P Marcus, Mary Ann Marcinkiewicz, and
Beatrice Santorini. 1993. Building a large annotated
corpus of english: The Penn Treebank. Computa-
tional Linguistics 19(2):313–330.
Stephen Merity, Caiming Xiong, James Bradbury, and
Richard Socher. 2016. Pointer sentinel mixture mod-
els. CoRR abs/1609.07843.
Tomas Mikolov, Greg Corrado, Kai Chen, and Jeffrey
Dean. 2013. Efficient Estimation of Word Repre-
sentations in Vector Space. In Proc. Int’l Conf. on
Learning Representations (ICLR).
430
From the Paft to the Fiiture: a Fully Automatic NMT and Word
Embeddings Method for OCR Post-Correction
431
Proceedings of Recent Advances in Natural Language Processing, pages 431–436,
Varna, Bulgaria, Sep 2–4 2019.
pean Commission-funded IMPACT project4 gath- parallel data from our corpus containing OCR er-
ers 26 national libraries and commercial providers rors, then we will present the model designed to
to “take away the barriers that stand in the way carry out the actual error correction.
of the mass digitization of the European cultural
heritage” by improving OCR technology and ad- 3.1 Extracting Parallel Data
vocating for best practices. To extract a parallel corpus of OCR errors and
Dong and Smith (2018) present an unsuper- their correctly spelled counterparts out of our cor-
vised method for OCR post-correction. As op- pus, we use a simple procedure consisting of mea-
posed to our character-level approach, they use suring the similarity of the OCR errors with their
a word-level sequence-to-sequence approach. As correct spelling candidates. The similarity is mea-
such a model requires training data, they gather the sured in two ways, on the one hand an erroneous
data automatically by using repeated texts. This form will share a similarity in meaning with the
means aligning the OCRed text automatically with correct spelling as they are realizations of the same
matched variants of the same text from other cor- word. On the other hand, an erroneous form is
pora or within the OCRed text itself. In contrast, bound to share similarity on the level of charac-
our unsupervised approach does not require any ters, as noted by Hill and Hengchen (2019) in their
repetition of text, but rather repetition of individ- study of OCR typically failing on a few characters
ual words. on the corpus at hand.
Different machine translation approaches have
In order to capture the semantic similarity, we
been used in the past to solve the similar prob-
use Gensim (Řehůřek and Sojka, 2010) to train a
lem of text normalization, which means converting
Word2Vec (Mikolov et al., 2013) model.5 As this
text written in a non-standard form of a language
model is trained on the corpus containing OCR er-
to the standard form in order to facilitate its pro-
rors, when queried for the most similar words with
cessing with existing NLP tools. SMT (statistical
a correctly spelled word as input, the returned list
machine translation) has been used previously, for
is expected to contain OCR errors of the correctly
instance, to normalize historical text (Pettersson
spelled word together with real synonyms, the key
et al., 2013) to modern language and to normalize
finding which we will exploit for parallel data ex-
modern Swiss German dialects (Samardzic et al.,
traction.
2015) into a unified language form. More recently
As an example to illustrate the output of
with the rise of the NMT, research has emerged
the Word2Vec model, a query with the word
in using NMT to normalize non-standard text, for
friendship yields friendlhip, friendihip, friend-
example work on normalization of medieval Ger-
flip, friend-, affection, friendthip, gratitude, affe-
man (Korchagina, 2017) and on historical English
tion, friendflhip and friendfiip as the most similar
(Hämäläinen et al., 2018).
words. In other words, in addition to the OCR
All of the normalization work cited above on us-
errors of the word queried for, other correctly-
ing machine translation for normalization has been
spelled, semantically similar words (friend-, affec-
based on character-level machine translation. This
tion and gratitude) and even their erroneous forms
means that words are split into characters and the
(affetion) are returned. Next, we will describe our
translation model will learn to translate from char-
method (as shown in Algorithm 1) to reduce noise
acter to character instead of word to word.
in this initial set of parallel word forms.
3 Model As illustrated by the previous example, we need
a way of telling correct and incorrect spellings
As indicated by the related work on text normal- apart. In addition, we will need to know which
ization, character-level machine translation is a vi- incorrect spelling corresponds to which correct
able way of normalizing text into a standard va- spelling (affetion should be grouped with affection
riety. Therefore, we will also use character-level instead of friendship).
NMT in building our sequence-to-sequence OCR For determining whether a word is a correctly
post-correction model. However, such a model spelled English word, we compare it to the lem-
requires parallel data for training. First, we will
present our method of automatically extracting 5
Parameters: CBOW architecture, window size of 5, fre-
quency threshold of 100, 5 epochs. Tokens were lowercased
4
http://www.impact-project.eu and no stopwords were removed.
432
mas of the Oxford English Dictionary (OED).6 source all >=2 >=3 >=4 >=5
W2V all 29013 28910 27299 20732 12843
If the word exists in the OED, it is spelled cor- W2V freq
11730 11627 10373 7881 5758
rectly. However, as we are comparing to the OED >100,000
lemmas, inflectional forms would be considered BNC 7692 7491 6681 5926 4925
as errors, therefore, we lemmatize the word with Table 1: Sizes of the extracted parallel datasets
spaCy7 (Honnibal and Montani, 2017). If neither
the word nor its lemma appear in the OED, we
consider it as an OCR error. Word2Vec model and list the words that are cor-
For a given correct spelling, we get the most rectly spelled. We use this list of correctly spelled
similar words from the Word2Vec model. We then words in the model to do the extraction. How-
group these words into two categories: correct En- ever, as this list introduces noise to the parallel
glish words and OCR errors. For each OCR error, data, we combat this noise by producing another
we group it with the most similar correct word on list of correctly spelled words that have occurred
the list. This similarity is measured by using Lev- over 100,000 times in ECCO. For these two word
enshtein edit distance (Levenshtein, 1966). The lists, one containing all the correct words in the
edit distances of the OCR errors to the correct model and the other filtered with word frequen-
words they were grouped with are then computed. cies, we produce parallel datasets consisting of
If the distance is higher than 3 – a simple heuris- words longer or equal to 1, 2, 3, 4 and 5. The
tic, based on ad-hoc testing –, we remove the OCR idea behind these different datasets is that longer
error from the list. Finally, we have extracted a words are more likely to be matched correctly with
small set of parallel data of correct English words their OCR error forms, and also frequent words
and their different erroneous forms produced by will have more erroneous forms than less frequent
the OCR process. ones.
In addition, we use the frequencies from the
Algorithm 1: Extraction of parallel data British National Corpus (The BNC Consortium,
Draw words w from the input word list; 2007) to produce one more dataset of words occur-
for w do ring in the BNC over 1000 times to test whether
Draw synonyms sw in the word the results can be improved with frequencies ob-
embedding model tained from a non-noisy corpus. This BNC dataset
for synonym sw do is also used to produce multiple datasets based on
if sw is correctly spelled then the length of the word. The sizes of these auto-
Add sw to correct forms f ormsc matically extracted parallel datasets are shown in
end Table 1.
else
Add sw to error forms f ormse 3.2 The NMT Model
end We use the automatically extracted parallel
end datasets to train a character level NMT model
for error e in f ormse do for each dataset. For this task, we use Open-
group e with the correct form in
NMT8 (Klein et al., 2017) with the default pa-
f ormsc by Levmin
rameters except for the encoder where we use a
if Lev(e,c) > 3 then
BRNN (bi-directional recurrent neural network)
remove(e)
instead of the default RNN (recurrent neural net-
end
work) as BRNN has been shown to provide a per-
end
formance gain in character-level text normaliza-
end
tion (Hämäläinen et al., 2019). We use the de-
fault of two layers for both the encoder and the
We use the extraction algorithm to extract the decoder and the default attention model, which is
parallel data by using several different word lists. the general global attention presented by Luong
First, we list all the words in the vocabulary of the et al. (2015). The models are trained for the de-
fault number of 100,000 training steps with the
6
http://www.oed.com.
7 8
Using the en_core_web_md model. Version 0.2.1 of opennmt-py
433
source all >=2 >=3 >=4 >=5
False No False No False No False No False No
Correct Correct Correct Correct Correct
positive output positive output positive output positive output positive output
W2V all 0,510 0,350 0,140 0,500 0,375 0,125 0,520 0,325 0,155 0,490 0,390 0,120 0,525 0,390 0,085
W2V freq
0,515 0,305 0,180 0,540 0,310 0,150 0,510 0,340 0,150 0,540 0,315 0,145 0,515 0,330 0,155
>100,000
BNC 0,580 0,285 0,135 0,555 0,300 0,145 0,570 0,245 0,185 0,550 0,310 0,140 0,550 0,315 0,135
434
Acknowledgements Guillaume Klein, Yoon Kim, Yuntian Deng, Jean
Senellart, and Alexander M. Rush. 2017. Open-
We would like to thank the COMHIS group10 for NMT: Open-Source Toolkit for Neural Machine
their support, as well as GALE for providing the Translation. In Proc. ACL.
group with ECCO data. Natalia Korchagina. 2017. Normalizing medieval ger-
man texts: from rules to deep learning. In Proceed-
ings of the NoDaLiDa 2017 Workshop on Processing
References Historical Language, pages 12–17.
Yvonne Adesam, Dana Dannélls, and Nina Tahmasebi. Vladimir I Levenshtein. 1966. Binary codes capable
2019. Exploring the quality of the digital historical of correcting deletions, insertions, and reversals. In
newspaper archive kubhist. Proceedings of DHN. Soviet physics doklady, 8, pages 707–710.
Rui Dong and David Smith. 2018. Multi-input atten- Minh-Thang Luong, Hieu Pham, and Christopher D
tion for unsupervised OCR correction. In Proceed- Manning. 2015. Effective approaches to attention-
ings of the 56th Annual Meeting of the Association based neural machine translation. arXiv preprint
for Computational Linguistics (Volume 1: Long Pa- arXiv:1508.04025.
pers), pages 2363–2372. Tomas Mikolov, Kai Chen, Greg Corrado, and Jef-
frey Dean. 2013. Efficient estimation of word
Senka Drobac, Pekka Sakari Kauppinen, Bo Krister Jo- representations in vector space. arXiv preprint
han Linden, et al. 2017. OCR and post-correction of arXiv:1301.3781.
historical finnish texts. In Proceedings of the 21st
Nordic Conference on Computational Linguistics, Stephen Mutuvi, Antoine Doucet, Moses Odeo, and
NoDaLiDa, 22-24 May 2017, Gothenburg, Sweden. Adam Jatowt. 2018. Evaluating the impact of OCR
Linköping University Electronic Press. errors on topic modeling. In International Con-
ference on Asian Digital Libraries, pages 3–14.
Greta Franzini, Mike Kestemont, Gabriela Rotari, Springer.
Melina Jander, Jeremi K Ochab, Emily Franzini,
Joanna Byszuk, and Jan Rybicki. 2018. Attribut- Vivi Nastase and Julian Hitschler. 2018. Correction
ing authorship in the noisy digitized correspondence of OCR word segmentation errors in articles from
of Jacob and Wilhelm Grimm. Frontiers in Digital the ACL collection through neural machine transla-
Humanities, 5:4. tion methods. In Proceedings of the 11th Language
Resources and Evaluation Conference, Miyazaki,
Mika Hämäläinen, Tanja Säily, Jack Rueter, Jörg Japan. European Language Resource Association.
Tiedemann, and Eetu Mäkelä. 2018. Normalizing
early English letters to present-day English spelling. Eva Pettersson, Beáta Megyesi, and Jörg Tiedemann.
In Proceedings of the Second Joint SIGHUM Work- 2013. An SMT approach to automatic annotation
shop on Computational Linguistics for Cultural of historical text. In Proceedings of the workshop
Heritage, Social Sciences, Humanities and Litera- on computational historical linguistics at NODAL-
ture, pages 87–96, Santa Fe, New Mexico. Associa- IDA 2013; May 22-24; 2013; Oslo; Norway. NEALT
tion for Computational Linguistics. Proceedings Series 18, 087, pages 54–69. Linköping
University Electronic Press.
Mika Hämäläinen, Tanja Säily, Jack Rueter, Jörg Michael Piotrowski. 2012. Natural language process-
Tiedemann, and Eetu Mäkelä. 2019. Revisiting ing for historical texts. Synthesis lectures on human
NMT for normalization of early English letters. In language technologies, 5(2):1–157.
Proceedings of the 3rd Joint SIGHUM Workshop
on Computational Linguistics for Cultural Heritage, Radim Řehůřek and Petr Sojka. 2010. Software Frame-
Social Sciences, Humanities and Literature, pages work for Topic Modelling with Large Corpora. In
71–75, Minneapolis, USA. Association for Compu- Proceedings of the LREC 2010 Workshop on New
tational Linguistics. Challenges for NLP Frameworks, pages 45–50, Val-
letta, Malta. ELRA. http://is.muni.cz/
Mark J. Hill and Simon Hengchen. 2019. Quantify- publication/884893/en.
ing the impact of dirty OCR on historical text analy-
sis: Eighteenth Century Collections Online as a case Kepa Joseba Rodriquez, Mike Bryant, Tobias Blanke,
study. Digital Scholarship in the Humanities: DSH. and Magdalena Luszczynska. 2012. Comparison of
named entity recognition tools for raw OCR text. In
Matthew Honnibal and Ines Montani. 2017. spaCy 2: KONVENS, pages 410–414.
Natural Language Understanding with Bloom Em-
beddings, Convolutional Neural Networks and In- Tanja Samardzic, Yves Scherrer, and Elvira Glaser.
cremental Parsing. To appear. 2015. Normalising orthographic and dialectal vari-
ants for the automatic processing of Swiss German.
10
https://www.helsinki.fi/en/ In Proceedings of the 7th Language and Technology
researchgroups/computational-history Conference.
435
David A. Smith and Ryan Cordell. 2019. A research
agenda for historical and multilingual optical char-
acter recognition. Technical report, Northeastern
University.
Sandeep Soni, Lauren Klein, and Jacob Eisenstein.
2019. Correcting whitespace errors in digitized
historical texts. In Proceedings of the 3rd Joint
SIGHUM Workshop on Computational Linguistics
for Cultural Heritage, Social Sciences, Humanities
and Literature, pages 98–103, Minneapolis, USA.
Association for Computational Linguistics.
Simon Tanner, Trevor Muñoz, and Pich Hemy Ros.
2009. Measuring mass text digitization quality and
usefulness. D-lib Magazine, 15(7/8):1082–9873.
The BNC Consortium. 2007. The British Na-
tional Corpus, version 3 (BNC XML Edition).
Distributed by Bodleian Libraries, University of
Oxford, on behalf of the BNC Consortium.
Http://www.natcorp.ox.ac.uk/.
436
Investigating Terminology Translation in Statistical and Neural Machine
Translation: A Case Study on English-to-Hindi and Hindi-to-English
437
Proceedings of Recent Advances in Natural Language Processing, pages 437–446,
Varna, Bulgaria, Sep 2–4 2019.
For experimentation we used the IIT Bom- concerned, HEPS and HENS produce moderate
bay English-Hindi parallel corpus (Kunchukuttan BLEU scores (34.1 BLEU and 39.9 BLEU) on
et al., 2017). For building additional LMs for the test set. As expected, translation quality in
Hindi and English we use the HindEnCorp mono- the morphologically-rich to morphologically-poor
lingual corpus (Bojar et al., 2014) and monolin- language improves.
gual data from the OPUS project (Tiedemann,
2012), respectively. Corpus statistics are shown 3 Creating Gold Standard Evaluation
in Table 1. We selected 2,000 sentences (test set) Set
for the evaluation of the MT systems and 996 sen-
tences (development set) for validation from the To evaluate terminology translation with our MT
Judicial parallel corpus (cf. Table 1) which is a systems, we manually annotated the test set by
juridical domain corpus (i.e. proceedings of legal marking term-pairs on the source- and target-sides
judgments). The MT systems were built with the of the test set (cf. Table 1) with a view to creat-
training set shown in Table 1 that includes the re- ing a gold standard evaluation set. The annotation
maining sentences of the Judicial parallel corpus. process is performed using our own bilingual term
annotation tool, TermMarker. If there is a source
term present in the source sentence, its transla-
Table 1: Corpus Statistics. tion equivalent (i.e. target term) is found in the
target sentence, and the source–target term-pair is
English–Hindi parallel corpus marked. The annotators are native Hindi evalua-
Sentences Words (En) Words (Hi) tors with excellent English skills. They were in-
Training set 1,243,024 17,485,320 18,744,496
(Vocabulary) 180,807 309,879 structed to mark those words as terms that belong
Judicial 7,374 179,503 193,729 to legal or judicial domains. The annotators were
Development set 996 19,868 20,634 also instructed to mark those sentence-pairs from
Test set 2,000 39,627 41,249
the test set that contain errors (e.g. mistransla-
Monolingual Corpus Sentences Words
tions, spelling mistakes) in either source or target
Used for PB-SMT Language Model
English 11M 222M
sentences. The annotators reported 75 erroneous
Hindi 10.4M 199M sentence-pairs which we discarded from the test
Used for NMT Back Translation set. In addition, 655 sentence-pairs of the test set
English 1M 20.2M did not contain any terms. We call the remain-
Hindi 903K 14.2M
ing 1,270 sentence-pairs our gold-testset. Each
sentence-pair of gold-testset contains at least one
We present the comparative performance of the aligned source-target term-pair. We have made the
PB-SMT and NMT systems in terms of BLEU gold-testset publicly available to the research com-
score (Papineni et al., 2002) in Table 2. Addition- munity.1
ally, we performed statistical significance tests us-
ing bootstrap resampling methods (Koehn, 2004). Annotation Suggestions from Bilingual Ter-
The confidence level (%) of the improvement ob- minology While manually annotating bilingual
tained by one MT system with respect to the an- terms in the judicial domain test set, we took sup-
other MT system is reported. As can be seen port from a bilingual terminology that was auto-
matically created from the Judicial corpus (cf. Ta-
Table 2: Performance of MT systems on BLEU. ble 1). For automatic bilingual term extraction we
followed the approach of Haque et al. (2018). We
System BLEU System BLEU found 3,064 English terms and their target equiva-
EHPS 28.8 HEPS 34.1 lents (3,064 Hindi terms) in the source- and target-
EHNS 36.6 (99.9%) HENS 39.9 (99.9%)
sides of gold-testset, respectively.
438
glish word ‘affidavit’ has multiple target equiv- does not consider various nuances of term trans-
alents (LIVs (lexical and inflectional variations)) lation errors. We propose an error typology taking
in Hindi even if the translation domain is legal terminology translation into consideration. First,
or juridical: ‘shapath patr’, ‘halaphanaama’, ‘ha- we translated the test set sentences with our MT
laphanaame’, or ‘halaphanaamo’. The term ‘sha- systems, and sampled 300 translations from the
path patr’ is the lexical variation of Hindi term whole translation set. Then, the terminology trans-
‘halaphanaama’. The base form ‘halaphanaama’ lations were manually inspected, noting the pat-
could have many inflectional variations (e.g. ‘ha- terns of the term translation-related errors. From
laphanaame’, ‘halaphanaamo’) given the sen- our observations we found that the terminology
tence’s syntactic and morphological profile (e.g. translation-related errors can be classified into
gender, case). eight primary categories. As far as the term trans-
For each term we check whether the term has lation quality of an MT system is concerned, our
any additional LIVs pertaining to the juridical do- proposed typology could provide a better perspec-
main and relevant to the context of the sentence. If tive as to how the MT system lacks quality in
this is the case, we include the relevant variations translating domain terms. The categories are as
as legitimate alternatives term. follows: (i) reorder error (RE): the translation of
We again exploit the method of Haque et al. a source term forms the wrong word order in the
(2018) for obtaining variation suggestions for a target, (ii) inflectional error (IE): the translation of
term. The automatically extracted bilingual ter- a source term inflicts a morphological error, (iii)
minology of Haque et al. (2018) comes with the partial error (PE): the MT system correctly trans-
four highest-weighted target terms for a source lates part of a source term into the target and com-
term. If the annotator accepts an annotation sug- mits an error for the remainder of the source term,
gestion (source–target term-pair) from the bilin- (iv) incorrect lexical selection (ILS): the transla-
gual terminology, the remaining three target terms tion of a source term is an incorrect lexical choice,
are considered as alternative suggestions of the tar- (v) term drop (TD): the MT system omits the
get term. source term in translation, (vi) source term copied
Two annotators took part in the annotation task, (STC): a source term or part of it is copied ver-
and two sets of annotated data were obtained. The batim to target, (vii) disambiguation issue in tar-
term-pairs of gold-testset are finalised on the ba- get (DIT): although the MT system makes a po-
sis of the annotation agreement by the two annota- tentially correct lexical choice for a source term,
tors, i.e. we keep those source–target term-pairs its translation-equivalent does not carry the mean-
in gold-testset for which both annotators agree ing of the source term, and (viii) other error (OE):
that the source and target entities are terms and there is an error in relation to the translation of a
aligned. On completion of the annotation pro- source term, whose category, however, is beyond
cess, inter-annotator agreement was computed us- all remaining error categories. The proposed ter-
ing Cohen’s kappa (Cohen, 1960) at word-level. minology translation error typology is illustrated
For each word we count an agreement whenever in Figure 1 (cf. Appendix A).
both annotators agree that it is a term (or part of
Apart from the above error categories, we have
term) or non-term entity. We found the kappa co-
a class for a source term being correctly translated
efficient to be very high (i.e. 0.95) for the annota-
into the target, i.e. the MT system produces a cor-
tion task. This indicates that our terminology an-
rect translation (CT) for a source term. As pointed
notation is of excellent quality.
out in Section 3, we wanted to see how diverse
The final LIV list for a term is the union of the
an MT model can be in translating domain terms,
LIV lists created by the annotators. This helps
and how close the translation of a source term can
make the resulting LIV lists exhaustive.
be to the reference terms or its LIVs or to what
4 Terminology Translation Typology extent (e.g. syntactically and morphologically) it
differs from them. For this reason, we divide the
In order to annotate errors in (automatic) transla- CT class into seven sub-classes, and define them
tions, MT users often exploit the MQM (Multidi- below: (i) CT given the reference term (CTR): the
mensional Quality Metric) error annotation frame- translation of a source term is the reference term,
work (Lommel et al., 2014). One of the error types (ii) CT given one of the LIVs (CTV): the trans-
in the MQM toolkit is terminology (i.e. incon- lation of a source is one of the LIVs of the refer-
sistent with termbase, inconsistent use of termi- ence term, (iii) variation missing (VM): a source
nology) which is an oversimplified attribute and term is correctly translated into the target, but the
439
translation is neither the reference term nor any We considered the correct and incorrect categories
of its LIVs, (iv) correct inflected form (CIF): a for the calculation, i.e. we count an agreement
source term is correctly translated into the target, whenever both evaluators agree that it is a correct
but the translation is neither the reference term nor (or incorrect) term translation, with agreement by
any of its LIVs. However, the base form of the chance = 1/2. We found that the kappa coefficient
translation of the source term is identical to the for this ranges from 0.97 to 1.0. Thus, our man-
base form of either the reference term or one of ual term translation classification quality can be
the LIVs of the reference term, (v) correct reorder labeled as excellent.
form (CRF): a source term is correctly translated
into the target, and the translation includes those 6 Terminology Translation Evaluation in
words that either the reference term or one of the PB-SMT and NMT
LIVs has, but the word order of the translation is
different to that of the the reference term or one of This section provides a comparative evaluation
the LIVs, (vi) correct reorder and inflected form of the ability of PB-SMT and NMT to translate
(CRIF): this class is a combination of both CIF and terminology accurately. In Table 5, we report
CRF, and (vii) other correct (OC): a source term the statistics of terminology translations from the
is correctly translated into the target, whose cate- English-to-Hindi MT task. We see that EHPS
gory, however, is beyond the all remaining correct and EHNS incorrectly translate 303 and 253 En-
categories. glish terms (out of total 3,064 terms) (cf. last
row of Table 5), respectively, into Hindi, resulting
5 Manual Evaluation Plan in 9.9% and 8.3% terminology translation errors,
respectively. We use approximate randomization
This section presents our manual evaluation plan. (Yeh, 2000) to test the statistical significance of
Translations of the source terms of gold-testset the difference between two systems, and report
were manually validated and classified in accor- the significance-level (p-value) in the last column
dance with the set of fine-grained errors and cor- of Table 5. We found that the difference between
rect categories described above. This was accom- the error rates is statistically significant. In Table
plished by the human evaluator. The manual eval- 6, we report the statistics of terminology transla-
uation was carried out with a GUI that randomly tions for the Hindi-to-English MT task. We see
displays a source sentence and its reference trans- that HEPS and HENS incorrectly translate 396 and
lation from gold-testset, and the automatic trans- 353 Hindi terms (cf. last row of Table 6), respec-
lation by one of the MT systems. For each source tively, into English, resulting in 12.9% and 11.5%
term the GUI highlights the source term and the terminology translation errors, respectively. As
corresponding reference term from the source and can be seen from Table 6, the difference between
reference sentences, respectively, and displays the the error rates is statistically significant. When we
LIVs of the reference term, if any. The GUI lists compare these scores with those from Table 5, we
the error and correct categories described in Sec- see that these scores are slightly higher compared
tion 4. The evaluator, a native Hindi speaker with to those for the English-to-Hindi task. Surpris-
the excellent English and Hindi skills, was in- ingly, the terminology translation quality from the
structed to follow the following criteria for eval- morphologically-rich to the morphologically-poor
uating the translation of a source term: (a) judge language deteriorates compared to the overall MT
correctness / incorrectness of the translation of the quality (cf. Section 2).
source term in hypothesis and label it with an ap-
propriate category listed in the GUI, (b) do not 6.1 Comparison with Fine-Grained Category
need to judge the whole translation, but instead This section discusses the numbers and highlights
look at the local context to which both source term phenomena for the fine-grained categories, start-
and its translation belong, and (c) take the syn- ing with those that involve correct terminology
tactic and morphological properties of the source translations.
term and its translation into account.
The manual classification process was com- CTV & VM We see from Tables 5 and 6 that the
pleted for all MT system types. We measure numbers under the CTV (correct term given one of
agreement in manual classification of terminol- the LIVs class are much higher in the English-to-
ogy translation. For this, we randomly selected Hindi task (695 and 662) compared to those in the
an additional 100 segments from gold-testset and Hindi-to-English task (241 and 245). CTV is mea-
hired another evaluator having the similar skills. sured as the count of instances where a source term
440
is (i) correctly translated into the target translation As stated in Section 4, CRIF (correct reorder
and (ii) the translation-equivalent of that term is and inflected form) is the combination of the above
one of the LIVs of the reference term. As can be two types: CRF and CIF. As an example, con-
seen from Table 1, the training set vocabulary size sider a portion of the source Hindi sentence ‘vi-
is much higher in Hindi compared to that in En- vaadagrast vaseeyat hindee mein taip kee gaee hai
glish since the former is a morphologically-rich ...’ and the English reference translation ‘the will
and highly inflected language, which is probably in dispute is typed in hindi ...’ from gold-testset.
the reason why these numbers are much higher in Here, ‘vivaadagrast vaseeyat’ is a Hindi term and
the English-to-Hindi task. its English equivalent is ‘will in dispute’. The
In a few cases, the human evaluator found that translation of the source sentence by the Hindi-
the source terms are correctly translated into the to-English NMT system is ‘the disputed will have
target, but the translations are neither the reference been typed in hindi ...’. We see that the transla-
terms nor any of its LIVs. The manual evaluator tion of the source term (‘vivaadagrast vaseeyat’)
marked those instances with VM (variation miss- is ‘disputed will’ which is correct. We also see
ing) (cf. Tables 5 and 6). These can be viewed as that its word order is different to that of the refer-
annotation mistakes since the annotator omitted to ence term (‘will in dispute’); and the morpholog-
add relevant LIVs for the reference term into gold- ical form of (part of) the translation is not identi-
testset. In future, we aim to make gold-testset as cal to that of (part of) the reference term. As is
exhaustive as possible by adding missing LIVs for the case with CRF and CIF, the manual evaluator
the respective reference terms. marks such term translations as CRIF.
When translation of a source term is correct but
CRF, CIF, CRIF & OC We start this section
its category is beyond the all remaining correct
by highlighting the problem of word order in term
categories, the manual evaluator marks that term
translation, via a translation example from gold-
translation as OC (other correct). In our man-
testset. The Hindi-to-English NMT system cor-
ual evaluation task, we encountered various such
rectly translates a Hindi source term ‘khand nyaay
phenomena, and detail some of those below. (1)
peeth ke nirnay’ (English reference term: ‘division
term transliteration: the translation-equivalent of
bench judgment’) into the following target trans-
a source term is the transliteration of the source
lation (English): “it shall also be relevant to refer
term itself. We observed this happening only
to article 45 - 48 of the judgment of the division
when the target language is Hindi. In practice,
bench”. The manual evalautor marks this term
many English terms (transliterated form) are of-
translation as CRF (correct reorder form) since the
ten used in Hindi text (e.g. ‘decree’ as ‘dikre’,
term ‘judgment of the division bench’ was not in
‘tariff orders’ as ‘tarif ordars’), (2) terminology
the LIV list for the reference term, ‘division bench
translation coreferred: translation-equivalent of a
judgment’.
source term is not found in the hypothesis, how-
We show another example from the Hindi-to-
ever, it is correctly coreferred in target translation,
English translation task. This time, we highlight
and (3) semantically coherent terminology trans-
the issue of inflection in term translation. As an
lation: the translation-equivalent of a source term
example, we consider a source Hindi term ‘ab-
is not seen in the hypothesis, but its meaning is
hikathan’ from gold-testset. Its reference term
correctly transferred into the target. As an exam-
is ‘allegation’, and the LIV list of the reference
ple, consider the source Hindi sentence “sabhee
term includes two lexical variations for ‘allega-
apeelakartaon ne aparaadh sveekaar nahin kiya
tion’: ‘accusation’ and ‘complaint’. A portion of
aur muqadama chalaaye jaane kee maang kee”,
the reference translation is ‘an allegation made by
and reference English sentence “all the appel-
the respondent ...’. A portion of the translation
lants pleaded not guilty to the charge and claimed
produced by the Hindi-to-English NMT system is
to be tried” from gold-testset.2 Here, ‘aparaadh
‘it was alleged by the respondent ...’. In this trans-
sveekaar nahin’ is a Hindi term and its English
lation, we see the Hindi term ‘abhikathan’ is trans-
translation is ‘pleaded not guilty’. The Hindi-to-
lated into ‘alleged’ which is a correct translation of
English NMT system produces the following En-
the Hindi legal term ‘abhikathan’ as per the syn-
glish translation “all the appellants did not accept
tax of the target translation. As above, the man-
the crime and sought to run the suit” for the source
ual evalautor marked these term translations as
sentence. In this example, we see the meaning of
CIF (correct inflected form) since the translation-
equivalent of this term is not found in the LIV list 2
In this example, the reference English sentence is the
of the reference term. literal translation of the source Hindi sentence.
441
the source term ‘aparaadh sveekaar nahin’ is pre- can be seen from the last columns of Tables 5 and
served in the target translation. 6, the differences in these numbers in PB-SMT
and NMT are statistically significant.
Table 3: CIF, CRF, CRIF and OC in PB-SMT and
IE As far as the inflectional error type is con-
NMT.
cerned, the Hindi-to-English PB-SMT system
PB-SMT NMT makes nearly twice as many mistakes as the Hindi-
English-to-Hindi 122 (4%) 98 (3.2%) to-English NMT system (118 vs 76) (cf. Tables 5
Hindi-to-English 90 (2.9%) 138 (4.5%) and 6), which is statistically significant. We see
a different picture in the English-to-Hindi direc-
We recall the rule that we defined while forming tion, i.e. the numbers of morphological errors are
the LIV list for a reference term from Section 3. nearly the same, both in PB-SMT and NMT (77
Our annotators considered only those inflectional vs 79). We found no statistically significant differ-
variations for a reference term that would be gram- ence between them.
matically relevant to the context of the reference PE The numbers (cf. Tables 5 and 6) of partial
translation in which they would appear. In prac- term translation errors in PB-SMT and NMT are
tice, translation of a source sentence can be gener- almost the same regardless of the translation di-
ated in numerous ways. It is possible that a partic- rections. We found that the differences in these
ular inflectional variation of a reference term could numbers are not statistically significant.
be grammatically relevant to the context of the tar-
get translation, which, when it replaces the refer- ILS PB-SMT appears to be more error-prone
ence term in the reference translation, may (syn- than NMT as far as a term’s lexical selection is
tactically) misfit the context of the reference trans- concerned. EHPS commits 77 incorrect lexical
lation. As far as CRF and CRIF are concerned, choices which is 35 more than EHNS. The same
a similar story might be applicable to the transla- trend is observed with the Hindi-to-English direc-
tion of a multiword term. A multiword term may tion. HEPS and HENS commit 139 and 90 incor-
be translated into the target in various ways (as rect lexical choices, respectively. We found that
shown above, ‘division bench judgment’ as ‘judg- the differences in these numbers in PB-SMT and
ment of the division bench’, and ‘disputed will’ as NMT are statistically significant.
‘will in dispute’). In reality, it would be an im- TD Comparing the numbers of the term drop
possible task for the human annotator to consider category from Tables 5 and 6, we see that the
all possible such variations for a multiword refer- numbers of term omission by the PB-SMT and
ence term. Additionally, as above, we saw more NMT systems are almost the same (53 versus 56)
diverse translations with the domain terms under in the English-to-Hindi translation task. We found
the OC category. In Table 3, we report the com- no statistically significant difference in these num-
bined numbers under the above categories (CRF, bers. In contrast, in the Hindi-to-English transla-
CRIF, CIF and OC), with their percentage with re- tion task, HENS drops terms more than twice as
spect to the total number of terms. We see that often as HEPS (86 versus 38). This time, we found
translations of a notable portion of source terms in that the difference in these numbers is statistically
each translation task are diverse. Therefore, inves- significant.
tigating the automation of the terminology transla-
tion evaluation process (Haque et al., 2019), these STC & OE Now we focus on discussing var-
phenomena have to be taken into consideration. ious aspects with the STC (source term copied)
and OE (other error) classes, starting with the
RE Now, we turn our focus to the error classes, English-to-Hindi task. We counted the number of
starting with RE (reordering error). We compare source terms of gold-testset that are not found in
the results under RE from Tables 5 and 6, and the source-side of the training corpus (cf. Table
we see that NMT commits many fewer terminol- 1). We see that 88 source terms (out of a total
ogy translation-related reordering errors than PB- of 3,064 terms) are not found in the training data,
SMT. 15 REs are caught in the English-to-Hindi with almost all being multiword terms. Neverthe-
PB-SMT task compared to 5 in the English-to- less, only 5 unique words (i.e. adjudicary, hals-
Hindi NMT task. The same trend is observed with bury, presuit, decretal, adj) that are either single-
the reverse direction, with 18 reordering errors word terms or words of multiword terms are not
seen in the Hindi-to-English PB-SMT task com- found in the training data. In other words, these
pared to 5 in the Hindi-to-English NMT task. As are out-of-vocabulary (OOV) items.
442
Table 4: STC in English-to-Hindi PB-SMT and ing the strength of the open-vocabulary translation
NMT. technique (Sennrich et al., 2016). However, this
method also has down-sides. For example, some
STC (PB-SMT) translation (NMT) class (NMT)
of the term translations under the OE category in
adjudicatory role nyaay - nirnay keea PE the NMT task are non-existent wordforms of the
koee bhoomika
decretal TD target language, for which the open-vocabulary
halsbury ’s laws halbury ke kaanoonon PE translation technique is responsible. This phe-
presuit poorva vaad RE nomenon is also corroborated by Farajian et al.
adjudicatory TD
learned adj vidvat edeeje CTR (2017) while translating technical domain terms.
learned adj kaabil edeeje CTV We discuss the OE class further below.
mrtp act mrtp adhiniyam CTR
testatrix testrex OE We see from Table 5 that the human evalu-
concealments rahasyon OE ator has marked 24 term translations with OE
res judicata nyaayik roop OE in NMT. In this category we observed that the
subjudice vichaaraadheen CTV
translations of the source terms are usually either
strange words that have no relation to the meaning
of the source term, repetitions of other translated
We recall Table 5 where we see that the man- words or terms, entities that are non-existent word-
ual evaluator has marked 12 term translations with forms of the target language, or words with typo-
STC since in those cases the PB-SMT system graphical errors. As far as PB-SMT is concerned,
copied source terms (or a part of source terms) we see from Table 5 that the evaluator also tagged
verbatim into the target. In Table 4, we show 12 term translations with OE, most of which are
those source terms in the PB-SMT task that be- related to typographical errors.
long to the STC class. The first column of the
table shows source terms with the term itself or Now we turn our focus on the Hindi-to-English
part of it in bold, which means those words are task. We counted the number of those source
copied verbatim into target. We see from the ta- terms from gold-testset that are not found in the
ble that the OOV terms (i.e. adjudicary, halsbury, source-side (Hindi) of the training corpus (cf. Ta-
presuit, decretal, adj), in most cases, are respon- ble 1). We see that 160 source terms (out of a to-
sible for the term translations being marked with tal of 3,064 terms) are not found in the training
the STC tag. In one instance we found that a part data, most of which are, in fact, multiword terms.
of the English term (‘mrtp’) (cf. row 8 of Table However, only 18 unique Hindi words that are ei-
4) itself was present in the target-side of the train- ther single-word terms or words within multiword
ing corpus. This could be the possible reason why terms are not found in the training data. As in
‘mrtp’ is seen in the target translation. Each of English-to-Hindi translation task, in this task we
the remaining source terms (last 4 rows of Table found that the OOV items are largely responsi-
4) include words that are copied directly into the ble for the term translations being marked as STC.
target translation despite the fact that they are not We also examined how the Hindi-to-English NMT
OOVs. This is a well-known problem in PB-SMT system performed with those 17 source terms that
and rarely happens with the low frequency words were marked as STC. We see that HENS makes
of the training corpus. In short, these source terms a mistake on 13 occasions and correctly trans-
(last 4 entries of Table 4) either alone or with the lates on 4 occasions. The error types are spread
adjacent words of the test set sentences (i.e. as a over different categories: TD (2), OE (6), PE (1)
part of phrase) are not found in the source-side of and ILS (4). We observed that 3 out of 4 source
the PB-SMT phrase table. terms of the STC category for which the Hindi-to-
Now we see how NMT performed with the 12 English NMT system produces correct translations
source terms above; their translations with EHNS are OOV items. Here, we again see the strength
and the corresponding manual class are shown in of the open-vocabulary translation technique for
the second and third columns of Table 4, respec- the translation of novel terms. In the Hindi-to-
tively. We see that out of 12 translations EHNS English translation task, we found that the termi-
made a mistake on 8 occasions and correctly trans- nology translations under the OE category, as in
lated on 4 occasions. The errors are spread over English-to-Hindi translation, are roughly related
different categories (e.g. TD, OE, PE). Unsurpris- to odd translations, non-existent wordforms of the
ingly, we see NMT is capable of correctly trans- target language, typological mistakes and repeti-
lating rare and even unknown words, by exploit- tion of other translated words or terms.
443
DIT We see from Table 5 and Table 6 that the commit fewer lexical, reordering and morpholog-
manual evaluator marked 3 and 1 term transla- ical errors than the PB-SMT systems. The dif-
tions as DIT (disambiguation issue in target) in ferences in error rates of the former (lexical se-
English-to-Hindi and Hindi-to-English PB-SMT lection and reordering errors) types are statisti-
tasks, respectively. We found that the MT sys- cally significant in both MT tasks, and the dif-
tems made correct lexical choices for the source ference of the morphological error rates is statisti-
terms, although the meanings of their target- cally significant in the Hindi-to-English task. The
equivalents in the respective translations are dif- morphological errors are seen relatively more of-
ferent to those of the source terms. This can be ten in PB-SMT than in NMT when translation is
viewed as a cross-lingual disambiguation prob- performed from a morphologically-rich language
lem. For example, one of the three source terms (Hindi) to the a morphologically-poor language
from English-to-Hindi translation task is ‘victim’ (English). The opposite picture is observed in the
(reference translation ‘shikaar’) and the English- case of term omission in translation, with NMT
to-Hindi PB-SMT system makes a correct lexical omitting more terms in translation than PB-SMT.
choice (‘shikaar’) for ‘victim’, although the mean- We found that the difference in term omission-
ing of ‘shikaar’ is completely different in the target related error rates in PB-SMT and NMT are statis-
translation, i.e. here, its meaning is equivalent to tically significant in the Hindi-to-English task, i.e.
English ‘hunt’. again from the morphologically-rich language to
the morphologically-poor language. Another im-
Pairwise Overlap We report the numbers of portant finding from our analysis is that NMT is
pairwise overlaps, i.e. the number of instances in able to correctly translate unknown terms, by ex-
which NMT and PB-SMT have identical classifi- ploiting the strength of the open-vocabulary trans-
cation outcomes. We recall Table 5 & 6 whose lation technique, which, as expected, are copied
fourth columns show the numbers of pairwise verbatim into the target in PB-SMT. We also found
overlap for categories. The small number of over- that the majority of the errors made by the PB-
lapping instances in each category indicates that SMT system are complementary to those made
term translation errors from the PB-SMT system by the NMT system. In NMT, we observed that
are quite different from those from the NMT sys- translations of source terms are occasionally found
tem. As can be seen from the last row of Table 5 to be strange words that have no relation to the
& 6, the numbers of overlaps in the combination source term, non-existent wordforms of the tar-
of all error classes are 86 and 115, respectively, get language, and/or repetition of other translated
which are nearly one third or fourth of the num- words. This study also shows that a notable por-
ber of errors committed by the NMT and PB-SMT tion of the term translations by the MT systems
systems alone, indicating that the majority of the are diverse, which needs to be taken into consid-
errors in PB-SMT are complementary with those eration while investigating the automation of the
in NMT. This finding on terminology translation is terminology translation evaluation process.
corroborated by Popović (2017), who finds com- As far as future work is concerned, we plan
plementarity with the various issues relating to the to test terminology translation with different lan-
translations of NMT and PB-SMT. guage pairs and domains.
444
References Koehn, P., Hoang, H., Birch, A., Callison-Burch, C.,
Federico, M., Bertoldi, N., Cowan, B., Shen, W.,
Bahdanau, D., Cho, K., and Bengio, Y. (2015). Neural Moran, C., Zens, R., Dyer, C., Bojar, O., Constantin,
machine translation by jointly learning to align and A., College, W., and Herbst, E. (2007). Moses:
translate. In 3rd International Conference on Learn- Open source toolkit for statistical machine transla-
ing Representations (ICLR 2015), San Diego, CA. tion. In ACL 2007, Proceedings of the Interactive
Bentivogli, L., Bisazza, A., Cettolo, M., and Federico, Poster and Demonstration Sessions, pages 177–180,
M. (2016). Neural versus phrase-based machine Prague, Czech Republic.
translation quality: a case study. In Proceedings of Koehn, P., Och, F. J., and Marcu, D. (2003). Statisti-
the 2016 Conference on Empirical Methods in Nat- cal phrase-based translation. In HLT-NAACL 2003:
ural Language Processing, pages 257–267, Austin, conference combining Human Language Technology
Texas. conference series and the North American Chap-
Bojar, O., Diatka, V., Rychlý, P., Straňák, P., Suchomel, ter of the Association for Computational Linguistics
V., Tamchyna, A., and Zeman, D. (2014). HindEn- conference series, pages 48–54, Edmonton, AB.
Corp – Hindi-English and Hindi-only corpus for ma- Kunchukuttan, A., Mehta, P., and Bhattacharyya, P.
chine translation. In Proceedings of the Ninth Inter- (2017). The IIT Bombay English–Hindi parallel
national Language Resources and Evaluation Con- corpus. CoRR, 1710.02855.
ference (LREC’14), pages 3550–3555, Reykjavik,
Iceland. Lommel, A. R., Uszkoreit, H., and Burchardt, A.
(2014). Multidimensional Quality Metrics (MQM):
Burchardt, A., Macketanz, V., Dehdari, J., Heigold, G., A framework for declaring and describing transla-
Peter, J.-T., and Williams, P. (2017). A linguistic tion quality metrics. Tradumática: tecnologies de la
evaluation of rule-based, phrase-based, and neural traducció, (12):455–463.
mt engines. The Prague Bulletin of Mathematical
Linguistics, 108(1):159–170. Macketanz, V., Avramidis, E., Burchardt, A., Helcl,
J., and Srivastava, A. (2017). Machine trans-
Cohen, J. (1960). A coefficient of agreement for nom- lation: Phrase-based, rule-based and neural ap-
inal scales. Educational and Psychological Mea- proaches with linguistic evaluation. Cybernetics
surement, 20(1):37–46. and Information Technologies, 17(2):28–43.
Farajian, M. A., Turchi, M., Negri, M., Bertoldi, N., Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J.
and Federico, M. (2017). Neural vs. phrase-based (2002). Bleu: a method for automatic evaluation
machine translation in a multi-domain scenario. In of machine translation. In ACL-2002: 40th Annual
Proceedings of the 15th Conference of the European Meeting of the Association for Computational Lin-
Chapter of the Association for Computational Lin- guistics, pages 311–318, Philadelphia, PA. ACL.
guistics: Volume 2, Short Papers, pages 280–284,
Valencia, Spain. Popović, M. (2017). Comparing language related is-
sues for nmt and pbmt between German and En-
Haque, R., Hasanuzzaman, M., and Way, A. (2019). glish. The Prague Bulletin of Mathematical Linguis-
TermEval: An automatic metric for evaluating ter- tics, 108(1):209–220.
minology translation in MT. In Proceedings of CI-
CLing 2019, the 20th International Conference on Sennrich, R., Haddow, B., and Birch, A. (2016). Neu-
Computational Linguistics and Intelligent Text Pro- ral machine translation of rare words with subword
cessing, La Rochelle, France. units. In Proceedings of the 54th Annual Meeting of
the Association for Computational Linguistics (Vol-
Haque, R., Penkale, S., and Way, A. (2018). ume 1: Long Papers), pages 1715–1725, Berlin,
TermFinder: log-likelihood comparison and phrase- Germany.
based statistical machine translation models for
bilingual terminology extraction. Language Re- Specia, L., Harris, K., Blain, F., Burchardt, A., Mack-
sources and Evaluation, 52(2):365–400. etanz, V., Skadiņa, I., Negri, M., and Turchi, M.
(2017). Translation quality and productivity: A
Hassan, H., Aue, A., Chen, C., Chowdhary, V., Clark, study on rich morphology languages. In Proceed-
J., Federmann, C., Huang, X., Junczys-Dowmunt, ings of MT Summit XVI, the 16th Machine Transla-
M., Lewis, W., Li, M., Liu, S., Liu, T., Luo, R., tion Summit, pages 55–71, Nagoya, Japan.
Menezes, A., Qin, T., Seide, F., Tan, X., Tian, F.,
Wu, L., Wu, S., Xia, Y., Zhang, D., Zhang, Z., and Tiedemann, J. (2012). Parallel data, tools and inter-
Zhou, M. (2018). Achieving human parity on au- faces in OPUS. In Proceedings of the 8th Inter-
tomatic chinese to english news translation. CoRR, national Conference on Language Resources and
abs/1803.05567. Evaluation (LREC’2012), pages 2214–2218, Istan-
bul, Turkey.
Koehn, P. (2004). Statistical significance tests for ma-
chine translation evaluation. In Lin, D. and Wu, Toral, A. and Way, A. (2018). What level of quality can
D., editors, Proceedings of the 2004 Conference on neural machine translation attain on literary text?
Empirical Methods in Natural Language Processing In Translation Quality Assessment, pages 263–287.
(EMNLP), pages 388–395, Barcelona, Spain. Springer.
445
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J.,
Jones, L., Gomez, A. N., Kaiser, L., and Polo-
sukhin, I. (2017). Attention is all you need. CoRR,
abs/1706.03762.
Vaswani, A., Zhao, Y., Fossum, V., and Chiang, D. Table 6: PB-SMT vs NMT: Hindi-to-English.
(2013). Decoding with large-scale neural language
PB-SMT NMT ∩ p-value
models improves translation. In Proceedings of the
2013 Conference on Empirical Methods in Natu- CTR 2,313 2,295 2,075
ral Language Processing, pages 1387–1392, Seattle, CTV 241 245 147
Washington, USA. VM 24 33 5
CRF 13 11 4
Yeh, A. (2000). More accurate tests for the statisti- CIF 75 107 48
cal significance of result differences. In Proceed- CRIF 2
ings of the 18th conference on Computational lin- OC 2 18
guistics - Volume 2, COLING 2000, pages 947–953, CT 2,668 2,711 2,483
Saarbrücken, Germany.
RE 18 5 1 0.008
IE 118 76 21 0.0009
A Supplementary Material PE 65 73 31 0.42
ILS 139 90 35 0.0001
TD 38 86 6 0.0001
Table 5: PB-SMT vs NMT: English-to-Hindi. STC 17
OE 23
PB-SMT NMT ∩ p-value DIT 1
CTR 1,907 2,015 1662 ERROR 396 353 115 0.04
CTV 695 662 466
VM 35 36 10
CRF 4 7 4
CIF 112 87 31
CRIF
OC 8 4
CT 2,761 2,811 2614
RE 15 5 0.044
IE 79 77 30 0.91
PE 52 47 19 0.61
ILS 77 44 9 0.001 Figure 1: Terminology Translation Typology.
TD 53 56 9 0.83
STC 12 RE
OE 12 24 2
DIT 3 IE
ILS
Error
TD
STC
DIT
OE
Term Translation
CTR
CTV
VM
CT CRF
CIF
CRIF
OC
446
Beyond English-Only Reading Comprehension:
Experiments in Zero-Shot Multilingual Transfer for Bulgarian
447
Proceedings of Recent Advances in Natural Language Processing, pages 447–459,
Varna, Bulgaria, Sep 2–4 2019.
Finally, we address the resource scarceness in These datasets brought a variety of models and ap-
low-resource languages and the absence of ques- proaches. The usage of external knowledge has
tion contexts in our dataset by extracting relevant been an interesting topic, e.g., Chen et al. (2017a)
passages from Wikipedia articles. used Wikipedia knowledge for answering open-
domain questions, Pan et al. (2018) applied entity
Our contributions are as follows: discovery and linking as a source of prior knowl-
edge. Sun et al. (2019b) explored different read-
• We introduce a new dataset for reading com-
ing strategies such as back and forth reading, high-
prehension in a low-resource language such
lighting, and self-assessment. Ni et al. (2019) fo-
as Bulgarian. The dataset contains a total of
cused on finding essential terms and removing dis-
2,636 multiple-choice questions without con-
traction words, followed by reformulation of the
texts from matriculation exams and online
question, in order to find better evidence before
quizzes. These questions cover a large vari-
sending a query to the MRC system. A simpler ap-
ety of science topics in biology, philosophy,
proach was presented by Clark et al. (2016), who
geography, and history.
leveraged information retrieval, corpus statistics,
• We study the effectiveness of zero-shot trans- and simple inference over a semi-automatically
fer from English to Bulgarian for the task of constructed knowledge base for answering fourth-
multiple-choice reading comprehension, us- grade science questions.
ing Multilingual and Slavic BERT (Devlin Current state-of-the-art approaches in machine
et al., 2019), fine-tuned on large corpora, reading comprehension are grounded on transfer
such as RACE (Lai et al., 2017). learning and fine-tuning of language models (Pe-
ters et al., 2018; Conneau et al., 2018; Devlin et al.,
• We design a general-purpose pipeline1 for ex- 2019). Yang et al. (2019a) presented an open-
tracting relevant contexts from an external domain extractive reader based on BERT (De-
corpus of unstructured documents using in- vlin et al., 2019). Radford et al. (2018) used
formation retrieval. generative pre-training of a Transformer (Vaswani
The rest of this paper is organized as follows: et al., 2017) as a language model, transferring it to
The next section presents related work. Sec- downstream tasks such as natural language under-
tion 3 describes our approach. Details about the standing, reading comprehension, etc.
newly-proposed multiple-choice Bulgarian dataset Finally, there has been a Bulgarian MRC dataset
are given in Section 4. All experiments are de- (Peñas et al., 2012). It was used by Simov et al.
scribed in Section 5. Finally, Section 6 concludes (2012), who converted the question-answer pairs
and points to possible directions for future work. to declarative sentences, and measured their sim-
ilarity to the context, transforming both to a bag
2 Related Work of linguistic units: lemmata, POS tags, and depen-
2.1 Machine Reading Comprehension dency relations.
The growing interest in machine reading com-
2.2 (Zero-Shot) Multilingual Models
prehension (MRC) has led to the release of var-
ious datasets for both extractive (Nguyen et al., Multilingual embeddings helped researchers to
2016; Trischler et al., 2017; Joshi et al., 2017; Ra- achieve new state-of-the-art results on many NLP
jpurkar et al., 2018; Reddy et al., 2019) and non- tasks. While many pre-trained model (Grave et al.,
extractive (Richardson et al., 2013; Peñas et al., 2018; Devlin et al., 2019; Lample and Conneau,
2014; Lai et al., 2017; Clark et al., 2018; Mihaylov 2019) are available, the need for task-specific data
et al., 2018; Sun et al., 2019a) comprehension. in the target language still remains. Learning such
Our work primarily focuses on the non-extractive models is language-independent, and representa-
multiple-choice type, designed by educational ex- tions for common words remain close in the latent
perts, since their task is very close to our newly- vector space for a single language, albeit unrelated
proposed dataset, and are expected to be well- for different languages. A possible approach to
structured and error-free (Sun et al., 2019a). overcome this effect is to learn an alignment func-
1
The dataset and the source code are available at http: tion between spaces (Artetxe and Schwenk, 2018;
//github.com/mhardalov/bg-reason-BERT Joty et al., 2017).
448
Moreover, zero-shot application of fine-tuned 3.1 Context Retriever
multilingual language models (Devlin et al., 2019; Most public datasets for reading comprehension
Lample and Conneau, 2019) on XNLI (Conneau (Richardson et al., 2013; Lai et al., 2017; Sun
et al., 2018), a corpus containing sentence pairs et al., 2019a; Rajpurkar et al., 2018; Reddy et al.,
annotated with textual entailment and translated 2019; Mihaylov et al., 2018) contain not only
into 14 languages, has shown very close results to questions with possible answers, but also an ev-
such by a language-specific model. idence passage for each question. This limits
Zero-shot transfer and multilingual models had the task to question answering over a piece of
been a hot topic in (neural) machine transla- text, while an open-domain scenario is much more
tion (MT) in the past several years. Johnson challenging and much more realistic. Moreover,
et al. (2017) introduced a simple tweak to a stan- a context in which the answer can be found is
dard sequence-to-sequence (Sutskever et al., 2014) not easy to retrieve, sometimes even for a domain
model by adding a special token to the encoder’s expert. Finally, data scarceness in low-resource
input, denoting the target language, allowing a languages poses further challenges for finding re-
zero-shot learning for new language pairs. Re- sources and annotators.
cent work in zero-resource translation outlined In order to enable search for appropriate pas-
different strategies for learning to translate with- sages for non-English questions, we created an in-
out having a parallel corpus between the two tar- verted index from Wikipedia articles using Elas-
get languages. First, a many-to-one approach was ticsearch.2 We used the original dumps for the en-
adopted by Firat et al. (2016) based on building a tire Wikipage,3 and we preprocessed the data leav-
corpus from a single language paired with many ing only plain textual content, e.g., removing links,
others, allowing simultaneous training of multiple HTML tags, tables, etc. Moreover, we split the ar-
models, with a shared attention layer. A many- ticle’s body using two strategies: a sliding window
to-many relationship between languages was later and a paragraph-based approach. Each text piece
used by Aharoni et al. (2019), in an attempt to train with its corresponding article title was processed
a single Transformer (Vaswani et al., 2017) model. by applying word-based tokenization, lowercas-
Pivot-language approaches can also be used ing, stop-words removal, stemming (Nakov, 2003;
to overcome the lack of parallel corpora for the Savoy, 2007), and n-gram extraction. Finally, the
source–target language pair. Chen et al. (2017b) matching between a question and a passage was
used a student-teacher framework to train an NMT done using cosine similarity and BM25 (Robert-
model, using a third language as a pivot. A simi- son and Zaragoza, 2009).
lar idea was applied to MRC by Asai et al. (2018),
who translated each question to a pivot language, 3.2 BERT for Multiple-Choice RC
and then found the correct answer in the target lan- The recently-proposed BERT (Devlin et al., 2019)
guage using soft-alignment attention scores. framework is applicable to a vast number of NLP
tasks. A shared characteristic between all of them
3 Model is the form of the input sequences: a single sen-
tence or a pair of sentences separated by the [SEP]
Our model has three components: (i) a context re- special token, and a classification token ([CLS])
trieval module, which tries to find good explana- added at the beginning of each example. In con-
tory passages for each question-answer pair, from trast, the input for multiple-choice reading com-
a corpus of non-English documents, as described prehension questions is assembled by three sen-
in Section 3.1, (ii) a multiple-choice reading com- tence pieces, i.e., context passage, question, and
prehension module pre-trained on English data possible answer(s). Our model follows a simple
and then applied to the target language in a zero- strategy of concatenating the option (candidate an-
shot fashion, i.e., without further training or ad- swer) at the end of a question. Following the nota-
ditional fine-tuning, to a target (non-English) lan- tion of Devlin et al. (2019), the input sequence can
guage, as described in Section 3.2, and (iii) a vot- be written as follows:
ing mechanism, described in Section 3.3, which
combines multiple passages from (i) and their [CLS] Passage [SEP] Question + Option [SEP]
scores from (ii) in order to obtain a single (most 2
http://www.elastic.co/
probable) answer for the target question. 3
http://dumps.wikimedia.org/
449
In our experiments below, we adopt a simple sum-
ming strategy. We evaluate each result from the
context retriever against the question and the pos-
sible options (see Section 3.2 for more details),
thus obtaining a list of raw probabilities. We
found empirically that explanatory contexts as-
sign higher probability to the related answer, while
general or uninformative passages lead to stratifi-
cation of the probability distribution over the an-
swer options. We formulate this as follows:
exp(BERT (p, q + aj ))
P r(aj |p; q) = P , (1)
j0 exp(BERT (p, q + aj 0))
450
Domain #QA-pairs #Choices Len Question Len Options Vocabulary Size
12th Grade Matriculation Exam
Biology 437 4 10.4 2.6 2, 414 (12, 922)
Philosophy 630 4 8.9 2.9 3, 636 (20, 392)
Geography 612 4 12.8 2.5 3, 239 (17, 668)
History 542 4 23.7 3.6 5, 466 (20, 456)
Online History Quizzes
Bulgarian History 229 4 14.0 2.8 2, 287 (10, 620)
PzHistory 183 3 38.9 2.4 1, 261 (7, 518)
Overall 2, 633 3.9 15.7 2.9 13, 329 (56, 104)
RACE Train - Mid and High School
RACE-M 25, 421 4 9.0 3.9 32, 811
RACE-H 62, 445 4 10.4 5.8 125, 120
Overall 87, 866 4 10.0 5.3 136, 629
Table 1: Statistics about our Bulgaria dataset compared to the RACE dataset.
451
Finally, we examine the vocabulary richness of the 5.2 Wikipedia Retrieval and Indexing
two datasets. The total number of unique words is
Here, we discuss the retrieval setup (see Sec-
shown in the last column of Table 1 (Vocab Size).
tion 3.1 for more details). We use the Bulgarian
For our dataset, there are two numbers per row: the
dump of Wikipedia from 2019-04-20, with a total
first one shows statistics based on the question–
of 251,507 articles. We index each article title and
answer pairs only, while the second one, enclosed
body in plain text, which we call a passage. We
in parentheses, measures the vocabulary size in-
further apply additional processing for each field:
cluding the extracted passages by the Context Re-
triever. The latter number is a magnitude estimate • ngram: word-based 1–3 grams;
rather then a concrete number, since its upper limit • bg: lowercased, stop-words removed (from
is the number of words in Wikipedia, and it can Lucene), and stemmed (Savoy, 2007);
vary for different retrieval strategies. • none: bag-of-words index.
5 Experiments and Evaluation We ended up using a subset of four fields
from all the possible analyzer-field combinations,
5.1 BERT Fine-Tuning namely title.bg, passage, passage.bg, and pas-
We divide the fine-tuning into two groups of mod- sage.ngram. We applied Bulgarian analysis on the
els (i) Multilingual BERT, and (ii) Slavic BERT. title field only as it tends to be short and descrip-
Table 3 below presents the results in the multiple- tive, and thus very sensitive to noise from stop-
choice comprehension task on the dev dataset words, which is in contrast to questions that are
from RACE (Lai et al., 2017). formed mostly of stop-words, e.g., what, where,
when, how.
#Epoch RACE-M RACE-H Overall For indexing the Wikipedia articles, we adopt
BERT 1 64.21 53.66 56.73 two strategies: sliding window and paragraph. In
BERT 2 68.80 57.58 60.84 the window-based strategy, we define two types of
BERT 3 69.15 58.43 61.55 splits: small, containing 80-100 words, and large,
Slavic 2 53.55 44.48 47.12 of around 300 words. In order to obtain indexing
Slavic 3 57.38 46.88 49.94 chunks, we define a window of size K, and a stride
equal to one forth of K. Hence, each K4 charac-
Table 3: Accuracy measured on the dev RACE ters, which is the size of the stride, are contained
dataset after each training epoch. into four different documents. The paragraph-
based strategy divides the article by splitting it
Multilingual BERT As our initial model, we
using one or more successive newline characters
use BERTbase , Multilingual Cased which is pre-
([\n]+) as a delimiter. We avoid indexing entire
trained on 104 languages, and has 12-layers, 768-
documents due to their extensive length, which
hidden units per layer, 12-heads, and a total of
can be far beyond the maximum length that BERT
110M parameters. We further fine-tune the model
can take as an input, i.e., 320 word pieces (see
on RACE (Lai et al., 2017) for 3 epochs saving a
Section 5.1 for the more details). Note that ex-
checkpoint after each epoch. We use a batch size
tra steps are needed in order to extract a proper
of 8, a max sequence size of 320, and a learning
passage from the text. Moreover, the amount of
rate of 1e-5.
facts in the Wikipedia articles that are unrelated to
Slavic BERT The Slavic model4 was built us- our questions give rise to false positives since the
ing transfer learning from the Multilingual BERT question is short and term-unspecific.
model to four Slavic languages: Bulgarian, Czech, Finally, we use a list of top-N hits for each can-
Polish, and Russian. In particular, the Multilin- didate answer. Thus, we have to execute an addi-
gual BERT model was fine-tuned on a stratified tional query for each question + option combina-
dataset of Russian news and Wikipedia articles for tion, which may result in duplicated passages, thus
the other languages. We use this pre-trained Slavic introducing an implicit bias towards the candidates
BERT model, and we apply the same learning pro- they support. In order to mitigate this effect, dur-
cedure as for Multilingual BERT. ing the answer selection phase (see Section 3.3),
4
http://github.com/deepmipt/ we remove all duplicate entries, keeping a single
Slavic-BERT-NER instance.
452
Setting Accuracy Zero-Shot Transfer. Here, we assess the perfor-
mance of our model when applied to Bulgarian
Random 24.89
multiple-choice reading comprehension. Table 4
Train for 3 epochs – presents an ablation study for various components.
+ window & title.bg & pass.ngram 29.62 Each line denotes the type of the model, and the
+ passage.bg & passage 39.35 addition (+) or the removal (–) of a characteristic
– title.bg 39.69 from the setup in the previous line. The first line
+ passage.bg^2 40.26 shows the performance of a baseline model that
+ title.bg^2 40.30 chooses an option uniformly at random from the
+ bigger window 36.54 list of candidate answers for the target question.
+ paragraph split 42.23 The following rows show the results for experi-
+ Slavic pre-training 33.27 ments conducted with a model trained for three
Train for 1 epoch best 40.26 epochs on RACE (Lai et al., 2017).
Train for 2 epochs best 41.89
Our basic model uses the following setup:
Table 4: Accuracy on the Bulgarian testset: abla- Wikipedia pages indexed using a small sliding
tion study when sequentially adding/removing dif- window (400 characters, and stride of 100 charac-
ferent model components. ters), and context retrieval over two fields: Bulgar-
ian analyzed title (text.bg), and word n-grams over
5.3 Experimental Results the passage (passage.ngram). This setup yields
29.62% accuracy, and it improves over the ran-
Here, we discuss the accuracy of each model on dom baseline by 4.73% absolute. We can think
the original English MRC task, followed by ex- of it as a non-random baseline for further exper-
periments in zero-shot transfer to Bulgarian. iments. Next, we add two more fields to the IR
query: passage represented as a bag of words
English Pre-training for MCRC. Table 3
(named passage), and Bulgarian analyzed (pas-
presents the change in accuracy on the original En-
sage.bg), which improves the accuracy by addi-
glish comprehension task, depending on the num-
tional 10%, arriving at 39.35%. The following
ber of training epochs. In the table, “BERT” refers
experiment shows that removing the title.bg field
to the Multilingual BERT model, while “Slavic”
does not change the overall accuracy, which makes
stands for BERT with Slavic pre-training. We fur-
it an insignificant field for searching. Further, we
ther fine-tune the models on the RACE dataset.
add double weight on passage.bg, (shown as ^2),
Next, we report their performance in terms of ac-
which yields 1% absolute improvement.
curacy, following the notation from (Lai et al.,
2017). Note that the questions in RACE-H are From the experiments described above, we
more complex than those in RACE-M. The lat- found the best combination of query fields to be
ter has more word matching questions and fewer title.bulgarian^2, passage.ngram, passage, pas-
reasoning questions. The final column in the ta- sage.bulgarian^2, where the title has a minor con-
ble, Overall, shows the accuracy calculated over tribution, and can be sacrificed for ease of com-
all questions in the RACE testset. We train both putations and storage. Fixing the best query
setups for three epochs and we report their per- fields, allowed us to evaluate other indexing strate-
formance after each epoch. We can see a pos- gies, i.e., bigger window (size 1,600, stride 400)
itive correlation between the number of epochs with accuracy 36.54%, and paragraph splitting,
and the model’s accuracy. We further see that the with which we achieved our highest accuracy of
Slavic BERT performs far worse on both RACE- 42.23%. This is an improvement of almost 2.0%
M and RACE-H, which suggests that the change absolute over the small sliding window, and 5.7%
of weights of the model towards Slavic languages over the large one.
has led to catastrophic forgetting of the learned Next, we examined the impact of the Slavic
English syntax and semantics. Thus, it should be BERT. Surprisingly, it yielded 9% absolute drop
expected that the adaptation to Slavic languages in accuracy compared to the multi-lingual BERT.
would yield decrease in performance for English. This suggests that the latter already has enough
What matters though is whether this helps when knowledge about Bulgarian, and thus it does not
testing on Bulgarian, which we explore next. need further adaptation to Slavic languages.
453
We can see that the highest accuracy is observed
for history, particularly for online quizzes, which
are not designed by educators and are more of a
word-matching nature rather then a reasoning one
(see Table 2). Finally, geography appears to be
the hardest category with only 38.73% accuracy:
3.5% absolute difference compared to the second-
worst category. The performance for this subject
is also affected differently by changes in query re-
sult length: the peak is at lengths 5 and 10, while
there is a drop for length 2. A further study of the
model’s behavior can be found in Appendix B.
454
References Orhan Firat, Baskaran Sankaran, Yaser Al-Onaizan,
Fatos T. Yarman Vural, and Kyunghyun Cho. 2016.
Roee Aharoni, Melvin Johnson, and Orhan Firat. 2019. Zero-resource translation with multi-lingual neural
Massively multilingual neural machine translation. machine translation. In Proc. of the Conference on
In Proceedings of the Conference of the North Empirical Methods in Natural Language Process-
American Chapter of ACL. Minneapolis, MN, USA, ing. Austin, TX, USA, EMNLP ’16, pages 268–277.
NAACL-HLT ’19, pages 3874–3884.
Mikel Artetxe and Holger Schwenk. 2018. Mas- Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Ar-
sively multilingual sentence embeddings for zero- mand Joulin, and Tomas Mikolov. 2018. Learning
shot cross-lingual transfer and beyond. arXiv word vectors for 157 languages. In Proceedings of
preprint arXiv:1812.10464 . the Conference on Language Resources and Evalu-
ation. Miyazaki, Japan, LREC ’18.
Akari Asai, Akiko Eriguchi, Kazuma Hashimoto, and
Yoshimasa Tsuruoka. 2018. Multilingual extractive Jeremy Howard and Sebastian Ruder. 2018. Universal
reading comprehension by runtime machine transla- language model fine-tuning for text classification. In
tion. arXiv preprint arXiv:1809.03275 . Proceedings of the 56th Annual Meeting of the Asso-
ciation for Computational Linguistics. Melbourne,
Danqi Chen, Adam Fisch, Jason Weston, and Antoine
Australia, ACL ’18, pages 328–339.
Bordes. 2017a. Reading Wikipedia to answer open-
domain questions. In Proceedings of the Meeting of
the Association for Computational Linguistics. Van- Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim
couver, Canada, ACL ’17, pages 1870–1879. Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat,
Fernanda Viégas, Martin Wattenberg, Greg Corrado,
Yun Chen, Yang Liu, Yong Cheng, and Victor O.K. Macduff Hughes, and Jeffrey Dean. 2017. Google’s
Li. 2017b. A teacher-student framework for zero- multilingual neural machine translation system: En-
resource neural machine translation. In Proceed- abling zero-shot translation. Transactions of the As-
ings of the 55th Annual Meeting of the Association sociation for Computational Linguistics 5:339–351.
for Computational Linguistics. Vancouver, Canada,
ACL ’17, pages 1925–1935. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke
Zettlemoyer. 2017. TriviaQA: A large scale dis-
Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, tantly supervised challenge dataset for reading com-
Ashish Sabharwal, Carissa Schoenick, and Oyvind prehension. In Proceedings of the Annual Meeting
Tafjord. 2018. Think you have solved question an- of the Association for Computational Linguistics.
swering? Try ARC, the AI2 reasoning challenge. Vancouver, Canada, ACL ’17, pages 1601–1611.
arXiv preprint arXiv:1803.05457 .
Peter Clark, Oren Etzioni, Tushar Khot, Ashish Sab- Shafiq Joty, Preslav Nakov, Lluís Màrquez, and Israa
harwal, Oyvind Tafjord, Peter Turney, and Daniel Jaradat. 2017. Cross-language learning with ad-
Khashabi. 2016. Combining retrieval, statistics, and versarial neural networks. In Proc. of the Confer-
inference to answer elementary science questions. ence on Computational Natural Language Learning.
In Proceedings of the 13th AAAI Conference on Ar- Vancouver, Canada, CoNLL ’17, pages 226–237.
tificial Intelligence. Phoenix, AZ, USA, AAAI ’16,
pages 2580–2586. Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang,
and Eduard Hovy. 2017. RACE: Large-scale ReAd-
Alexis Conneau, Ruty Rinott, Guillaume Lample, Ad- ing comprehension dataset from examinations. In
ina Williams, Samuel Bowman, Holger Schwenk, Proceedings of the 2017 Conference on Empirical
and Veselin Stoyanov. 2018. XNLI: Evaluating Methods in Natural Language Processing. Copen-
cross-lingual sentence representations. In Proceed- hagen, Denmark, EMNLP ’17, pages 785–794.
ings of the 2018 Conference on Empirical Methods
in Natural Language Processing. Brussels, Belgium, Guillaume Lample and Alexis Conneau. 2019. Cross-
EMNLP ’18, pages 2475–2485. lingual language model pretraining. arXiv preprint
arXiv:1901.07291 .
Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Car-
bonell, Quoc Le, and Ruslan Salakhutdinov. 2019.
Transformer-XL: Attentive language models beyond Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish
a fixed-length context. In Proceedings of the Meet- Sabharwal. 2018. Can a suit of armor conduct elec-
ing of the Association for Computational Linguis- tricity? A new dataset for open book question an-
tics. Florence, Italy, ACL ’19, pages 2978–2988. swering. In Proceedings of the Conference on Em-
pirical Methods in Natural Language Processing.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Brussels, Belgium, EMNLP ’18, pages 2381–2391.
Kristina Toutanova. 2019. BERT: Pre-training of
deep bidirectional transformers for language under- Preslav Nakov. 2003. Building an inflectional stem-
standing. In Proceedings of the 2019 Annual Con- mer for Bulgarian. In Proceedings of the 4th Inter-
ference of the North American Chapter of the Asso- national Conference Conference on Computer Sys-
ciation for Computational Linguistics. Minneapolis, tems and Technologies: E-Learning. Rousse, Bul-
MN, USA, NAACL-HLT ’19, pages 4171–4186. garia, CompSysTech ’03, pages 419–424.
455
Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Matthew Richardson, Christopher J.C. Burges, and
Saurabh Tiwary, Rangan Majumder, and Li Deng. Erin Renshaw. 2013. MCTest: A challenge dataset
2016. MS MARCO: A human generated machine for the open-domain machine comprehension of
reading comprehension dataset. In Proceedings of text. In Proceedings of the 2013 Conference on Em-
the Workshop on Cognitive Computation: Integrat- pirical Methods in Natural Language Processing.
ing Neural and Symbolic Approaches. Barcelona, Seattle, WA, USA, EMNLP ’13, pages 193–203.
Spain, CoCo@NIPS ’16.
Stephen Robertson and Hugo Zaragoza. 2009. The
Jianmo Ni, Chenguang Zhu, Weizhu Chen, and Ju- probabilistic relevance framework: BM25 and be-
lian McAuley. 2019. Learning to attend on essen- yond. Found. Trends Inf. Retr. 3(4):333–389.
tial terms: An enhanced retriever-reader model for Jacques Savoy. 2007. Searching strategies for the Bul-
open-domain question answering. In Proceedings garian language. Inform. Retrieval 10(6):509–529.
of the Conference of the North American Chapter
of ACL. Minneapolis, MN, USA, NAACL-HLT ’19, Kiril Ivanov Simov, Petya Osenova, Georgi Georgiev,
pages 335–344. Valentin Zhikov, and Laura Tolosi. 2012. Bulgarian
question answering for machine reading. In CLEF
Rodrigo Nogueira and Kyunghyun Cho. 2019. Pas- Working Note Papers. Rome, Italy.
sage re-ranking with BERT. arXiv preprint
arXiv:1901.04085 . Kai Sun, Dian Yu, Jianshu Chen, Dong Yu, Yejin
Choi, and Claire Cardie. 2019a. DREAM: A chal-
Xiaoman Pan, Kai Sun, Dian Yu, Heng Ji, and Dong lenge data set and models for dialogue-based read-
Yu. 2018. Improving question answering with ex- ing comprehension. Transactions of the Association
ternal knowledge. arXiv preprint:1902.00993 . for Computational Linguistics 7:217–231.
Kai Sun, Dian Yu, Dong Yu, and Claire Cardie. 2019b.
Anselmo Peñas, Eduard Hovy, Pamela Forner, Álvaro Improving machine reading comprehension with
Rodrigo, Richard Sutcliffe, Corina Forascu, Yassine general reading strategies. In Proceedings of the
Benajiba, and Petya Osenova. 2012. Overview of North American Chapter of ACL. Minneapolis, MN,
QA4MRE at CLEF 2012: Question answering for USA, NAACL-HLT ’19, pages 2633–2643.
machine reading evaluation. In CLEF Working Note
Papers. Rome, Italy, pages 1–24. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural net-
Anselmo Peñas, Christina Unger, and Axel- works. In Proceedings of the 27th Annual Con-
Cyrille Ngonga Ngomo. 2014. Overview of CLEF ference on Neural Information Processing Systems.
question answering track 2014. In Information Montreal, Canada, NIPS ’14, pages 3104–3112.
Access Evaluation. Multilinguality, Multimodality,
Yi Tay, Luu Anh Tuan, and Siu Cheung Hui. 2018.
and Interaction. pages 300–306.
Multi-range reasoning for machine comprehension.
arXiv preprint arXiv:1803.09074 .
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt
Gardner, Christopher Clark, Kenton Lee, and Luke Adam Trischler, Tong Wang, Xingdi Yuan, Justin Har-
Zettlemoyer. 2018. Deep contextualized word rep- ris, Alessandro Sordoni, Philip Bachman, and Ka-
resentations. In Proceedings of the 2018 Confer- heer Suleman. 2017. NewsQA: A machine compre-
ence of the North American Chapter of the Associ- hension dataset. In Proceedings of the 2nd Work-
ation for Computational Linguistics. New Orleans, shop on Representation Learning for NLP. Vancou-
LA, USA, NAACL-HLT ’18, pages 2227–2237. ver, Canada, RepL4NLP ’19, pages 191–200.
Alec Radford, Karthik Narasimhan, Tim Salimans, and Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Ilya Sutskever. 2018. Improving language under- Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
standing by generative pre-training . Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. In Proceedings of the Annual Conference
Alec Radford, Jeffrey Wu, Rewon Child, David Luan, on Neural Information Processing Systems. Long
Dario Amodei, and Ilya Sutskever. 2019. Language Beach, CA, USA, NIPS ’17, pages 5998–6008.
models are unsupervised multitask learners . Wei Yang, Yuqing Xie, Aileen Lin, Xingyu Li, Luchen
Tan, Kun Xiong, Ming Li, and Jimmy Lin. 2019a.
Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. End-to-end open-domain question answering with
Know what you don’t know: Unanswerable ques- BERTserini. In Proceedings of the Conference of
tions for SQuAD. In Proceedings of the Meeting of the North American Chapter of ACL. Minneapolis,
the Association for Computational Linguistics. Mel- MN, USA, NAACL-HLT ’19, pages 72–77.
bourne, Australia, ACL ’18, pages 784–789.
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Car-
Siva Reddy, Danqi Chen, and Christopher D. Manning. bonell, Ruslan Salakhutdinov, and Quoc V Le.
2019. CoQA: A conversational question answering 2019b. XLNet: Generalized autoregressive pre-
challenge. Transactions of the Association for Com- training for language understanding. arXiv preprint
putational Linguistics 7:249–266. arXiv:1906.08237 .
456
Appendix The questions in Table 6 are from five different
categories: biology, philosophy, geography, his-
A Per-Category Results tory, and online quizzes. Each of them has its own
Table 5 gives an overview, including per-category specifics and gives us an opportunity to illustrate a
breakdown, of our parameter tuning experiments. different model behavior.
We present the results for some interesting ex- The first question is from the biology domain,
periments rather then for a full grid search. The and we can see that the text is very general, and
first row shows a random baseline for each cate- so is the retrieved context. The latter talks about
gory. In the following rows, we compare different hair rather than coat, and the correct answer (D)
types of indexing: first, we show the results for morphological adaptation is not present in the re-
a small sliding window (400-character window, trieved text. On the other hand, all the terms are
and 100-character stride), followed by a big win- only connected to it, and hence the model assigns
dow (1,600-character window, and 400-character high probability to this answer option.
stride), and finally for paragraph indexing. We use For the second question, from the philosophy
the same notation as in Section 5. The last group in domain, there are two related contexts found. The
the table (Paragraph) shows the best-performing first one is quite short, noisy, and it does not give
model, where we mark in bold the highest accu- much information in general. The second para-
racy for each category. For completeness, we also graph manages to extract the definition of rela-
show the accuracy when using the Slavic BERT tivism and to give good supporting evidence for
model for prediction, which yields a 10% drop on the correct answer, namely that there is no abso-
average compared to using the Multilingual BERT, lute good and evil (B). As a result, this option is
for each of the categories. assigned high probability. Nevertheless, the incor-
rect answer here is only one moral law that is valid
B Case Study for all (A) is assigned an even higher probability
In Table 6, we present the retrieved evidence pas- and it wins the voting.
sages for the example questions in Table 2: we In the third example, from the domain of ge-
omit the answers, and we only show the ques- ography, we see a large number of possible con-
tions and the contexts. Each example is sepa- texts, due to the long and descriptive answers. We
rated by a double horizontal line, where the first can make two key observations: (i) the query is
row is the question starting with “Q:”, and the fol- drawn in very different directions by the answers,
lowing rows contain passages returned by the re- and (ii) there is no context for Southwestern re-
triever. For each context, we normalize the raw gion, and thus, in the second option, the result is
scores from the comprehension model using Eq. 1 for Russia, not for Bulgaria. The latter passage
to obtain a probability distribution. We then select pushes the probability mass to an option that talks
an answer using arg max, according to Eq. 2. In about transportation (D), which is incorrect. For-
the table, we indicate the correctness of each pre- tunately, the forth context has an almost full term
dicted answer using one of the following symbols overlap with the correct answer (B), and thus gets
before the question: very high probability assigned to it: 72%.
The fourth question, from the history domain,
3 The question is answered correctly. asks to point out a missing concept, but the query
is dominated by the question, and especially by
7 An incorrect answer has the highest score. underscores, leading to a single hit, counting only
? Two or more answers have the highest score. symbols, without any words. As expected, the
model assigned uniform probability to all classes.
We show the top retrieved result in order to il- The last question, a history quiz, is a factoid
lustrate the model scores over different evidence one, and it lacks a reasoning component, unlike
passages and the quality of the articles. The the previous examples. The query returned a sin-
queries are formed by concatenating the question gle direct match. The retrieved passage contains
with an answer option, even though this can lead to the correct answer exactly: option Bogoridi (C).
duplicate results since some answers can be quite Thereby, the comprehension model assigns to it a
similar or the question’s terms could dominate the very high probability of 68%.
similarity score.
457
#docs Overall biology-12th philosophy-12th geography-12th history-12th history-quiz
Random
0 24.89 26.09 24.44 24.18 25.87 24.03
Window Small
title.bulgarian, passage.bulgarian
1 39.95 40.27 40.63 34.97 42.99 41.99
2 40.22 40.27 40.63 35.95 42.62 42.72
5 40.22 38.90 40.63 38.07 41.51 42.48
10 38.66 40.50 39.84 35.46 39.30 38.83
20 36.84 37.53 39.05 33.82 38.75 34.71
title.bulgarian, passage.ngram
1 28.94 29.06 32.06 27.29 27.49 28.40
2 29.09 29.06 33.33 25.00 28.78 29.13
5 29.05 27.46 32.06 26.63 30.63 27.67
10 29.62 29.06 32.54 26.96 30.07 29.13
20 29.43 31.81 32.70 26.63 28.60 27.18
title.bulgarian, passage.ngram, passage, passage.bulgarian
1 38.32 38.22 40.00 34.48 39.48 40.05
2 39.08 37.07 40.32 34.48 40.59 44.17
5 39.35 40.96 39.84 34.64 41.33 41.26
10 38.63 40.50 40.63 33.50 40.41 38.83
20 36.54 38.67 37.94 31.37 37.45 38.59
passage.ngram, passage, passage.bulgarian^2
1 39.69 40.27 40.63 35.13 42.07 41.26
2 40.26 39.82 40.95 35.95 42.62 42.96
5 39.57 39.59 39.37 37.25 40.96 41.50
10 38.70 41.19 39.52 35.78 39.30 38.35
20 37.14 39.36 37.78 35.29 38.38 34.95
title.bulgarian^2, passage.ngram, passage, passage.bulgarian^2
1 39.84 40.27 40.79 35.13 42.25 41.75
2 40.30 40.27 40.63 36.11 42.80 42.72
5 40.26 39.13 40.63 38.40 41.14 42.48
10 38.74 40.50 39.68 35.62 39.48 39.08
20 37.07 37.76 39.05 34.64 38.56 34.95
Window Big
title.bulgarian^2, passage.ngram, passage, passage.bulgarian^2
1 31.22 28.38 33.97 29.41 30.81 33.25
2 33.12 31.58 37.46 31.21 33.95 29.85
5 36.04 35.70 38.10 33.82 37.82 34.22
10 36.54 37.30 36.03 33.99 39.30 36.65
20 35.62 34.55 39.68 31.05 38.38 33.74
Paragraph
title.bulgarian^2, passage.ngram, passage, passage.bulgarian^2
1 41.82 41.42 42.06 38.07 40.96 48.54
2 42.23 42.56 43.17 35.62 42.99 49.27
5 41.59 43.25 40.32 38.73 40.04 48.06
10 39.46 40.96 38.41 36.93 39.85 42.72
20 37.52 39.13 37.62 34.64 38.56 38.59
Slavic BERT
1 33.19 30.89 33.17 28.76 32.29 43.45
2 33.27 31.58 31.90 31.21 35.24 37.62
5 31.14 30.21 30.16 29.25 31.00 36.65
10 30.42 29.29 29.68 29.74 31.92 31.80
20 29.66 28.60 29.37 28.43 32.10 29.85
Table 5: Evaluation results for the Bulgarian multiple-choice reading comprehension task: comparison
of various indexing and query strategies.
458
Context P rA P rB P rC P rD
? Q: Point out the concept that is missed in the text of the Turnovo
Constitution: . . .
1) 0.26 0.26 0.26 0.22
Table 6: Retrieved unique top-1 contexts for the example questions in Table 2. The passages are retrieved
using queries formed by concatenating a question with an answer option.
459
Tweaks and Tricks for Word Embedding Disruptions
{Amir.Hazem,Nicolas.Hernandez}@univ-nantes.fr
460
Proceedings of Recent Advances in Natural Language Processing, pages 460–464,
Varna, Bulgaria, Sep 2–4 2019.
tence embeddings using StackExchange Philoso- tecture and the training procedure of word embed-
phy data set over several in-domain 1 and pre- dings (Nematzadeh et al., 2017). CBOW model
trained embedding models on question to ques- for instance, predicts the target word based on a
tion similarity task and show significant improve- mean average weights of its context words. While,
ments. The used pre-trained models are skipgram skip-gram maximizes the average log-probability
(Sg) (Mikolov et al., 2013), Glove (Pennington of each word’s context (Mikolov et al., 2013).
et al., 2014), dependency relation (Deps) (Levy Also, the process of weights computation is of-
and Goldberg, 2014), and the character Skipgram ten based on batches which leads to a mean er-
(ChSg) and character CBOW (ChC)2 (Bojanowski ror minimization instead of a specific attention to
et al., 2016). each single training example. This optimization
process might lead to some computation decisions
2 Word Embedding Disruptions that create margin values, potentially not relevant
and so creates dimension disruptions. Even with-
Word embedding models are vector-spaces in out using batches, and in order to allow efficient
which words are represented as points in an N- training, evaluating the normalization factor of the
dimensional Euclidean space. Regardless of the Softmax (Mikolov et al., 2013) for instance, intro-
complexity of the embedding models, the main duces approximations. We don’t claim that one of
concern remains weights estimation. At the end the above cited reasons is the main cause of dis-
of the training process, the obtained model is ex- ruptions, however we hypothesize that several pa-
pected to map and to efficiently represent the train- rameters may lead to training side effects that lead
ing data set. This supposes that each dimension to disruption values. If it is difficult to remove dis-
has a degree of representativeness of a given word. ruptions, we propose two tweaking techniques to
Which means that each single dimensional value reduce their effects.
can potentially affect the word’s position in the N-
dimensional space. This also means that extreme
values that we call disruptions might have a bigger
impact on the word’s position. This phenomenon
is amplified by the mathematical properties and
the additive process of bottom-up representation.
If we consider for instance a 3-dimensional space
in which one dimensional value is drastically high, (a) Skip-gram (b) CBOW
the position in the 3-D space will be attracted by
this dimension. This is not problematic on its own
if the 3-D model well maps the data. However, if
it is not the case, this might weaken the quality of
the word’s embedding vector.
Multiple reasons lend support to the idea that
disruptions are more likely to be side effects of (c) Glove (d) Deps
the training process rather than being discrimi-
native values. The first reason comes from the Figure 1: 300 dimensional word embedding dis-
characteristics of the training data set. Regard- tributions on the StackExchange Philosophy data
less of the size which in most cases greatly af- set.
fects the quality of the embedding models, not
all words are equally distributed and even using Figure 1 illustrates the distribution values of
down-sampling and other sophisticated techniques some well-known state-of-art embedding models.
to reduce the size impact, infrequent words will We first observe that all the embedding models fol-
always be under-characterized at least for embed- low a Gaussian distribution with a mean around
ding models not involving character n-gram mod- zero. The main differences concern the standard
eling. The second reason comes from the archi- deviation, minimum and maximum values and the
1 density. Also, Table 1 reports the statistics of
In-domain embedding models are embeddings trained on
the in-domain philosophy corpus. in-domain and pre-trained embeddings. We ob-
2
ChC is not available, and is only used as in-domain. serve that each model shows different characteris-
461
In-domain embedding models Pre-trained embedding models
Sg CBOW Glove ChC ChSG W2V Glove6B Glove42B Bow5C Bow5W Deps ChSG
µ -0.01 -0.003 -0.0008 0.001 -0.011 -0.003 -0.003 0.005 -0.0002 -0.0007 0.0004 0.007
σ 0.35 0.47 0.12 0.77 0.17 0.13 0.38 0.29 2 0.05 0.05 0.05 0.295
min -2.07 -5.51 -3.06 -8.01 -2.55 -4.06 -3.06 -5.09 -0.31 -0.358 -0.34 -10.1
max 2.20 7.57 2.57 9.98 2.26 4.18 3.52 3.25 0.32 0.354 0.29 13.6
disrupt (%) 7.56 17.3 7.11 15.8 22.1 8.1 18.3 17.1 36.6 37.9 38.4 21.5
Cove (%) 100 100 100 100 100 52.1 55.2 66.1 52.0 52.0 51.9 56.4
Table 1: Statistics over several in-domain and pre-trained embedding models on the Philosophy data set. The mean value
over the entire set of embeddings (µ), its corresponding standard deviation (σ), minimum (min) and maximum (max) values,
the percentage of disruptions (disrupt) and vocabulary coverage of the embedding models on the Philosophy data set (Cove).
tics which can be comparable in some cases (Sg µ - σ for min intervals. In the dimensional space
and ChSg) or totally different in other cases such this procedure tends to center the words position in
as Sg and CBOW. Also, we see that substantial di- the space and to reduce the impact of disruptions.
mensional values that we define as disruptions are From the bottom-up sentence representation side,
beyond the standard deviation. this can be interpreted as ignoring some dimen-
sions in the additive process. As can be seen in Ta-
3 Tweaks Approach ble 1, the mean is often around zero and this value
In order to reduce disruption impact on bottom- will not have any effect on the position of the sen-
up sentence representation, we introduce the mean tence in the space when we sum-up the words of
correction trick. A tweaking process that adjusts a given sentence. We consider two ways of mean
extreme values and center them around the mean computation. The first is computed over the en-
embedding value. Let’s consider N the number of tire set of word embeddings and the second one
embedding dimensions, V the corpus vocabulary is the computation of a mean per dimension. We
size and S the set of disruptions with S ∈ N × V . contrast both techniques in our experiments.
All the S values are replaced by the mean. 4 Data Description
Algorithm 1 Mean Tweak approach The data set was extracted from the philosophy
Require: Emb = SG, CBOW, Glove, ... community question answering forum StackEx-
Require: µ ← M ean(Emb) change. Basically, each post is composed of a
Require: σ ← StandardDeviation(Emb) pair of questions and one or several answers and
Require: α ∈ ]min, µ − σ] comments. Our data set contains 5.7k posts and
Require: β ∈ [µ + σ, max[ 1.1M tokens. We took 10% of questions for dev
and 10% for test set (575 questions).
1: function T WEAK(Embw ,µ,α, β) Philosophy question pair example:
2: for i ∈ Dim(Embw ) do Q1: were there any pre-enlightenment philosoph-
3: if Embw [i] 6∈ [α, β] then ical consideration of the naturalistic fallacy or re-
4: Embw [i] ← µ late concept?
5: end if Q2: naturalistic fallacy was described and named
6: end for by g.e. Moore at the beginning of the 20th century.
7: return Embw but have there been pre-enlightenment philoso-
8: end function pher who have treat the concept?
462
In-domain embedding models Pre-trained embedding models
Approach Sg CBOW Glove ChC ChSG W2V(Sg) Glove6B Glove42B Bow5C(CBOW) Bow5W(Sg) Deps ChSG
Baseline 63.6 40.8 46.5 36.8 49.6 49.6 51.0 50.1 53.1 52.0 50.3 54.5
+Norm 60.3 52.2 54.0 48.0 50.9 50.2 51.9 51.9 52.6 51.8 49.8 55.3
+MeanAll 63.9 56.1 58.7 56.4 59.7 54.1 55.2 57.3 59.2 55.8 56.4 57.8
+MeanDim 63.8 63.1 61.5 59.0 59.0 55.0 58.1 59.1 58.4 57.8 56.9 59.4
+MeanALL+Norm 60.7 53.4 54.6 55.3 59.7 54.3 55.8 56.7 59.0 55.7 56.8 57.9
+MeanDim+Norm 60.9 60.1 59.2 58.5 58.3 54.7 58.4 59.0 57.6 58.0 56.3 59.5
Table 2: Results (MAP%) of bottom-up sentence representation on the test question-to-question similarity task using the
Philosophy data set, 5 in-domain 300 dimensions embedding models (CBOW and Skipgram (Mikolov et al., 2013)), Glove
(Pennington et al., 2014), Character n-gram CBOW (ChC) and character Skip-gram (ChSG) (Bojanowski et al., 2016) and 7
pre-trained 300 dimensions models (W2V(sg) trained on googlenews (Mikolov et al., 2013), Glove6B trained on wikipedia and
Gigaword, Glove42B trained on Common Crawl (Pennington et al., 2014), Bow5C(CBOW), Bow5W(Sg) and Deps trained on
wikipedia (Levy and Goldberg, 2014) and Character skipgram (ChSG) trained on wikipedia (Bojanowski et al., 2016)).
ing step, we apply the L2 norm (+N orm) and Map (%) 70
our two tweaking techniques that is: +M eanAll
60
for a mean computation over the entire corpus and
+M eanDim, for a per dimension mean adjust- 50
ment. We also contrast the use of the L2 norm on
40
the top of our tweaking techniques that we refer
to as +M eanAll + N orm and +M eanDim + 30100 300 500 800
N orm. Dimension size (dim)
463
of embedding dimensions. Our results over in- Jeffrey Pennington, Richard Socher, and Christopher D
domain and pre-trained models showed significant Manning. 2014. Glove: Global vectors for word
representation. In Empirical Methods in Natural
improvements for question similarity task.
Language Processing (EMNLP). pages 1532–1543.
http://www.aclweb.org/anthology/D14-1162.
7 Acknowledgments
Richard Socher, Eric H. Huang, Jeffrey Pennington,
The current work was supported by the ANR 2016 Andrew Y. Ng, and Christopher D. Manning. 2011.
PASTEL (ANR-16-CE33-0007) project4 . Dynamic Pooling and Unfolding Recursive Autoen-
coders for Paraphrase Detection. In Advances in
Neural Information Processing Systems 24.
464
Meta-Embedding Sentence Representation for Textual Similarity
{Amir.Hazem,Nicolas.Hernandez}@univ-nantes.fr
According to Enkvist (1987): ”a model is a sim- Embedding models at the word level representa-
plified representation of reality. It is simplified be- tions have been widely explored in many appli-
cause it aims at reproducing a selection of rele- cations (Bengio et al., 2003; Collobert and We-
vant elements of reality rather than all of real- ston, 2008; Mikolov et al., 2013a; Pennington
ity at once.”. If several word embedding mod- et al., 2014; Bojanowski et al., 2016). Naturally,
els (Mikolov et al., 2013a; Pennington et al., they have been extended to sentence, paragraph
2014; Yin and Schütze, 2016; Arora et al., 2017) and document level representations (Socher et al.,
capture a selection of relevant features, differ- 2011; Mikolov et al., 2013a; Le and Mikolov,
ent embedding sets can cover different charac- 2014; Kalchbrenner et al., 2014; Kiros et al.,
teristics which can also be complementary (Yin 2015b; Wieting et al., 2016; Arora et al., 2017),
and Schütze, 2016). In order to capture a wide thanks to the continuous advances of deep neu-
range of features, it is useful to perform models ral embedding methods such as Recurrent Neu-
combination (ensemble models). The representa- ral Networks (RNN), Long Short Term Mem-
465
Proceedings of Recent Advances in Natural Language Processing, pages 465–473,
Varna, Bulgaria, Sep 2–4 2019.
ory (LSTM) and Convolutional Neural Networks ilarity tasks. A noticeable remark is that their
(CNN). Some sentence embedding representations approach outperformed sophisticated supervised
can be seen as a direct inspiration from word em- methods such as RNN’s and LSTM’s. Finally,
bedding models. For instance, while the skip- some approaches are supervised and need labelled
gram model (Mikolov et al., 2013a) predicts the data, such as DictRep (Hill et al., 2015) which
surrounding words given a source word, in the uses structured data to map dictionary definitions
same way, SkipThought (Kiros et al., 2015a) and of words with their pre-trained embeddings. With
FastSent (Hill et al., 2016) models predict sur- the encouraging results and simplicity of bottom-
rounding sentences given a source sentence. Also, up approaches, we focus in this paper on this type
the paragraph DBOW model (Le and Mikolov, of approaches and show their potential while used
2014) learns representations for variable length jointly with meta-embeddings.
pieces of texts and learns to predict the surround-
ing words based on contexts sampled from para- 3 Sentence Meta-Embedding
graphs. Recently, Pagliardini et al. (2018) in- Representation
troduced Sent2Vec, an approach based on word To deal with textual similarity, we propose a new
vectors along with n-gram embeddings simultane- approach that we refer to as meta-embedding sen-
ously to represent sentences. tence representation (MetaSentEmb). In the next
sections we first recall the principle of ensemble
Another type of sentence embedding represen-
approach from which we drawn our inspiration,
tation, also called bottom-up approach, represents
then we give the details of our approach.
sentences by a weighted sum of the embedding
vectors of their individual words. This naive ap- 3.1 Ensemble Approach
proach turned out to be competitive and outper- The principle of the ensemble approach is to com-
formed sophisticated approaches based on RNNs bine different models in order to catch the strength
and LSTMs in many natural language processing of each individual model. The main combina-
applications (Mikolov et al., 2013a; Wieting et al., tion techniques that have shown their effective-
2016; Arora et al., 2017; Hazem and Morin, 2017). ness are: vector addition (Garten et al., 2015) and
Mikolov et al. (2013a) for instance demonstrated vector concatenation (Garten et al., 2015; Yin and
the effectiveness of their model on the phrase anal- Schütze, 2016). For vector addition, given two
ogy task. They used the hierarchical softmax and embedding models, the procedure consists of ap-
subsampling using large amount of data. Wiet- plying a simple dimension-wise vector addition1 .
ing et al. (2016) have shown that a simple but su- For vector concatenation, given two embedding
pervised word averaging model of sentence em- models of dimensions dim1 and dim2, the result-
beddings leads to better performance on the para- ing concatenated embedding vector will be of size
phrase pairs data set (PPDB). However, the per- dim1 + dim2. The vectors have to be normalized
formance of their approach is closely related to before concatenation. Usually L2 norm is per-
the supervision from the date set, while without formed2 . Yin and Schütze (2016) performed a
supervision, their approach did not perform well weighted concatenation of 5 embedding models.
on textual similarity tasks. More recently, Arora They also experienced the SVD on top of weighted
et al. (2017) proposed a new sentence embedding concatenation vectors of dimension 950. This re-
method where they first compute a weighted av- sulted in a reduced model of 200 dimensions.
erage sum of the word embedding vectors of sen-
tences, and then, remove the projections of the av- 3.2 Proposed Approach
erage vectors on their first principal components. The bottom-up sentence embedding representa-
Like Mikolov et al. (2013a) and Wieting et al. tion consists of representing each given sentence
(2016), their approach is based on word embed- (or piece of text of any length) by an embed-
ding sum, but the difference is remarkable on the ding vector which is the sum of the vector embed-
weighted schema and on the use of principal com- ding of each word of the sentence (Mikolov et al.,
ponent analysis (PCA) method to remove the cor-
1
relation of sentence vectors dimensions. They This technique can not be applied when embeddings are
not of the same dimension size.
significantly achieved better performance than the 2
L2 norm can be performed either at dimension level (as
unweighted average on a variety of textual sim- suggested by Glove authors) or at vector length level.
466
2013b; Wieting et al., 2016; Arora et al., 2017). Research Paraphrase Corpus (MSRPC) used for
This representation is illustrated in the following paraphrase detection and (iv) the Ubuntu Dialogue
equation: Corpus used for Next Utterance Ranking (NUR).
n
X 4.1 Embedding Models
Senti = (Embedding(wj )) (1)
j=1 To study the impact of external data and con-
text representation, we chose different embedding
with Senti a given sentence i and n the number of
models. In addition to the word2vec model trained
words in Senti . Embedding(wj ) corresponds to
on Google News (Mikolov et al., 2013a), we
the embedding model used to represent each word
used the two Glove models respectively trained on
of the sentence Senti . We refer to this baseline ap-
Wikipedia+GigaWord (Glove6B) and on Com-
proach as SentEmb for sentence embedding rep-
mon Crawl (Glove42B) (Pennington et al., 2014).
resentation. A variant of this representation is the
We also used three Wikipedia pre-trained mod-
use of a weighted sum as presented in (Wieting
els (Levy and Goldberg, 2014), that is, two lin-
et al., 2016) for instance.
ear bag of word contexts and one dependency-
In this work, we extend the baseline representa-
based context. The bag of word models use a
tion (SentEmb) and propose M etaSentEmb, a
context size of 5 (Bow5C corresponds to CBow
meta-embedding sentence representation. We aim
and Bow5W to skipgram). The dependency-based
at improving sentence representation based on the
model used syntactic relations (Deps). Finally,
sum of its word embeddings. As we mainly oper-
we experienced the recent proposed character n-
ate at the word level representation, we study dif-
gram model (Bojanowski et al., 2016) by using the
ferent word meta-embedding techniques for sen-
character Skip-gram model trained on Wikipedia
tence representation. We basically use an ensem-
(ChSG). A summary of the pre-trained out-
ble approach to represent each word, which means
of-domain embedding sets is presented in Table
that each word has its own meta-embedding.
1. We also trained embedding models (CBOW,
Then, we sum each meta-embedding word of a
Skipgram, Glove and character n-gram models)
given sentence to obtain a meta-embedding sen-
on in-domain data sets (Qatar Living and Sick
tence representation (equation 2).
corpus of SemEval, MSPR for paraphrasing and
n
X Ubuntu for NUR). We respectively noted in do-
Senti = (Ensemble(wj )) (2) main trained embeddings as CBow, SkipGram,
j=1 Glove, CharSG and CharCBOW .
with Senti a given sentence i and n the number
4.2 Data Sets
of words in Senti . Ensemble(wj ) corresponds
to the ensemble technique used to represent word 4.2.1 Qatar Living Corpus
meta-embeddings. Ensemble(wj ) can be the ad- The Qatar Living corpus is a community ques-
ditive or the concatenation technique. We refer tion answering data set made of original and re-
to our proposed approach as M etaSentEmb for lated questions and their n corresponding answers.
sentence meta-embedding representation. Each The training and development data sets consist
sentence is pre-processed (Tokenization, part-of- of 317 original questions and 3,169 related ques-
speech tagging and lemmatization). Depending tions3 . The test sets of 2016 and 2017 respec-
on the targeted task, stop-words can be removed tively consist of 70 original/700 related questions
and part of speech filtering can be applied (keep- and 88 original/880 related questions. The Se-
ing only nouns, verbs and adjectives for instance). mEval (2016/2017) question-to-question similar-
ity shared task (Task3, SubtaskB) consists of iden-
4 Data and Tasks Description tifying for each original question, its correspond-
In this section, we briefly outline the differ- ing related questions over 10 candidates (Nakov
ent textual resources used for our experiments, et al., 2016, 2017). The question-to-question sim-
namely: (i) the Qatar Living corpus used in Se- ilarity task of SemEval offers an appropriate and
mEval 2016/2017 for question similarity task, (ii) interesting framework for evaluating our meta-
the Sick corpus used in SemEval 2014 for tex- 3
http://alt.qcri.org/semeval2016/
tual entailment and relatedness, (iii) the Microsoft task3/index.php?id=data-and-tools
467
Model Vocab Dim Training Data
word2vec 93k 300 Google News (Mikolov et al., 2013a)
Glove6B 400k 50-300 Wikipedia-Gigaword (Pennington et al., 2014)
Glove42B 1.9M 300 Common Crawl (Pennington et al., 2014)
Bow, Deps 175K 300 Wikipedia (Levy and Goldberg, 2014)
CharSG 175K 300 Wikipedia (Bojanowski et al., 2016)
embedding approach since an evaluation of mul- lation exists between two given sentences. By con-
tiple approaches including sentence embeddings trast to the question similarity and entailment pre-
have been already performed. diction tasks, sentence embedding similarity for
paraphrasing might not be appropriate for evalu-
4.2.2 Sick Corpus ation since the MSRP corpus includes many sen-
The sick data set consists of 10,000 English sen- tence pairs which are not paraphrases but contain
tence pairs annotated for relatedness in meaning many similar words. Sentence similarity based ap-
(a score form 1 to 5) and for entailment (Neutral, proaches should fail in this case to detect para-
Entailment or Contradiction). The SemEval 2014 phrases. However, showing the behaviour and the
shared task (Task1) consists of predicting whether performance of sentence similarity approaches in-
two given sentences are entailed, contradictions or cluding meta-embeddings on such a task may offer
neutral. Using sentence embeddings as well as some clues and may constitute a baseline for more
meta-embeddings for entailment prediction is an sophisticated approaches.
appropriate textual similarity task for evaluation,
however, dealing with contradictions and neutral
sentences is more difficult than a binary classifi- 4.2.4 Ubuntu Dialogue Corpus
cation which consists of predicting whether sen-
tences are entailed or not. In any case and for the The Ubuntu Dialogue Corpus is a large freely
sake of comparison, we perform the same evalua- available multi-turn dialogue data set (Lowe
tion as the state of the art approaches by keeping et al., 2015) constructed from the Ubuntu chat
the three classes (Neutral, Entailment or Contra- logs4 . The corpus (Human-Human chat) con-
diction) instead of two classes (Entailment or Not sists of approximately 930,000 two person di-
Entailment). As this work is mainly dedicated to alogues, 7,100,000 utterances5 and 100,000,000
the evaluation of sentence representations in sen- words. The task of NUR consists of retrieving the
tence similarity, we only focus on the entailment most probable utterance among a database of ex-
part and don’t consider the relatedness (we only isting human productions given a similar context.
report the results of the accuracy). This task offers a key challenge for sentence sim-
ilarity approaches since the relations between dia-
4.2.3 Microsoft Research Paraphrase Corpus
logue utterances are more generic. Here also, eval-
The Microsoft Research Paraphrase (MSRP) Cor- uating sentence similarity based approaches on a
pus (Dolan et al., 2004) is composed of 5,801 news different task, should give some insights about
paraphrase sentence pairs extracted from the web. their behaviour and to what extent it might help
Each sentence pair has been annotated by humans utterance prediction.
as being in paraphrase relationship (label=1) or not
(label=0). 67% of sentence pairs are positive ex-
amples (in paraphrase relationship) and 33% are 4
The first version can be found in http:
negative examples which make the corpus unbal- //irclogs.ubuntu.com/. A newer version has
anced. The corpus has been divided into 4,076 been recently released in https://github.com/
rkadlec/ubuntu-ranking-dataset-creator
training pairs and 1,725 test pairs. The paraphras- 5
All the replies and initial questions are referred to as ut-
ing task consists of identifying if a paraphrase re- terances
468
Input Tasks
Training data Models SemEval16 SemEval17 MSRP SICK NUR
Map (%) Map (%) Accuracy Accuracy 1 in 10 R@1
1*Wiki/GigaWord Glove6B 74.63 (8) 43.12 (5) 69.7 (5) 64.3 (2) 63.6 (4) (4)
1*Common Crawl Glove42B 74.93 (7) 41.68 (9) 66.9 (8) 61.5 (8) 59.4 (11) (10)
1*Google News word2vec 74.42 (9) 42.38 (8) 67.5 (7) 63.5 (5) 62.0 (7) (8)
4*wikipidia Bow5C 74.08 (10) 42.93 (6) 66.4 (9) 47.1 (12) 58.8 (12) (11)
Bow5W 75.64 (2) 45.69 (2) 70.1 (4) 63.9 (3) 62.6 (5) (3)
Deps 75.02 (6) 44.33 (4) 67.9 (6) 63.1 (7) 60.2 (8) (6)
CharSG 75.08 (5) 42.91 (7) 71.8 (2) 64.4 (1) 60.1 (9) (7)
5* In-domain Glove 73.04 (11) 41.57 (10) 64.9 (11) 54.4 (11) 62.3 (6) (12)
Cbow 72.78 (12) 40.12 (11) 65.1 (10) 60.1 (10) 66.1 (2) (12)
SkipGram 76.16 (1) 45.58 (3) 70.3 (3) 63.9 (3) 68.5 (1) (1)
CharCBOW 75.13 (4) 45.23 (4) 63.4 (12) 61.1 (9) 59.8 (10) (9)
CharSG 75.21 (3) 46.75 (1) 72.1 (1) 63.3 (6) 64.2 (3) (2)
Table 2: Results of SentEmb for five distinct textual relation detection tasks (question-to-question with
SemEval16 and SemEval17, paraphrase with MSRP, entailment with SICK and Next Utterance Ranking
with UDC) using different pre-trained out-of-domain and in-domain embedding models. The numbers
in brackets refer to the model rank in the given task, except for the last column which ranks the models
regardless of the task. The score of the three best models for each task are in bold.
5 Results and Discussion domain training data set. Surprisingly, the in-
domain CBow which is globally one of the two
We conducted two sets of experiments. The first worst models achieves the second best position
one aims at providing insights about the behaviour for the next utterance ranking task (NUR). This
of pre-trained and in-domain embeddings used in- can be explained by the large size of the in-
dividually. The second one aims at studying the domain Ubuntu data set while compared to other
contribution of ensemble models. in-domain data sets.
Table 2 shows that, regardless of the task,
the skipgram models (in-domain SkipGram, Table 3 reports the results of MetaSentEmb ap-
in-domain CharSG, out-of-domain Wikipedia proach for the four tasks using several pre-trained
Bow5W) outperform the other models. In addi- embedding combinations. Overall, we observe
tion, the two best models are in-domain and the that pre-trained embedding combination is use-
two following are out-of-domain. The fourth posi- ful in the majority of tasks (except the SICK
tion is hold by the Wikipedia/GigaWord Glove6B task where no significant improvements were ob-
model. Among the worst models, we observe two served). That said, not all the combinations are
CBOW models (in-domain Cbow out-of-domain efficient. It depends on the tasks and on the na-
Bow5C) as well as the in-domain Glove and the ture of the training data sets. For instance, in the
out-of-domain Glove42B models. Having said question-to-question similarity task (Semeval) the
that, even if the differences between the extrem- best meta-embedding models combination were
ities are notable, the coefficients of variability be- Glove6B with Glove42B (76.2% using addition
tween two successive ranked scores are often very and 76.4% using concatenation on 2016 edition)
low. A closer look at the results shows that the and CharSG with Glove42B (76.5% using addi-
out-of-domain and in-domain character skipgram tion and 76.2% using concatenation on 2016 edi-
models (CharSG) performed best for the para- tion), while for 2017 edition the best models were
phrase prediction task (MSRP). The entailment Glove6B with word2vec (47.3% using concatena-
detection task (SICK) is the only one for which tion) and Deps combined with Glove42B (47.4%
best models are largely out-of-domain. This can using addition). It is to note that Deps concate-
be explained by the very small size of the in- nated to Bow5W obtained similar results with
469
Tasks Pre-trained embedding models
Glove6B ChSG w2v Deps
Gl42 w2c B5C B5W Deps ChSG Gl42 w2v B5C B5W Deps Gl42 B5C B5W Deps Gl42 B5C B5W
SE16A 76.2 75.2 75.5 75.6 74.8 74.6 76.5 73.9 74.8 75.8 73.57 75.0 74.1 75.4 74.8 75.4 74.2 75.1
SE16C 76.4 76.2 75.6 76.1 74.4 75.0 76.2 74.8 74.7 75.8 74.7 75.6 74.6 76.1 74.2 75.3 74.8 75.3
SE17A 44.1 44.5 43.4 46.4 45.3 45.0 43.8 42.4 44.1 45.8 45.4 44.6 43.8 46.5 46.1 47.41 46.1 46.5
SE17C 45.9 47.3 45.2 46.1 45.8 45.3 45.0 43.7 44.4 46.3 46.1 45.3 43.4 46.1 45.7 45.8 45.4 47.1
M SP RA 69.2 68.4 68.6 68.1 70.0 68.1 72.6 68.9 68.0 69.1 70.3 70.7 69.2 67.8 71.5 70.6 69.1 70.8
M SP RC 69.4 67.9 68.6 67.0 70.4 68.0 72.7 69.2 69.1 68.0 71.5 71.5 69.2 68.0 71.0 70.3 69.2 70.3
SICKA 64.8 64.4 64.2 64.2 64.3 64.4 64.2 63.6 62.2 63.7 63.9 63.3 62.7 63.4 63.8 64.1 63.4 63.9
SICKC 64.3 64.3 64.0 63.9 64.3 64.4 64.3 63.5 62.4 63.6 64.1 63.5 63.1 63.5 63.9 64.1 63.5 63.9
N U RA 65.1 65.0 61.3 63.2 62.0 64.9 62.9 61.6 57.8 60.0 58.4 64.2 59.8 61.5 60.2 60.2 60.9 62.1
N U RC 65.6 65.3 61.1 63.6 62.1 65.1 64.4 61.5 57.8 60.2 58.5 64.4 59.6 62.0 60.3 62.2 61.1 62.2
Table 3: Results of the meta-embedding SentEmb, using addition (A ) and concatenation (C ) along with
pre-trained embeddings. SE16 and SE17 stand respectively for SemEval16 and SemEval17. Models
names were also digested to match the page setup: Glove42B (Gl42), word2vec (w2v), CharSG (ChSG),
Bow5C (B5C) and Bow5W (B5W).
Table 4: Meta-Embedding results using addition (A ) and concatenation (C ) along with in-domain em-
bedding models. SE16 and SE17 stand respectively for SemEval16 and SemEval17.
470
Data Tasks
Models SemEval16 SemEval17 MSRP SICK NUR
Map (%) Map (%) Accuracy Accuracy 1 in 10 R@1
Best@1 76.7 (1) 47.2 (3) 80.4 84.6 55.2
Best@2 76.0 46.9 77.4 83.6 48.8
Best@3 75.8 46.6 76.8 83.1 37.9
SentEmb 76.16 46.75 72.2 64.4 68.5
MetaSentEmbADD (In) 75.5 46.0 70.7 63.9 66.5
MetaSentEmbConcat (In) 74.1 44.0 69.4 63.1 66.7
MetaSentEmbADD (Out) 76.5 (2) 47.4 (1) 72.6 64.8 65.1
MetaSentEmbConcat (Out) 76.4 (3) 47.3 (2) 72.7 64.3 65.6
Table 5: Results obtained by the 3 best state of the art models proposed during the official competitions of
each task. Results obtained by the SentEmb baseline and with our proposed MetaSentEmb using addition
and concatenation techniques over pre-trained and in-domain embeddings as well as their combinations.
Bow5W (rank 3 overall the four tasks) was less Globally, except for the NUR task, the meta-
efficient while combined to other models, on embedding configurations using pre-trained mod-
the contrary, Glove42b which performed poorly els are slightly better than the ones using in-
individually (rank 10 overall the four tasks), domain models and enhance the performance of
turned out to be very efficient while combined to the SentEmb model. In particular, they outperform
other models. the SentEmb baseline and are ranked among the
To study the impact of in-domain embeddings, three best models for the SemEval tasks. Concern-
we report in Table 4 the results of MetaSentEmb ing the NUR task, the SentEmb baseline and all
while using embeddings trained on the in-domain the meta-embedding models outperform the three
data set of each task. According to the results, best state of the art models. For this specific task,
we observe the same tendency as for pre-trained the combination of in-domain models give bet-
embeddings. However, the improvements seem to ter results than out-of-domain models. Concern-
be task dependent. For instance, the best obtained ing the MSRP and the SICK tasks, while meta-
results were 76.5% for pre-trained versus 75.5% embedding models build on out-of-domain cor-
(Semeval 2016) and 47.4% for pre-trained versus pora achieve better results than in-domain models,
46.0% (Semeval 2017) while for NUR task, the none of them succeeded in beating the state of the
in-domain embedding obtained better results with art models.
66.5% versus 65.6% for the pre-trained models. If additional efforts are certainly needed to un-
That said for the NUR results, the different is not derstand the weak results on the entailment task
significant. Generally speaking, the results of Ta- and the different errors over all the evaluations,
bles 3 and 4 confirm the usefulness of using ex- our observations through an error analysis showed
ternal data in addition to various embedding mod- different findings, depending on the task of course
els and also put forward the possibility to combine but also on the proposed method itself which is
embeddings trained on both in-domain and exter- quite naive, especially for tasks like paraphras-
nal data sets. ing or entailment. First, MetaSentEmb performed
Table 5 reports the 3 best state of the art re- well on the question-to-question similarity task
sults obtained during the official competition of and was competitive with regards to the best Se-
each task. Also, it contrasts the SentEmb base- mEval systems. This is certainly due to the ade-
line with our proposed MetaSentEmb using addi- quacy of the task with our way of measuring sen-
tion and concatenation techniques over pre-trained tence similarity. The questions in the Qatar Liv-
(Out) and in-domain (In) embeddings as well as ing corpus contain few ambiguities and the main
their combinations. Below the header, the first errors were due to the specific forum vocabulary
horizontal frame reports the state of the art re- and mistakes that can be done by users. Also, one
sults. The second frame depicts results for simi- notable remark is the size of the original and re-
larity measures and the last frame contains results lated questions which is very important. Our way
of classification-based approaches. to deal with that was to filter stop-words and keep
471
only nouns, verbs and adjectives to limit the im- References
pact of long sentences. Using POS-tagging also Sanjeev Arora, Liang Yingyu, and Ma Tengyu. 2017.
provided tagging errors that introduced some er- A simple but tough to beat baseline for sentence
rors of our system. Second, MetaSentEmb did embeddings. In Proceedings of the 17th Inter-
not compete with the three best systems on the national Conference on Learning Representations
(ICLR’17). pages 1–11.
paraphrasing task (MSPR), however, if we com-
pare the results of MetaSentEmb with state of Yoshua Bengio, Rjean Ducharme, Pascal Vincent, and
art sentence embedding representations such as Christian Jauvin. 2003. A neural probabilistic lan-
FastSent (72.2%) and Skipthough (73.0%) (Hill guage model. JOURNAL OF MACHINE LEARN-
ING RESEARCH 3:1137–1155.
et al., 2016), our approach obtained similar results
(72.7%) with much simpler training. This finding Piotr Bojanowski, Edouard Grave, Armand Joulin, and
is encouraging while no particular attention was Tomas Mikolov. 2016. Enriching word vectors
given to the characteristics of paraphrase. Con- with subword information. CoRR abs/1607.04606.
http://arxiv.org/abs/1607.04606.
cerning textual entailment, MetaSentEmb failed to
improve the performance of SentEmb. Here also R. Collobert and J. Weston. 2008. A unified architec-
the particularities of the small data set as well as ture for natural language processing: Deep neural
the prediction of three classes including contra- networks with multitask learning. In International
Conference on Machine Learning, ICML.
diction and neutral sentences may explain the low
results. Finally, for the NUR task, our approach Bill Dolan, Chris Quirk, and Chris Brockett. 2004. Un-
turned out to be very efficient. Utterance char- supervised construction of large paraphrase corpora:
acteristics, at least for the Ubuntu corpus exhibit Exploiting massively parallel news sources. In Pro-
ceedings of the 20th International Conference on
strong similarities which are certainly better cap- Computational Linguistics. Association for Compu-
tured by our meta-embedding approach. tational Linguistics, Stroudsburg, PA, USA, COL-
ING ’04.
6 Conclusion
N Enkvist. 1987. Text linguistics for the applier: An
In this paper we introduced the first bottom-up orientation.. In U. Connor R. Kaplan, Writing
meta-embedding sentence representation for tex- across languages: Analysis of L2 Text (pp. 23-44),
Reading, MA: Addison-Wesley.
tual similarity. We have explored a variety of
pre-trained and in-domain embedding models and Justin Garten, Kenji Sagae, Volkan Ustun, and Morteza
there impact on question-to-question similarity, Dehghani. 2015. Combining distributed vector rep-
paraphrasing, textual entailment and next utter- resentations for words. In Proceedings of the 1st
Workshop on Vector Space Modeling for Natural
ance ranking tasks. We have also proposed meta- Language Processing. Denver, CO, USA, pages 95–
embedding sentence representations based on vec- 101. http://www.aclweb.org/anthology/W15-1513.
tors addition and concatenation and have shown
under which conditions they can be jointly used Amir Hazem and Emmanuel Morin. 2017. Bilin-
gual word embeddings for bilingual terminol-
for better performance. If further investigations ogy extraction from specialized comparable cor-
are needed, the preliminary results lend support pora. In Proceedings of the Eighth International
the idea that using meta-embeddings improve the Joint Conference on Natural Language Process-
performance of bottom-up sentence-based embed- ing (Volume 1: Long Papers). Asian Federation
of Natural Language Processing, pages 685–693.
ding approaches and offer an appropriate way to http://aclweb.org/anthology/I17-1069.
deal with textual similarity. One notable advan-
tage of our approach is its simplicity, especially Felix Hill, Kyunghyun Cho, and Anna Korhonen.
when using pre-trained embeddings since no com- 2016. Learning distributed representations of sen-
tences from unlabelled data. CoRR abs/1602.03483.
putational cost is incurred. http://arxiv.org/abs/1602.03483.
472
Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov, NAACL 2018 - Conference of the North American
Richard S Zemel, Antonio Torralba, Raquel Urta- Chapter of the Association for Computational Lin-
sun, and Sanja Fidler. 2015a. Skip-thought vectors. guistics.
arXiv preprint arXiv:1506.06726 .
Jeffrey Pennington, Richard Socher, and Christopher D
Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov, Manning. 2014. Glove: Global vectors for word
Richard Zemel, Raquel Urtasun, Antonio Torralba, representation. In Empirical Methods in Natural
and Sanja Fidler. 2015b. Skip-thought vectors. In Language Processing (EMNLP). pages 1532–1543.
C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, http://www.aclweb.org/anthology/D14-1162.
and R. Garnett, editors, Advances in Neural Infor-
mation Processing Systems 28, Curran Associates, Richard Socher, Eric H. Huang, Jeffrey Pennington,
Inc., pages 3294–3302. Andrew Y. Ng, and Christopher D. Manning. 2011.
Dynamic Pooling and Unfolding Recursive Autoen-
Quoc V. Le and Tomas Mikolov. 2014. Distributed coders for Paraphrase Detection. In Advances in
representations of sentences and documents. CoRR Neural Information Processing Systems 24.
abs/1405.4053.
John Wieting, Mohit Bansal, Kevin Gimpel, and Karen
Omer Levy and Yoav Goldberg. 2014. Dependency- Livescu. 2016. Towards universal paraphrastic sen-
based word embeddings. In Proceedings of the tence embeddings. Internationa Conference on
52nd Annual Meeting of the Association for Com- Learning Representations, CoRR abs/1511.08198.
putational Linguistics, ACL 2014, June 22-27, 2014,
Baltimore, MD, USA, Volume 2: Short Papers. pages Wenpeng Yin and Hinrich Schütze. 2016. Learning
302–308. http://aclweb.org/anthology/P/P14/P14- word meta-embeddings. In ACL (1). The Asso-
2050.pdf. ciation for Computer Linguistics. http://dblp.uni-
trier.de/db/conf/acl/acl2016-1.htmlYinS16.
Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle
Pineau. 2015. The ubuntu dialogue corpus: A large
dataset for research in unstructured multi-turn dia-
logue systems. arXiv preprint arXiv:1506.08909 .
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor-
rado, and Jeff Dean. 2013a. Distributed represen-
tations of words and phrases and their composition-
ality. In C. J. C. Burges, L. Bottou, M. Welling,
Z. Ghahramani, and K. Q. Weinberger, editors, Ad-
vances in Neural Information Processing Systems
26, pages 3111–3119.
Tomas Mikolov, Scott Wen-tau Yih, and Geoffrey
Zweig. 2013b. Linguistic regularities in continu-
ous space word representations. In Proceedings of
the 2013 Conference of the North American Chap-
ter of the Association for Computational Linguis-
tics: Human Language Technologies (NAACL-HLT-
2013). Association for Computational Linguistics.
Preslav Nakov, Doris Hoogeveen, Lluı́s Màrquez,
Alessandro Moschitti, Hamdy Mubarak, Timothy
Baldwin, and Karin Verspoor. 2017. SemEval-2017
task 3: Community question answering. In Proceed-
ings of the 11th International Workshop on Semantic
Evaluation. Association for Computational Linguis-
tics, Vancouver, Canada, SemEval ’17.
Preslav Nakov, Lluı́s Màrquez, Alessandro Moschitti,
Walid Magdy, Hamdy Mubarak, Abed Alhakim
Freihat, Jim Glass, and Bilal Randeree. 2016.
SemEval-2016 task 3: Community question answer-
ing. In Proceedings of the 10th International Work-
shop on Semantic Evaluation. Association for Com-
putational Linguistics, San Diego, California, Se-
mEval ’16.
Matteo Pagliardini, Prakhar Gupta, and Martin Jaggi.
2018. Unsupervised Learning of Sentence Embed-
dings using Compositional n-Gram Features. In
473
Emoji Powered Capsule Network to Detect Type and Target of Offensive
Posts in Social Media
474
Proceedings of Recent Advances in Natural Language Processing, pages 474–480,
Varna, Bulgaria, Sep 2–4 2019.
region of the input embeddings. However recent – Offensive (OFF): Posts containing any
introduction of capsule networks shows that while form of non-acceptable language. These
pooling works better in most of the scenarios, it posts can include insults, threats swear
nonetheless is losing valuable information (Hinton words etc. (Zampieri et al., 2019)
et al., 2018). The solution that has been brought
forward is Capsule Networks. How it overcomes • Subtask B: Categorization of Offensive Lan-
the weaknesses in max pooling layer will be dis- guage : Task’s goal was to categorize the type
cussed in Section 3. of offense.
Since the Capsule Networks are very new to the – Targeted Insult (TIN) : Posts that con-
field, they have not been used much in NLP tasks. tain targeted insults and threats.
However, their good performance in image classi- – Untargeted (UNT) : Posts containing
fication tasks motivated us to use them in offen- non targeted insults or threats.
sive language detection tasks too. To the best of
our knowledge, no prior work has been explored • Subtask C: Offensive Language Target Iden-
in offensive language identification with Capsule tification : Goal of the task was to categorize
Networks. Also, it might be important to explore the targets of insults/threats.
how the Capsule Networks performs in NLP do- – Individual (IND) : Insults that target in-
main. Additionally we analyzed that most of the dividuals.
social media posts contain not only text but emo-
– Group (GRP) : Insults that target a
jis too, which can be a contributing factor for of-
group of people.
fense. Therefore we propose a method to incor-
porate emoji knowledge to the Capsule Network – Other (OTH) : The target does not be-
architecture. Generally, this paper proposes a Cap- long to any category mentioned above.
sule Network architecture with emoji information Few examples from the training set is shown in
to detect offensive posts in social media. The rest Table 1.
of the paper is organised as follow. Section 2 As you can see, the nature of the three tasks are
would briefly describe the tasks and the dataset. different and it would interesting to explore how
Section 3 would describe the capsule network ar- one architecture can be used to capitalise all of the
chitecture we used and how we integrated emoji three tasks.
information to the architecture. After that we eval-
uate the system comparing with the architectures 3 Research Approach
provided in Zampieri et al. (2019). Finally, the
We first describe the existing approaches men-
conclusions are presented.
tioned in Zampieri et al. (2019). Then we will de-
2 Dataset and Task Description scribe the proposed capsule network architecture.
Dataset that we used was released for the Task 6 in 3.1 Existing Approaches
SemEval-2019 : OffensEval: Identifying and Cat- There are three approaches considered in Zampieri
egorizing Offensive Language in Social Media. It et al. (2019) which will be described in the fol-
has been collected from Twitter using its API and lowing list. We describe them briefly in this sub
searching for keywords and constructions that are section before introducing the capsule network ar-
often included in offensive messages, such as she chitecture.
is, to:BreitBartNews, gun control etc (Zampieri
et al., 2019). 1. SVM - A linear SVM trained on word un-
There were three tasks associated with the igrams. SVMs have achieved state-of-the-
shared task. art results for many text classification tasks
(Zampieri et al., 2018).
• Subtask A: Offensive language Detection :
2. BiLSTM - The model consists of (i) an input
Goal of the task was to discriminate between
embedding layer,(ii) a bidirectional LSTM
the following types of tweets:
layer (Schuster and Paliwal, 1997), and (iii)
– Not Offensive (NOT) : Posts that do not an average pooling layer of input features.
contain offense The concatenation of the LSTM layer and
475
id tweet a b c
44209 @USER @USER what a baby! URL NOT NULL NULL
97670 @USER Liberals are all Kookoo !!! OFF TIN OTH
74831 @USER Trump kicks dem butt - its so fun. OFF TIN IND
IM FREEEEE!!!! WORST EXPERIENCE
17259 OFF UNT NULL
OF MY FUCKING LIFE
the average pooling layer is further passed words that are fed will be out-of-vocabulary
through a dense layer, whose output is ulti- words. Therefore we used character embed-
mately passed through a softmax to produce dings (Mikolov et al., 2018) as it provides
the final prediction. The model is adapted embeddings for misspelling words and new
from a pre-existing model for sentiment anal- words. Also character embeddings handle in-
ysis (Rasooli et al., 2017). frequent words better than word2vec embed-
ding as later one suffers from lack of enough
3. CNN - A convolutional neural network based training opportunity for those rare words.
on the model proposed in Kim (2014). It con- We used fasttext embeddings pre trained on
sists of an (i) an input embedding layer, (ii) Common Crawl (Mikolov et al., 2018). Us-
a convolutional layer (Collobert et al., 2011) ing the model we represented each word as
and (iii) a max pooling layer (Collobert et al., a vector with a size of 300 values. The em-
2011) of input features. The output of the bedding layer is improved more with emoji
max pooling layer is further passed through a information, which will be described in Sec-
dropout (Srivastava et al., 2014) and softmax tion 3.3.
output.
2. Feature Extraction Layer - We used this
Both BiLSTM and CNN architectures above layer to extract long term temporal dependen-
have pooling layers which is a very primitive type cies within the text. We experimented both
of routing mechanism. The most active features LSTMs (Hochreiter and Schmidhuber, 1997)
in a local pool is routed to the higher layer and and GRUs (Chung et al., 2014) for this layer.
the higher-level detectors don’t have an impact in Due to the fact that we had a small number
the routing. However in the Capsule Network, of training examples, GRUs performed bet-
only those features that agree with high-level de- ter than LSTMs, capitalising on GRU’s abil-
tectors are routed. It has a superior dynamic rout- ity to exhibit better performance on smaller
ing mechanism. With this advantage we propose a datasets. For the final architecture we used a
novel capsule network architecture for aggression bi directional GRU layer with 50 time steps,
detection which will be described in the next sec- each getting initialised with glorot normal
tion. initialiser.
3.2 Proposed Architecture 3. Capsule Layer - The Capsule layer we used
The proposed architecture is depicted in Figure 1. is primarily composed of two sub-layers Pri-
The architecture consists of four layers. mary Capsule Layer and Convolutional Cap-
sule Layer.
1. Embedding Layer - We represent every text
xi , as a sequence of one-hot encoding of its (a) Primary Capsule Layer - The primary
words, xi = (w1 , w2 , ...wn ) of length n, which capsule layer is supposed to capture the
is the maximum length of the all of the texts instantiated parameters of the inputs, for
in the training set, with zero padding. Such example, in case of texts, local order
a sequence becomes the input to the embed- of words and their semantic representa-
ding layer. Most of the words exist in social tion is captured with the primary capsule
media texts are not proper words. If we used layer.
word embeddings to initialize the embedding (b) Convolutional Capsule Layer - The con-
matrix in the embedding layer, most of the volutional capsule layer outputs a lo-
476
Figure 1: Capsule Network
cal grid of vectors to capsules in ear- Fiser, 2016). A research conducted by (Barbi-
lier layers using different transforma- eri et al., 2017) has been showed that there is an
tion matrices for each capsule. This is unique and important relation between sequences
trained using dynamic routing algorithm of words and emojis. When analyze the top 10
described in Sabour et al. (2017) that emojis belong to both categories; Offensive and
overlooks words that are not important Not Offensive, in the selected dataset, it also
or unrelated in the text, like stopwords shows a clear distinction of emojis corresponding
and name mentions which are common to its category as shown in Figure 2. Due to the
in social media texts. extensive usage of emojis in social media and the
relationship lie between emojis and text, integra-
4. Dense Layers - Output of the Capsule layer tion of emojis can be used to improve the social
is flattened and then fed in to two dense lay- media offensive language detection.
ers. First dense layer had 100 units and was Since the proposed architecture is based on em-
activated with relu function. After apply- beddings, we decided to integrate emojis also us-
ing batch normalization to the output of first ing the embeddings. But most of the available
dense layer, it was fed in to the second dense pre-trained word embedding sets include few or
layer with 1 unit and and sigmoid activation. no emoji representations. Therefore in addition
to the character embeddings, separate embedding
Apart from the major sections in the architecture set; emoji2vec (Eisner et al., 2016) was chosen
described above, we used a spatial dropout (Tomp- for emojis. Emoji2vec consists of pre-trained em-
son et al., 2015) between the embedding layer and beddings for all Unicode emojis using their de-
the feature extraction layer and a dropout (Srivas- scriptions in the Unicode emoji standard. This
tava et al., 2014) between the two dense layers to maps emojis into 300-dimensional space similar
minimize over fitting of the network. The imple- to other available word embeddings; word2vec,
mentation was done using Keras (Chollet et al., glove, etc. to make the integration easy with
2015) and Python1 . word vectors. Emoji2vec embeddings were eval-
The next section would describe how we inte- uated based on sentiment analysis on tweets and
grated emoji knowledge to this architecture. it showed word2vec with emoji embeddings ad-
vances the classification accuracy while proving
3.3 Integrating Emojis that the emoji2vec embeddings are useful in social
Emojis are ideograms which are used with text natural language processing tasks.
to visually complement its meaning. In present, Following (Eisner et al., 2016), there were two
emojis are widely used by social media. A global pre-trained emoji2vec models2 . One model is
analysis done on Twitter has been found that based on the sum of vectors corresponds to the
19.6% of tweets contain emojis. Further it stated, words found in phrases which describe the emojis.
emojis are used by 37.6% of users (Ljubesic and As an extended version of it, other model feeds the
actual word embeddings to an LSTM layer. We
1
The code is available on ”https://github.com/
2
TharinduDR/Aggression-Identification” https://github.com/uclmr/emoji2vec
477
Figure 2: Top 10 emojis belong to Not Offensive (NOT) and Offensive (OFF) posts, Task 6 Dataset,
SemEval-2019
Table 2: Results for offensive language detection. We report Precision (P), Recall (R), and F1 for each
model/baseline on all classes (NOT, OFF), and weighted averages. Macro-F1 is also listed (best in bold).
† denotes the capsule net architecture integrated with emoji embeddings.
used 300 dimensional embedding spaces gener- learning rate once learning stagnates. Also the pa-
ated by both models; sum based and LSTM based rameters of the network was optimized using five
for this experiment. fold cross validation.
Two approaches were used to integrate emoji
embeddings with the above mentioned architec- 4 Evaluation
ture as follows:
The capsule network architecture we proposed
1. 300 dimensional embedding layer - Shared above was evaluated using the testing set provided
same 300 dimensional vector space for both for each of the subtask in SemEval-2019 Task 6:
word and emoji embeddings. OffensEval: Identifying and Categorizing Offen-
sive Language in Social Media.
2. 600 dimensional embedding layer - Used
concatenation layer and resulted 600 dimen-
4.1 Offensive Language Detection
sional vector space by both word and emoji
embeddings. The performance on identifying offensive (OFF)
and non-offensive (NOT) posts is reported in Table
Among the experiments we conducted us- 2. The Capsule Network we proposed outperforms
ing both emoji2vec models and integration ap- the RNN model, achieving a macro-F1 score of
proaches, combination of sum based emoji em- 0.81.
beddings with 600 dimensional embedding layer
and LSTM based emoji embeddings with 300 4.2 Categorization of Offensive Language
dimensional embedding layer resulted improve-
ments compared to word embeddings only ap- The results for the offensive language categoriza-
proaches. More details on experiment results are tion is shown in Table 3. In this subtask too
mentioned in Section 4. Capsule Network architecture outperforms all the
other models having a macro F1 score of 0.71.
3.4 Training
4.3 Offensive Language Target Identification
The network was trained on the training dataset
provided for SemEval-2019 Task 6: OffensEval: The results for the offensive language target iden-
Identifying and Categorizing Offensive Language tification is shown in Table 4. Capsule Network
in Social Media. It was trained using adam op- architecture outperforms all the other models hav-
timiser (Kingma and Ba, 2015), with a reduced ing a macro F1 score of 0.49.
478
TIN UNT Weighted Average
Model
P R F1 P R F1 P R F1 F1 Macro
SVM 0.91 0.99 0.95 0.67 0.22 0.33 0.88 0.90 0.88 0.64
BiLSTM 0.95 0.83 0.88 0.32 0.63 0.42 0.88 0.81 0.83 0.66
CNN 0.94 0.90 0.92 0.32 0.63 0.42 0.88 0.86 0.87 0.69
CapsuleNet † 0.96 0.94 0.94 0.33 0.67 0.44 0.90 0.88 0.87 0.71
All TIN 0.89 1.00 0.94 - 0.00 0.00 0.79 0.89 0.83 0.47
All UNT - 0.00 0.00 .11 1.00 0.20 .01 0.11 0.02 0.10
Table 3: Results for offensive language categorization. We report Precision (P), Recall (R), and F1 for
each model/baseline on all classes (TIN, UNT), and weighted averages. Macro-F1 is also listed (best in
bold). † denotes the capsule net architecture integrated with emoji embeddings.
.
Table 4: Results for offense target identification. We report Precision (P), Recall (R), and F1 for each
model/baseline on all classes (GRP, IND, OTH), and weighted averages. Macro-F1 is also listed (best in
bold). † denotes the capsule net architecture integrated with emoji embeddings.
As shown in Table 2, 3 and 4 capsule network Language in Social Media. Importantly our sys-
architecture outperformed all the other models in tem does not rely on language dependent features
all sub tasks. It is worth noticing that the un- so that it is portable for any other language too.
balanced nature of the dataset did not affect the The main conclusion of the paper is that even
performance of the capsule network architecture. though the capsule networks are not widely used
Also eventhough the capsule layer is seemingly in NLP domain, they can achieve state of the art
complex, results show that it does not need a large results. Also with the shown way of integrating
training set to optimize its parameters. emoji information to the network, results can im-
We did not fine tune the model analyzing data prove.
in this dataset since we wanted a general model In the future we hope to implement a multi pur-
capable of identifying offense. Hence, we did not pose capsule network architecture for several tasks
compare our results with the final results of the in NLP domain such as spam detection, gender
shared task. identification etc. We hope to further explore cap-
sule network architectures in various NLP tasks.
5 Conclusion
We have presented a novel capsule network archi- References
tecture to detect type and target of offensive posts
Francesco Barbieri, Miguel Ballesteros, and Horacio
in social media. Also we propose a method to Saggion. 2017. Are emojis predictable? Arxiv .
incorporate emoji knowledge to the architecture.
Our approach was able to improve on the baseline François Chollet et al. 2015. Keras. https://
keras.io.
system presented at SemEval-2019 Task 6: Of-
fensEval: Identifying and Categorizing Offensive Junyoung Chung, aglar Gülehre, Kyunghyun Cho, and
479
Yoshua Bengio. 2014. Empirical evaluation of gated Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann
recurrent neural networks on sequence modeling. LeCun, and Christoph Bregler. 2015. Efficient ob-
CoRR abs/1412.3555. ject localization using convolutional networks. 2015
IEEE Conference on Computer Vision and Pattern
Ronan Collobert, Jason Weston, Léon Bottou, Michael Recognition (CVPR) pages 648–656.
Karlen, Koray Kavukcuoglu, and Pavel P. Kuksa.
2011. Natural language processing (almost) from Marcos Zampieri, Shervin Malmasi, Preslav Nakov,
scratch. Journal of Machine Learning Research Ahmed Ali, Suwon Shon, James Glass, Yves Scher-
12:2493–2537. rer, Tanja Samardzic, Nikola Ljubesic, Jörg Tiede-
mann, Chris van der Lee, Stefan Grondelaers,
Ben Eisner, Tim Rocktäschel, Isabelle Augenstein, Nelleke Oostdijk, Dirk Speelman, Antal van den
Matko Bosnjak, and Sebastian Riedel. 2016. Bosch, Ritesh Kumar, Bornini Lahiri, and Mayank
emoji2vec: Learning emoji representations from Jain. 2018. Language identification and morphosyn-
their descriptions. Arxiv . tactic tagging: The second vardial evaluation cam-
paign. In VarDial@COLING 2018.
Thiago Galery and Efstathios Charitos. 2018. Aggres-
sion identification and multi lingual word embed- Marcos Zampieri, Shervin Malmasi, Preslav Nakov,
dings. In TRAC@COLING 2018. Sara Rosenthal, Noura Farra, and Ritesh Kumar.
2019. Predicting the type and target of offensive
Geoffrey E. Hinton, Sara Sabour, and Nicholas Frosst. posts in social media. In Proceedings of NAACL.
2018. Matrix capsules with em routing. In ICLR.
480
EoANN: Lexical Semantic Relation Classification Using an
Ensemble of Artificial Neural Networks
Abstract 1 Introduction
Researchers use wordnets as a knowledge Lexical semantic relation classification is the task
base in many natural language processing of identifying s semantic relation(s) which holds
tasks and applications, such as question between word pairs among a set of predefined
answering, textual entailment, discourse relation types. Relation classification can be done
classification, and so forth. Lexical in a supervised manner, using a dataset, labeled
semantic relations among words or
with a certain number of relation classes. In
concepts are important parts of knowledge
encoded in wordnets. As the use of
addition to classification with known relations,
wordnets becomes extensively widespread, there are some methods which go even further and
extending the existing ones gets more learn new semantic relations and suggest new
attention. Manual construction and relation categories (Shamsfard and Barforoosh,
extension of lexical semantic relations for 2003).
WordNets or knowledge graphs are very Relation identification plays an essential role in
time consuming. Using automatic relation many natural language processing application such
extraction methods can speedup this as question answering, recognizing textual
process. entailment and discourse understanding.
There are two main approaches for classification
In this study, we exploit an ensemble of
LSTM and convolutional neural networks
of lexical semantic relations; distributional and
in a supervised manner to capture lexical path-based (Wang et al., 2017).
semantic relations which can either be used Path-based approaches try to recognize the type
directly in NLP applications or compose of semantic relation between word pairs according
the edges of wordnets. The whole to their co-occurrence information in the corpus.
procedure of learning vector space These methods mainly use the dependency path
representation of relations is language between word pairs as their input feature (Snow et
independent. We used Princeton WordNet al., 2004; Riedel et al., 2013). As Ziph's law states
3.1, and FarsNet 3.0 (the Persian that most of the words in vocabulary rarely occur
wordnet), as gold standards to evaluate the
in the corpus (Powers, 1998) these methods have
predictive performance of our model and
the results are comparable on the two
some limitation for word pairs who do not co-occur
languages. Empirical results demonstrate in a context.
that our model outperforms the state-of- On the other hand according to the distributional
the-art models. hypothesis which states "words that occur in
similar contexts tend to have similar meanings"
(Harris, 1954), distributional approaches try to
recognize the relation between words based on
their separate occurrence in the corpus which can
481
Proceedings of Recent Advances in Natural
1 Language Processing, pages 481–486,
Varna, Bulgaria, Sep 2–4 2019.
be represented for example by their word According to human expert reviews, our
embedding vectors (Mikolov et al.,2013) and these model goes beyond relation discovery
methods have shown great performance (Baroni et and can be employed to correct the
al., 2012; Turney and Pantel, 2010; Roller et al., potential error in wordnet edges and
2014). suggest new missed relation instances.
In the last decades, several researches have been
conducted on discovering hypernymy as an The rest of this paper is structured in 6
example of lexical semantic relations, and a key sections:
part of taxonomies and state-of-the-art models Section 2 presents the existing approaches for the
show significant results (Shwarts et al., 2016; classification of lexical semantic relations; the next
Roller et al., 2016). one presents our model in detail, section 4
However, other types of relations have been less describes the data set we used for evaluating our
investigated. model, section 5 reports experimental results and
Several types of models have been used for the finally section 6 dedicated to the conclusion and
task of semantic relation classification, but the future works.
results are not sufficiently admissible (Vu and
Shwarts, 2018). 2 Related Work
In this paper, we use an ensemble of models to There are two main lexical semantic relation
improve prediction performance of relation extraction models, distributional and path-based
classification. The idea of ensemble methodology (pattern-based) (Wang et al., 2017) and also there
is to combine some weighted classifiers in order to are methods that use an integration of these two
obtain a more accurate one (Rokach et al. 2009). approaches (Shwartz et al., 2016).
Main building blocks of this combinational Distributional methods learn the relation
model are some inducers named weak learner between word pairs based on the disjoint
which perform slightly better than random. occurrence of them. These methods usually use a
According to Condorcet Jury theorem which states combination of word embedding vectors (Mikolov
" the ensemble of independent voters each of which et al., 2016) as their input features. Considering v1
performs better than random (p>0.5) has a and v2 being word embedding vector
probability of L>p to make the right decision." we corresponding to w1 and w2, most common
used different structured neural networks as the combinations are:
model's weak learners. concatenation of v1 and v2 (Concat)
The input of our model is a concatenation of
word embedding vectors corresponding to target the offset of v1 and v2 (Offset)
word pairs, and the output is a class label predicted
based on learned vector space distributional point-wise multiplication of v1 and v2
representation of the semantic relation which holds (Mult)
between them.
Our final model has the best validation F1 score squared difference between v1 and v2
of 0.894 in predicting the relation between FarsNet (Sqdiff)
(Shamsfard, et al., 2010) word pairs and 0.768 to Offset (Roller et al., 2014; Weeds et al., 2014;
predict Princeton Wordnet (Miller, 1995; Felbaum Fu et al., 2014), Concat (Baroni et al., 2012) and
1998) relation classes. Concat+Offset (Washio and Kato, 2018) is the
We summarize the contribution of this paper as most common type of feature vector combination
follow: which is used in this task. To capture the different
We propose EoANN, an ensemble of notion of interaction information about relation Vu
artificial neural networks for classifying and Shwartz (2018) add Mult, studied by Weeds et
all types of lexical semantic relations in al. (2014) and Sqdiff introduced by themselves as
target datasets, without any hand-crafted input feature and report Mult+Concat performs
features.
better than other combination.
Our model addresses the sparseness issue These methods mostly focus on lexical
and can classify word pairs which do not entailment and relation classes such as hypernym,
necessarily co-occur in the corpus. causality and other instances of relation which
482
2
exemplified inference and have a state-of-the-art stacking which involves training a learning
F1 score of 0.91. algorithm to combine the predictions of several
Path-based or pattern-based methods utilize learning algorithms.
features derived from the context in which word The advantage of stacking is to increase the
pairs co-occur. For example, the dependency path prediction power of the classifier. As the using of
between a word pair and observed predefined another neural network above the weak learners in
patterns are used as an informative feature to order to learn the final prediction imposes excess
classify the relation. The methods of this category overhead, we use the simplest stacking method
are limited to use only the word pairs that co-occur which is averaging. Averaging has no parameter, so
in corpus (Hearst et al., 1992; Snow et al., 2004; no training is needed.
Navigli and Velardi 2010; Shamsfard et al,. 2010; We transfer the input embedding vector of
Boella and Di Caro, 2013; Pavlick and Pasca, word pairs to dense-valued feature vectors, next
2017) feed these vectors to both ANN to compose their
Recently some approaches use an integration own distributional representation of them. At the
of these two methods and combine both final layer of each, a softmax classifier predicts the
distributional and dependency path information to label of input sample.
obtain better results. HypNet (shwarts et al., 2016) Finally, a weighted averaging mechanism is
is an examples of these approaches. used to decide the relation class in which input
words participate.
3 Our Model
3.1 Input of EoANN
In this paper, we propose a model to classify lexical
Our inputs are raw lexical entries (multi word
semantic relations between a word pair using their
expressions are excluded) of Wordnets. We first
word embedding vectors.
transform every single word to its embedding
The rarity of co-occurring every candidate
vectors using word embedding.
word pair which possibly involves in a semantic
Word embedding is a method to map words
relation leads us to exploit a method which does
and phrases from space with one dimension per
not necessarily need to see the word pair in a
word, to a continuous low dimensional vector
context together.
space. There are many word embedding
The output of our model is a class label
frameworks. We use Fasttext (Piotr et al., 2017)
prediction based on learned vector space
which represents words as the sum of the n-gram
distributional representation of the semantic
vectors. This method is actually an extension of the
relation which holds between target word pairs.
continuous skip-gram model (Mikolov et al.,
Although using a single deep neural network
2013), which considers sub-word information as
(as a distributional method) showed some
well. We denote the word embedding vector of
improvement in capturing semantic relations, in
word w by vw ∈ ℝ
order to get the advantage of the diversity among
Given R (a, b) as a sample of semantic relation
predictions of separately trained models, we use an
triple in target Wordnet, R is the class of relation
ensemble of two artificial neural networks.
which connects a to b and va and vb are the
The ensemble is a general statistical
embedding vectors corresponding to them. The
enhancing technique to improve the
input vector and labels of our classifiers is the
representational capacity of the model. This
concatenation of word vectors:
enhancement helps to find a hypothesis which is
ℎ1(𝑎, 𝑏) = [𝑣𝑎: 𝑣𝑏]
independent of the space of the model from which
𝑜𝑢𝑡(𝑎, 𝑏) = 𝑅
it starts to learn.
First, we train two neural networks separately
on data our labeled data and evaluate their test 3.2 Weak Learners Structure
results, then put these two models in an ensemble
and re-evaluate the result. Comparing two result We use both convolutional neural network and
sets shows 0.1 improvement in F1 score of learned LSTM network in the simplest structure as our
hypothesis. model base inducers. These two learners are
Models can be assembled in many different chosen because of their power in capturing of
ways like boosting, bagging and stacking. We use
483
3
hierarchical patterns and the extraction of the WordNet Hypernym, Hyponym, 634,330
temporal behavior. Entailment, Cause,
The simple CNN inducer is composed of 3 main Instance-Hypernym,
layers, a convolutional layer with 20 filters of size Instance-Hyponymy,
(1, 2), a pooling layer which is used to reduce the Member, Holonym,
Attribute
dimensions of feature map and finally a fully
connected layer that flattens the results and passes
Farsnet Hypernym, Hyponym, 322,554
it to a softmax classifier to decide which relation
Antonym, Instrument,
class the input belongs to. Domain, Instance-
LSTM neural network which we use as another Hypernym,
weak learner is composed of a fully connected Instance-Hyponym
layer to encode 2-dimentional input feature vector Location, Patient
to a dense flat vector, then passes its output to a ROOT09 Hypernym, co- 12,762
LSTM layer with 200 memory units and a softmax Hyponym, Random
classifier finally decides about data class label. EVALution Hypernym, Antonym, 7,378
Combiner is responsible for getting the final meronym, possession,
decision by combining individual classifiers Attribute, Part Of
predictions. This component holds a majority
Table 1: data sets we use for evaluating our model,
voting among classifiers and declares the ultimate
their main relation categories and the number of
predicted label.
relation instances of each
The rest of the parameters are as Fasttext default Model Data Set Classifier F1
configuration. feature
data set relation classes # of composition
instances EoANN Root09 LSTM+CNN 0.868
Concat
484
4
EVALution LSTM+CNN 0.655 References
Concat
LexNet Root09 RBF 0.814 George A. Miller. 1995. WordNet: a lexical database
Sum+SqDiff for English. Communications of the ACM,
EVALution RBF 0.6 38(11):39-41.
Concat+Mult Christiane Fellbaum 1998. WordNet: An Electronic
KSIM Root09 RBF 0.723 Lexical Database. Cambridge, MA: MIT Press.
Sum+SqDiff Mehrnoush Shamsfard, Akbar Hesabi, Hakimeh
EVALution RBF 0.505 Fadaei, Niloofar Mansoory, Ali Famian, Somayeh
Concat+Mult Bagherbeigi, Elham Fekri, Maliheh Monshizadeh
and Mostafa Assi. 2010. Semi-automatic
Table 2: best precision recall and F1 score for root09 development of farsnet; the Persian
and EVALution in 4 compared models wordnet. Proceedings of 5th global WordNet
conference, Mumbai, India; 29.
David Austen-Smith, Jeffrey S. Banks. 1996.
model dataset Model and Features F1 Information aggregation rationality and the
score Condorcet jury theorem, American Political
EoANN farsnet LSTM+CNN 0.894 Science Review, vol. 90, pages. 34-45.
Concat
Mehrnoush Shamsfard and AA Barforoush. 2003. An
Semi- farsnet Pattern-based 0.605 introduction to HASTI: an ontology learning
auto +structured_based+ system. Proceedings of the iasted international
statistical conference artificial intelligence and soft
computing, Acta Press, Galgary, Canada: 242-247.
Table 3: best F1 score on farsnet Rion Snow, Daniel Jurafsky, and Andrew Y Ng. 2004.
Learning syntactic patterns for automatic hypernym
discovery. In NIPS.
Ronan Collobert, Jason Weston, Leon Bottou, Michael
6 Conclusion and Future Works Karle, Koray Kavukcuoglu, Pavel Kuks. 2011.
Natural language processing (almost) from scratch.
Our objective in this research was to automatically Journal of machine learning research, pages 2493-
classify lexical semantic relation employing the 537
power of the simple but effective structured neural
Sebastian Riedel, Limin Yao, Andrew McCallum, and
networks, which have shown their proficiency in M. Benjamin Marlin. 2013. Relation extraction with
many tasks of natural language processing matrix factorization and universal schemas. In
(Collobert et al., 2011; Yao et al., 2013). Proceedings of NAACL-HLT 2013, pages 74–84.
We used both LSTM and convolutional
David M. Powers 1998. Applications and explanations
network to benefit the exhibition of temporal of Zipf's law. Association for Computational
behavior by first and the extraction of the Linguistics: 151–160.
hierarchical pattern by last.
Zellig S. Harris. 1954. Distributional
We also used the simplest distributional
Structure. WORD, 10(2-3):146-162.
feature as input and entrusted the extraction of the
most proper composition of features to the model. Tomas Mikolov, Ilya Sutskever, Kai Chen, Gregory S
In case of ROOT09 and EVALution our model Corrado, and Jeffrey Dean. 2013. Distributed
representations of words and phrases and their
has an improvement of 0.05 in F1 score from state compositionality. In NIPS, pages 3111–3119.
of the art (LexNet). And for FarsNet dataset we
have 0.11 improvement in F1 score. Marco Baroni, Raffaella Bernardi, Ngoc-Quynh Do,
The next step in extending lexical ontologies is and Chungchieh Shan. 2012. Entailment above the
word level in distributional semantics. In
to complete missed relation edges, then to learn Proceedings of EACL 2012, pages 23–32.
new relation classes, which can be added to the
target wordnet. Peter D Turney and Patrick Pantel. 2010. From
frequency to meaning: Vector space models of
semantics. Journal of artificial intelligence research,
37:141–188.
485
5
Julie Weeds, Daoud Clarke, Jeremy Reffin, David Guido Boella and Luigi Di Caro. 2013. Supervised
Weir, and Bill Keller. 2014. Learning to distinguish learning of syntactic contexts for uncovering
hypernyms and co-hyponyms. In Proceedings of definitions and extracting hypernym relations in
COL- ING 2014, the 25th International Conference text databases. In Machine learning and knowledge
on Computational Linguistics: Technical Papers, discovery in databases, pages 64–79. Springer.
pages 2249–2259, Dublin, Ireland. Dublin City
Ellie Pavlick and Marius Pasca 2017. Identifying
University and Association for Computational
1950s American jazz musicians: Fine-grained isa
Linguistics.
extraction via modifier composition. In Proceedings
Ruiji Fu, Jiang Guo, Bing Qin, Wanxiang Che, Haifeng of the 55th Annual Meeting of the Association for
Wang, and Ting Liu. 2014. Learning semantic Computational Linguistics (Volume 1: Long
hierarchies via word embeddings. In Proceedings of Papers), pages 2099–2109. Association for
the 52nd Annual Meeting of the Association for Computational Linguistics.
Computational Linguistics (Volume 1: Long
Tu Vu and Vered Shwartz. 2018. Integrating
Papers), pages 1199–1209, Baltimore, Maryland.
Multiplicative Features into Supervised
Association for Computational Linguistics.
Distributional Methods for Lexical
Koki Washio and Tsuneaki Kato. 2018. Neural Latent Entailment. arXiv preprint arXiv:1804.08845.
Relational Analysis to Capture Lexical Semantic
Lior Rokach. 2009. Ensemble-based
Relations in a Vector Space. arXiv preprint
classifiers. Artificial Intelligence Review, 33(1-
arXiv:1809.03401.
2):1-39.
Chengyu Wang, Xiaofeng He, and Aoying Zhou. 2017.
Bojanowski Piotr, Edouard Grave, Armand Joulin and
A short survey on taxonomy learning from text
Tomas Mikolov. 2017. Enriching word vectors with
corpora: Issues, resources and recent advances. In
subword information. Transactions of the
Proceedings of the 2017 Conference on Empirical
Association for Computational Linguistics, 5, 135-
Methods in Natural Language Processing, pages
146.
1190– 1203.
Enrico Santus, Frances Yung, Alessandro Lenci, and
Stephen Roller, Katrin Erk, and Gemma Boleda. 2014.
Chu-Ren Huang. 2015. Evalution 1.0: an evolving
Inclusive yet selective: Supervised distributional
semantic dataset for training and evaluation of
hypernymy detection. In Proceedings of COLING
distributional semantic models. In Proceedings of
2014, the 25th International Conference on
the 4th Workshop on Linked Data in Linguistics
Computational Linguistics: Technical Papers, pages
(LDL- 2015), pages 64–69.
1025– 1036, Dublin, Ireland.
Enrico Santus, Alessandro Lenci, Qin Lu, and Sabine
Vered Shwartz, Yoav Goldberg, and Ido Dagan. 2016.
Schulte im Walde. 2014. Chasing hypernyms in
Improving hypernymy detection with an integrated
vector spaces with entropy. In Proceedings of the
path-based and distributional method. In
14th Conference of the European Chapter of the
Proceedings of the 54th Annual Meeting of the
Association for Computational Linguistics, volume
Association for Computational Linguistics (Volume
2: Short Papers, pages 38–42, Gothenburg, Sweden.
1: Long Papers), pages 2389–2398, Berlin,
Association for Computational Linguistics.
Germany. Association for Computational
Linguistics. Omer Levy, Steffen Remus, Chris Biemann, and Ido
Dagan. 2015. Do supervised distributional meth-
Stephen Roller and Katrin Erk. 2016. Relations such as
ods really learn lexical inference relations? In
hypernymy: Identifying and exploiting Hearst
Proceedings of the 2015 Conference of the North
patterns in distributional vectors for lexical
American Chapter of the Association for
entailment. In Proceedings of the 2016 Conference
Computational Linguistics: Human Language
on Empirical Methods in Natural Language
Technologies, pages 970–976, Denver, Colorado.
Processing, pages 2163–2172, Austin, Texas.
Association for Computational Linguistics.
Association for Computational Linguistics.
Marti A Hearst. 1992. Automatic acquisition of
hyponyms from large text corpora. In Proceedings
of the 14th conference on Computational
linguistics, pages 539–545.
Roberto Navigli and Paola Velardi. 2010. Learning
word-class lattices for definition and hypernym ex-
traction. In ACL, pages 1318–1327.
486
6
Opinions Summarization: Aspect Similarity Recognition
Relaxes the Constraint of Predefined Aspects
487
Proceedings of Recent Advances in Natural Language Processing, pages 487–496,
Varna, Bulgaria, Sep 2–4 2019.
No. Sentence Aspect Similarity non-neural techniques, we refer the reader to Liu
The pc runs so fast. and Zhang (2012).
1 Yes
I like its performance and price. For extractive generic summarization, Cao et al.
The food is cheap. (2015) rank sentences in a parsing tree via a recur-
2 No
The shop’s location is good. sive neural network. However, the model requires
I love its pizza. handcrafted features as input. Cheng and Lapata
3 Yes
I bought food from the restaurant. (2016) propose an end-to-end model for extract-
ing words and sentences. In this system, a doc-
Table 1: Some samples of Aspect Similarity ument is encoded via convolutional and recurrent
Recognition layers, then an attention architecture is employed
to extract sentences and words. Follow this work,
units have at least one aspect in common, which Zhou et al. (2018) enhance the previous system by
is called Aspect Similarity Recognition - ASR jointly learning to score and select sentences. By
(Nguyen et al., 2018), rather than explicitly ex- integrating sentence scoring and selecting into one
tracting aspects of each text unit. Table 1 shows phase, as the model selects a sentence, the sen-
some samples of the ASR task. Follow this ob- tence is scored according to the partial output sum-
servation, we propose an aspect-based summariza- mary and current extraction state.
tion using ASR instead of aspect discovery. The To our knowledge, the first neural-based model
advantage of ASR is to learn patterns and relations of extractive opinions summarization is proposed
between two text units and not need to identify by Kågebäck et al. (2014), which uses an un-
the aspects of each unit, therefore it is potential folding recursive auto-encoder to learn phrase em-
to cross-domain application. Our contributions in beddings and measures similarity by Cosine and
this work are as follows: Euclidean distance. The limitation of this sys-
tem is to purely rely on semantic similarity with-
• We propose an attention-cell LSTM model out taking into account the aspect information.
(ACLSTM) for ASR which enhances the Yang et al. (2017) use the unsupervised neural
LSTM model via employing attention sig- attention-based aspect autoencoder (ABAE) (He
nals into the input gate and the memory cell. et al., 2017b) for presenting each aspect in an as-
ACLSTM shows improvements compared to pect embedding space. Then, the representative
the conventional attention models for both sentence for each aspect is selected via its dis-
settings of in-domain and cross-domain. tance with the centroid of that aspect. For sum-
marization, however, ABAE is not efficient com-
• We introduce a novel aspect-based summa- pared to K-mean in the aspects which occur more
rization using Aspect Similarity Recognition. frequently in the dataset. Angelidis and Lapata
According to the experiments, our method (2018) introduce seed words of each domain to
outperforms strong baselines on Opinosis the autoencoder ABAE. This weakly-supervised
corpus. We also evaluate our method in re- model which is trained under multi-task objec-
gard to domain adaptation. tive outperforms the unsupervised model for as-
pect extraction. Different from the previous work
The remainder of this paper is organized as fol-
in aspect-based opinions summarization, we apply
lows: Section 2 reviews the related research, Sec-
aspect similarity recognition (ASR) instead of as-
tion 3 describes the problem formulation, Section
pect extraction. ASR facilitates the problem of do-
4 and 5 respectively introduce the attention-cell
main adaptation in summarization.
LSTM for ASR and the proposed summarization
using ASR, Section 6 discusses the experiments
for ASR and summarization, and Section 7 con- 3 Problem Formulation
cludes our work and future work. Every product e contains a set of reviews Re =
{rie , ..., rne } expressing users’ opinions on that
2 Related Work
product. A review rie is viewed as a sequence of
In the scope of this paper, we focus on discussing sentences (s1 , ..., sm ). For each product e, our
neural-based systems for generic and opinions goal is to select the most salient sentences in re-
summarization. For a comprehensive literature of views Re for producing a summary. The proposed
488
Figure 1: The proposed framework for the aspect similarity task
approach is divided into subtasks as follows: how to combine the polarity, semantic and aspect
1. Sentiment prediction determines the overall similarity to produce a summary.
polarity ps ∈ [−1, +1] a sentence carries,
4 Attention Cell LSTM
where −1, +1 respectively indicate maxi-
mally negative and positive. According to According to Nguyen et al. (2018), recurrent neu-
Angelidis and Lapata (2018), highly positive ral networks efficiently capture aspect relation-
or negative opinions are more likely to con- ships. For dealing with the remaining difficulties
tain informative text than neutral ones. In of this task, the authors analyze the necessary of
our system, we use the ensemble sentiment an attention mechanism. For that reason, we aim
classifier proposed by Huy Tien and Minh Le to emphasize salient words as encoding sentences
(2017), which achieves strong performances over LSTM. A straightforward approach is to learn
at sentence level. attention signals by self-attention and then apply
these signals into inputs before feeding them into
2. Semantic textual similarity measures the
LSTM. In other words, these attention signals are
semantic similarity qij of two sentences i and
applied to all gates of a LSTM cell. However,
j, which plays an important role in identify-
we assume emphasized input makes the cell for-
ing the most informative sentences as well
get more information on the previous state (the
as redundant ones. We use the state-of-the-
forget gate’s function) while this state stores the
art multi-level comparison model (Tien et al.,
most salient information by the support of atten-
2018) for this task.
tion signals. This conflict causes the inefficiency
3. Aspect similarity Recognition (ASR) pre- of integrating attention signals with LSTM. There-
dicts a probability rij that two sentences i and fore, we propose a novel LSTM cell which pre-
j shares at least one aspect. This subtask fa- vents the state from forgetting too much salient in-
cilitates the elimination of redundant text in formation as employing attention signals for en-
summarization, especially for domain adap- coding sentences. For the ASR task, the proposed
tation. attention-cell LSTM outperforms the conventional
LSTM with/out using attention in both of settings:
4. Summarization Generation employs the in-domain and cross-domain.
three signals above for ranking sentences. By representing a word wi by a pre-trained
A concise and informative summary of a word embedding ewi , we construct a sentence S
product e is generated by selecting the most of n words as a sequence of n word embeddings
salient sentences from reviews Re . S = [ew1 , ew2 , ..., ewn ]. Contextual information
Section 4 describes in details the attention-cell is incorporated in the word embeddings over the
LSTM for the ASR task and Section 5 explains bidirectional GRU (Bahdanau et al., 2014) and
489
then the self-attention signal ai of wi is learned metrics are used to evaluate the relationship be-
as follows (from Yang et al. (2016)): tween two sentences es1 and es2 as follows (from
Nguyen et al. (2018)):
←
− ←−−− Cosine similarity:
hi = GRU (ewi ) (1)
→
− −−−→ es1 · es1
hi = GRU (ewi ) (2) dcosine = (14)
←− → − kes1 k kes2 k
hi = hi ⊕ hi (3)
ui = tanh(Wa hi + ba ) (4) Multiplication vector & Absolute difference:
490
1. In the first step t = 0, we score each sentence 3. We repeat step 2 until the number of selected
si ∈ D and select the most salient sentence sentences is reached or the most salient score
ŝ0 for the summary K: at the current step t is lower than a threshold.
To avoid missing topic words in a summary,
1 X in step 1 and 2, we only select sentences con-
aspt=0
si = rij (23)
|D| taining words belonging to the list of frequent
j∈D
1 X words on that topic. According to our obser-
simt=0
si = qij (24) vation, the topic words are the most frequent.
|D|
j∈D
salst=0
i
= (1 + α|psi |) ∗ aspt=0 t=0
si ∗ simsi 6 Experiments & Results
(25)
6.1 Aspect Similarity Recognition
ŝ0 = arg max{salst=0 } (26)
i
si ∈D We evaluate the attention-cell LSTM on ASRcor-
K t=1
= K ∪ {ŝ } 0
(27) pus (Nguyen et al., 2018), which contains sen-
t=1 0 tences from the SemEval 2016 dataset with two
D = D \ {ŝ } (28)
domains: RESTAURANT and LAPTOP. Each
At the step t = 0, the salient salsi is sample is a pair of sentences annotated as aspect
computed by the semantic similarity simsi , similarity (label = 1) or not aspect similarity
the aspect coverage simsi and the polarity (label = 0). Table 2 reports the statistic of AS-
psi . Different from the previous works, we RCorpus in details.
also take into account the aspect coverage in
RESTAURANT LAPTOP
which a sentence carrying more aspects has a
Train Dev Test Train Dev Test
higher salient score. In addition, the polarity
of a sentence contributes to its ranking by a Sentences 1239 469 587 1657 382 573
coefficient α ∈ [0, 1]. Sentence pairs 458K 68K 98K 447K 26K 44K
Similarity 229K 34K 49K 223K 13K 22K
2. In the next step t, the salient sentence ŝt is Not similarity 229K 34K 49K 223K 13K 22K
selected as follows:
Vocabulary 3769 3649
1 X
asptsi = rij (29) Table 2: Statistic of ASRCorpus
|Dt |
j∈Dt
1 X We compare our model to some strong baselines
simtsi = qij (30)
|Dt | t
as well as the conventional recurrent networks us-
j∈D
ing attention. We choose the optimal values of
¯ t=0 = (1 + α|ps |) ∗ aspt ∗ simt (31)
sal hyper-parameters in our model and baselines via
si i si si
a grid search on 30% of LAPTOP domain. Be-
To avoid the redundant information, we pe- cause the number of RESTAURANT’s categories
nalize each sentence si by the aspect similar- is smaller than LAPTOP’s, the performance of
ity acovst i and semantic similarity scovst i of RESTAURANT domain is better.
that sentence with the selected sentences, in Table 3 reports the experimental results. By em-
which β is a coefficient: ploying efficiently attention signals, the attention-
1 X cell LSTM outperforms the conventional recurrent
acovst i = rij (32) models using attention. As we analysis in Sec-
|K t | t j∈K tion 4, applying attention to all gates of a LSTM
1 X cell causes the conflict of remembering and for-
scovst i = qij (33)
|K t | t
getting. This drawback makes the training of the
j∈K
LSTM-attention model inefficient. Consequently,
¯ t − β ∗ acov t ∗ scov t (34)
salst i = sal si si si the trained LSTM-attention model predicts the
ŝt = arg max {salst i } (35) same label for all inputs.
si ∈Dt
We also evaluate how the models perform in
K t+1 = K ∪ {ŝ }
t t
(36) cross-domain setting where the models are trained
t+1 t t
D = D \ {ŝ } (37) on one domain dataset and tested on the other.
491
These results also prove that these approaches are 6.2 Opinion Summarization
potential to cross-domain application. We observe
The Opinosis dataset (Ganesan et al., 2010) in-
that a set of salient words in each domain is differ-
cludes user reviews of 51 different topics (e.g., ho-
ent. Therefore, the support of attention signals in
tel, car, product). Each topic includes between 50
domain adaptation is not significant compared to
and 575 sentences made by various authors and
the recurrent models without attention.
around 4 reference summaries created by human.
The corpus is suited for opinion summarization as
Method RES LAP →RES →LAP well as evaluating the ability of domain adapta-
Word Average 70.75 65.12 54.5 54.59 tion.
CNN 77.57 67.23 54.08 54.49 We use ROUGE to assess the agreement of
LSTM 79.4 70.21 59.1 57.59 generated summaries and gold summaries. Our
BiLSTM 79.2 71.14 59.2 57.95 experiments include ROUGE-1, ROUGE-2 and,
Attention 78.79 68 57.92 54.55 ROUGE-SU4, which base on one-gram, bi-gram
and skip-bigram co-occurrences respectively.
LSTM-attention 50 50 50 50
The model for each subtask in our summariza-
Attention-Cell LSTM 80 72.73 59.77 58.1
tion system is implemented as follows:
Attention-Cell BiLSTM 79.42 71.65 59.3 58
Table 3: The in-domain and cross-domain exper- • Sentiment prediction: the ensemble classifier
imental results on the two domains: RESTAU- (Huy Tien and Minh Le, 2017) is trained on
RANT and LAPTOP. ”→Y” denotes that models Stanford Sentiment Treebank (Socher et al.,
are tested on Y but trained on the other. Accuracy 2013) with the accuracy of 88.6%.
metric is used for evaluation. The results are sta-
• Semantic textual similarity : the multi-level
tistically significant at p < 0.05 via the pairwise
comparison model (Tien et al., 2018) is
t-test.
trained on STSbenchmark1 with the accuracy
of 82.45%.
To obtain deeper analysis, we inspect the
attention-cell LSTM’s performance on each class • Aspect similarity recognition: the attention-
(e.g., “similarity” and “not similarity”) by preci- cel LSTM is trained on the ASRcorpus of the
sion, recall and F1 scores reported in Table 4. In both domains with the accuracy of 76.2%.
both of the domains and settings, the model per-
forms better on “not similarity” class than “sim- • Summary generation: we set α = 1.67 and
ilarity” class in terms of F1 score. According to β = 0.1. The number of the most frequent
the results in cross-domain setting, we could con- words is three. These parameters are opti-
clude that the models learn rules, patterns for iden- mized over a set of 5 topics randomly se-
tifying aspect similarity rather than remembering lected from the Opinosis dataset. According
topic words and keywords in a particular domain. to the analysis of (Ganesan et al., 2010), the
size of a summary is two sentences.
Domain Class P recision Recall F1
For comparison, we use MEAD (Radev et al.,
Not Similarity 0.76 0.88 0.81
RES 2000) and CW-AddEuc (Kågebäck et al., 2014)
similarity 0.86 0.82 0.78
as baselines. MEAD is an extractive method
Not Similarity 0.68 0.87 0.76
LAP based on cluster centroids which selects the salient
similarity 0.82 0.59 0.68
sentences by a collection of the most important
Not Similarity 0.58 0.68 0.63 words. CW-AddEuc measures the Euclidean sim-
→RES
similarity 0.62 0.52 0.56 ilarity between two sentences by their continuous
Not Similarity 0.56 0.72 0.63 vector space. In addition, we also report the contri-
→LAP
similarity 0.61 0.44 0.51 bution of using aspect and sentiment information
in summarization. The results denoted OPTR and
Table 4: The attention-cell LSTM’s performance
OPTF in Table 5 describe the upper bound score
on each class.
1
http://ixa2.si.ehu.es/stswiki/index.php/STSbenchmark
492
ROUGE-1 ROUGE-2 ROUGE-SU4
Method
Recall Precision F Recall Precision F Recall Precision F
OPTR 57.86 21.96 30.28 22.96 12.31 15.33 29.5 13.53 17.7
OPTF 45.93 48.84 46.57 20.42 19.94 19.49 23.17 26.5 23.7
MEAD 49.32 9.16 15.15 10.58 1.84 3.08 23.16 1.02 1.89
CW-AddEuc 29.12 22.75 24.88 5.12 3.6 4.1 10.54 7.59 8.35
The proposed summarizer
Semantic 28.24 28.63 27.62 7.34 7.19 7 10.69 10.94 10.4
Semantic + Aspect 29.2 29.19 28.24 7.45 7.29 7.12 11.25 11.26 10.78
Aspect + Polarity 27.77 27.86 26.92 7.24 7.09 6.93 10.42 10.55 10.04
Semantic + Aspect + Polarity 28.56 28.31 27.5 7.06 6.84 6.71 10.92 10.83 10.4
of recall and F-score respectively. As the refer- we) are seldom present in a summary. These fac-
ence summaries of Opinosis are generated in ab- tors lead to an unexpected result of using polarity
stractive approach by humans, our generated sum- information in summarization although sentences
maries cannot fully match with the reference sum- carrying the most polarity are still informative.
maries. For example, the maximum recall which
an extractive method could achieve in ROUGE-1 Domain Class Semantic + Aspect
is 57.86%. More informative 33%
Tablet Less informative 13%
While MEAD selects long sentences (around 75
Equally informative 54%
words) containing a lot of salient words to achieve
More informative 17%
a high score in recall but low in precision, our
Others Less informative 8%
approach obtains a balance between these scores
Equally informative 72%
with quite shorter sentences (around 17 words).
Table 7: Informative test for using Semantic with
Positive sentences
Aspect against without Aspect.
I purchased a 2007 Camry because of the looks of the re-
designed model and because of the legendary Toyota qual-
ity and reliability. We expect that aspect signals support to gen-
The Concierge staff, exceptional and extremely helpful, erate an informative summary, which is a sum-
right from suggestions on transportation excursion options mary carrying salient information on various as-
to recommending an amazing restaurant. pects. However, the ROUGE metric measures the
When I checked in, I asked to be shown several rooms and number of matches between two pieces of text, so
the staff was happy to do so.
it is difficult to compare which one is more infor-
Negative sentences mative. Therefore, we execute an informative test
My wife does say the vehicle is not as comfortable for long to understand whether aspect signals help to gen-
trips as other cars we’ve owned.
erate a more informative summary. Given refer-
We had to go up a floor and into a service area to find ice. ence summaries and two summaries generated by
The rude and poor service started from the concierge who the system with/out using aspect signals respec-
was curt when I asked a question .
tively, three persons are asked to select one of the
Table 6: Some sentences carrying the most polar- three answers: which system’s summary is more
ity in the Opinosis dataset. informative, or both of them are equally informa-
tive. The inter-rater agreement Cohen’s Kappa
To analyze why sentiment signals cause neg- score for each pair of assessors is higher than 0.74.
ative impacts on the summarization generation, The overall answer is concluded by the majority
we inspect the most polarity sentences in the cor- vote scheme. In case of receiving three differ-
pus. Some typical sentences are listed in Table ent answers, that pair of summaries is assigned
6. We observe that most of these sentences ex- as equally informative. The result reported in Ta-
press individual experiences and too subjective to ble 7 includes domain specification (15 samples in
be selected for summarization. According to the T ablet and 36 samples in Others), which facili-
Opinosis dataset, overstrong words (i.e., rude, ex- tates the evaluation of domain adaptation. As the
tremely) and subjective words (i.e., my wife, I, ASR system is trained on the restaurant and laptop
493
Summary on the Comfort of Toyota Camry 2007
[1] The Camry offers interior comfort, while providing a quiet ride. Comfortable seating and easy to
Human drive.
[2] Overall very comfortable ride front and back. Nice and roomy.
[3] Its very comfortable and a quiet ride with low levels of noise.
Semantic The ride is quiet and comfortable. Very comfortable, quiet interior.
Semantic + Aspect The ride is quiet and comfortable. Very comfortable ride and seating.
Summary on the location of Holiday Inn London
[1] Location is excellent, very close to the Glouchester Rd. Tube stop.
Human [2] Excellent location. Near the tube station.
[3] The location is excellent. The hotel is very convenient to shopping, sightseeing, and restaurants.
It is located just minutes from the tube stations.
Semantic Great location but don’t bring the car! Great location great breakfast!
Semantic + Aspect Great location but don’t bring the car! Great location for the tube and bus!
Table 8: Human and system summaries for some products/services. For each topic, we list three sum-
maries by human.
dataset, we consider tablet’s topics in the Opinosis ROUGE metric, we conducted the informative test
corpus as in-domain and others as out-of-domain. for quality evaluation. However, for a large corpus
According to the informative test, the system with or multiple systems comparison, this test requires
aspect dominates in both of the domains (T ablet a huge amount of human effort. Therefore, it is
and Others). This result proves the contribution a high demand to have a reliable metric for sum-
of aspect signals and the domain adaptation of the maries evaluation without human involvement.
ASR system.
7 Conclusion
To obtain a better view of the advantages and
disadvantages in our system, we show some gen- In this work, we introduced a novel aspect-based
erated summaries against reference summaries in opinions summarization framework using aspect
Table 8. In extractive methods, the most salient similarity recognition. This subtask relaxes the
sentences are selected from different reviewers, so constraint of predefined aspects in conventional
it is possible to have repeated information in a aspect categorization tasks. For ASR tasks, we
summary. For instance in the case #1, the first sen- proposed an attention-cell LSTM to integrate ef-
tence mentions quiet and comfortable ride while ficiently attention signals into LSTM. This ap-
the second one contains ride and seating. Al- proach outperforms the baselines on both settings
though these sentences still have different opin- of in-domain and cross-domain. For summariza-
ions (i.e., quiet vs seating), the repeat of comfort- tion, we evaluated our system on the Opinosis cor-
able ride downgrades the generated summary’s pus. In addition to ROUGE metric, an informative
quality. For improvement, we suggest a post- test with human involvement was implemented to
processing for a more concise summary by fil- show the domain adaptation ability of our system
tering redundant information. As the proposed and how informative our generated summaries are.
aspect-based system ranks a sentence by not only In the corpus, we observe that sentences carrying
semantic cover but also aspect cover, it selects the the most polarity are not suited to summarization.
more salient opinions for summarization. For in- Therefore, employing sentiment for summariza-
stance, although both of the systems extract differ- tion needs deeper analysis. Due to the ASR task’s
ent features (e.g., interior vs seating, breakfast vs advantage, we believe that it has a high demand in
tube and bus), the opinions (i.e., seating, tube and some fundamental tasks of natural language pro-
bus) chosen by the system with aspect support are cessing such as information retrieval, and sentence
more suited to the reference summaries. comparison.
In each topic, although the reference summaries
Acknowledgements
and generated summary share most of the mean-
ing, they deliver information in different ways and This work was supported by JST CREST Grant,
words. This fact makes the quality evaluation of Japan. We are also immensely grateful to the re-
generated summaries difficult. In addition to the viewers for their comments.
494
References Nguyen Huy Tien and Nguyen Minh Le. 2017. An
ensemble method with sentiment features and clus-
Stefanos Angelidis and Mirella Lapata. 2018. Summa- tering support. In Proceedings of the Eighth In-
rizing opinions: Aspect extraction meets sentiment ternational Joint Conference on Natural Language
prediction and they are both weakly supervised. In Processing (Volume 1: Long Papers). Asian Feder-
EMNLP. ation of Natural Language Processing, pages 644–
653. http://aclweb.org/anthology/I17-1065.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua
Bengio. 2014. Neural machine translation by Mikael Kågebäck, Olof Mogren, Nina Tahmasebi, and
jointly learning to align and translate. CoRR Devdatt Dubhashi. 2014. Extractive summariza-
abs/1409.0473. tion using continuous vector space models. In Pro-
ceedings of the 2nd Workshop on Continuous Vector
David M. Blei, Andrew Y. Ng, and Michael I. Jordan. Space Models and their Compositionality (CVSC).
2003. Latent dirichlet allocation. J. Mach. Learn. Association for Computational Linguistics, pages
Res. 3:993–1022. 31–39.
Ziqiang Cao, Furu Wei, Li Dong, Sujian Li, and Ming Bing Liu. 2012. Sentiment analysis and opinion min-
Zhou. 2015. Ranking with recursive neural net- ing. Synthesis lectures on human language tech-
works and its application to multi-document summa- nologies 5(1):1–167.
rization. In Proceedings of the Twenty-Ninth AAAI
Conference on Artificial Intelligence. AAAI Press, Bing Liu and Lei Zhang. 2012. A Survey of Opin-
AAAI’15, pages 2153–2159. ion Mining and Sentiment Analysis, Springer US,
Boston, MA, pages 415–463.
Jianpeng Cheng and Mirella Lapata. 2016. Neural
summarization by extracting sentences and words. Samaneh Moghaddam and Martin Ester. 2013.
In Proceedings of the 54th Annual Meeting of the As- The flda model for aspect-based opinion min-
sociation for Computational Linguistics (Volume 1: ing: Addressing the cold start problem. In
Long Papers). Association for Computational Lin- Proceedings of the 22Nd International Con-
guistics, pages 484–494. ference on World Wide Web. ACM, New
York, NY, USA, WWW ’13, pages 909–918.
Kavita Ganesan, ChengXiang Zhai, and Jiawei Han. https://doi.org/10.1145/2488388.2488467.
2010. Opinosis: a graph-based approach to abstrac-
Ani Nenkova, Lucy Vanderwende, and Kathleen
tive summarization of highly redundant opinions. In
McKeown. 2006. A compositional context
Proceedings of the 23rd International Conference on
sensitive multi-document summarizer: Exploring
Computational Linguistics. Association for Compu-
the factors that influence summarization. In
tational Linguistics, pages 340–348.
Proceedings of the 29th Annual International
ACM SIGIR Conference on Research and De-
Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel
velopment in Information Retrieval. ACM, New
Dahlmeier. 2017a. An unsupervised neural atten-
York, NY, USA, SIGIR ’06, pages 573–580.
tion model for aspect extraction. In Proceedings of
https://doi.org/10.1145/1148170.1148269.
the 55th Annual Meeting of the Association for Com-
putational Linguistics (Volume 1: Long Papers). As- Huy Tien Nguyen, Quan-Hoang Vo, and Minh-Le
sociation for Computational Linguistics, Vancouver, Nguyen. 2018. A deep learning study of aspect sim-
Canada. ilarity recognition. In Proceedings of the 10th In-
ternational Conference on Knowledge and Systems
Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel Engineering.
Dahlmeier. 2017b. An unsupervised neural atten-
tion model for aspect extraction. In Proceedings of Dragomir Radev, Timothy Allison, Sasha Blair-
the 55th Annual Meeting of the Association for Com- Goldensohn, John Blitzer, Arda Çelebi, Stanko
putational Linguistics (Volume 1: Long Papers). As- Dimitrov, Elliott Drabek, Ali Hakim, Wai Lam,
sociation for Computational Linguistics, Vancouver, Danyu Liu, Jahna Otterbacher, Hong Qi, Horacio
Canada. Saggion, Simone Teufel, Michael Topper, Adam
Winkel, and Zhu Zhang. 2004. Mead - a plat-
Minqing Hu and Bing Liu. 2004. Mining and sum- form for multidocument multilingual text summa-
marizing customer reviews. In Proceedings of rization. In Proceedings of the Fourth International
the Tenth ACM SIGKDD International Conference Conference on Language Resources and Evaluation
on Knowledge Discovery and Data Mining. ACM, (LREC’04). European Language Resources Associ-
New York, NY, USA, KDD ’04, pages 168–177. ation (ELRA).
https://doi.org/10.1145/1014052.1014073.
Dragomir R. Radev, Hongyan Jing, and Malgorzata
Minqing Hu and Bing Liu. 2006. Opinion extraction Budzikowska. 2000. Centroid-based summarization
and summarization on the web. In Proceedings of of multiple documents: sentence extraction, utility-
the National Conference on Artificial Intelligence. based evaluation, and user studies. In NAACL-ANLP
volume 2. 2000 Workshop: Automatic Summarization.
495
Richard Socher, Alex Perelygin, Jean Y Wu, Jason
Chuang, Christopher D Manning, Andrew Y Ng,
and Christopher Potts. 2013. Recursive deep models
for semantic compositionality over a sentiment tree-
bank. In Proceedings of the conference on empirical
methods in natural language processing (EMNLP).
Citeseer, volume 1631, page 1642.
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky,
Ilya Sutskever, and Ruslan Salakhutdinov. 2014.
Dropout: A simple way to prevent neural networks
from overfitting. Journal of Machine Learning Re-
search 15:1929–1958.
Kai Sheng Tai, Richard Socher, and Christopher D.
Manning. 2015. Improved semantic representations
from tree-structured long short-term memory net-
works. In Proceedings of the 53rd Annual Meet-
ing of the Association for Computational Linguistics
and the 7th International Joint Conference on Natu-
ral Language Processing (Volume 1: Long Papers).
Association for Computational Linguistics, pages
1556–1566.
Huy Nguyen Tien, Minh Nguyen Le, Yamasaki To-
mohiro, and Izuha Tatsuya. 2018. Sentence mod-
eling via multiple word embeddings and multi-level
comparison for semantic textual similarity. CoRR
abs/1805.07882. http://arxiv.org/abs/1805.07882.
Yinfei Yang, Cen Chen, Minghui Qiu, and Forrest Bao.
2017. Aspect extraction from product reviews us-
ing category hierarchy information. In Proceedings
of the 15th Conference of the European Chapter of
the Association for Computational Linguistics: Vol-
ume 2, Short Papers. Association for Computational
Linguistics, pages 675–680.
Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He,
Alex Smola, and Eduard Hovy. 2016. Hierarchi-
cal attention networks for document classification.
In Proceedings of the 2016 Conference of the North
American Chapter of the Association for Computa-
tional Linguistics: Human Language Technologies.
Association for Computational Linguistics, pages
1480–1489. https://doi.org/10.18653/v1/N16-1174.
496
Discourse-Aware Hierarchical Attention Network for
Extractive Single-Document Summarization
Abstract ROOT
herent summary. However, current neu- S4: Meanwhile , West Ham also has bid 12million for
Ecuador ’s Enner Valencia.
ral network-based summarizers treat the
source document as just a sequence of Figure 1: Example of discourse dependency struc-
sentences and ignore the tree-like dis- ture.
course structure inherent in the docu-
ment. To incorporate the information of
a discourse tree structure into the neural The discourse structure consists of discourse re-
network-based summarizers, we propose a lations between units in the input, and discourse
discourse-aware neural extractive summa- information has been shown useful for summa-
rizer which can explicitly take into account rization tasks. An example of a Rhetorical Struc-
the discourse dependency tree structure ture Theory (RST) (Mann and Thompson, 1988)-
of the source document. Our discourse- based discourse structure, expressed as a depen-
aware summarizer can jointly learn the dency tree, is illustrated in Figure 1. In the figure,
discourse structure and the salience score each node corresponds to a sentence. Regarding
of a sentence by using novel hierarchical the relations between the sentences, sentence S2
attention modules, which can be trained elaborates the fact mentioned in sentence S1. In
on automatically parsed discourse depen- addition, S2 is further elaborated by S3. S4 is a
dency trees. Experimental results showed contrast to the mention S1. Such relations are es-
that our model achieved competitive or sential cues for generating a concise and coherent
better performances against state-of-the- summary. For example, elaborated sentences tend
art models in terms of ROUGE scores on to be more important than elaborating sentences,
the DailyMail dataset. We further con- and the elaborated sentences should be included in
ducted manual evaluations. The results the summary while the elaborating sentences are
showed that our approach also gained the not.
coherence of the output summaries. Several Integer Linear Programming (ILP)-
based summarizers (Hirao et al., 2013; Kikuchi
1 Introduction et al., 2014) use the discourse information given
by a discourse parser (Hernault et al., 2010). Thus,
Document summarization is the task of automati- the performance of the summarizers is strongly af-
cally shortening a source document while retain- fected by the performance of the discourse parsers.
ing its salient information. In this paper, we The performance of the parsers deteriorates espe-
present a recurrent neural network (RNN)-based cially when they are applied to documents of a
extractive summarizer taking into account the dis- domain different from the one which they were
course structure inherent in the source document. trained on.
497
Proceedings of Recent Advances in Natural Language Processing, pages 497–506,
Varna, Bulgaria, Sep 2–4 2019.
RNN-based approaches have achieved the state- 2 Related Work
of-the-art performance in document summariza-
tion (Cheng and Lapata, 2016; Nallapati et al., There have long been many attempts at tackling
2017). However, RNN-based summarizers treat extractive single-document summarization (Luhn,
the source document just as a sequence of sen- 1958), but there is still room for improvements in
tences, and ignore the discourse tree structure in- terms of ROUGE scores (Hirao et al., 2017). The
herent in the document. The lack of such infor- recent focus has been on RNN-based approaches
mation limits the ability to correctly compute rel- (Cheng and Lapata, 2016; Nallapati et al., 2017;
ative importance between sentences and reduces Narayan et al., 2018). We further extend the atten-
the coherence of output summaries. Cohan et al. tion mechanism used in RNN-based summarizers
(2018) might be the only exception to the above, to capture a discourse structure.
showing that the effectiveness of incorporating RNN-based approaches were introduced to nat-
discourse information into an RNN-based summa- ural language processing tasks by the pioneering
rizer for scientific papers by treating the source work by Bahdanau et al. (2015) and Luong et al.
document as a sequence of sections such as “In- (2015), originally for machine translation. Rush
troduction” or “Conclusion”. However, they were et al. (2015) applied the approach to a sentence
not able to show how the tree-like discourse struc- compression task. Nallapati et al. (2016) extended
ture is effective in RNN-based approaches for ex- the model to abstractive document summarization.
tractive single-document summarization. The DailyMail dataset (Hermann et al., 2015) has
been commonly used for training abstractive sum-
To effectively avoid the influence of parse er- marizers. Cheng and Lapata (2016) and Nallapati
rors and take advantage of the recent advances in et al. (2017) later proposed the methods to auto-
neural network-based approaches, we propose a matically annotate the binary labels, enabling us
model that jointly learns the discourse tree struc- to train extractive models. Cohan et al. (2018)
ture of the source document and a scoring func- demonstrated the usefulness of incorporating dis-
tion for sentence extraction. Our model represents course information into RNN-based summarizers.
the discourse tree structure as an attention distri- Unlike their model, our attention module explic-
bution and the probability of including a sentence itly captures the hierarchical tree structure inher-
in a summary as the softmax layer. In addition, ent in the document.
recursive attention modules in our model can con- Nallapati et al. (2016) and Yang et al. (2016)
sider multi-hop dependencies between sentences. also used a hierarchical attention that consists of
Therefore, our model can capture the relationships two simple attention modules; one is for words
between sentences effectively and create a sum- and the other is for sentences. Our attention
mary without losing the coherence between sen- mechanism differs from them in that ours cap-
tences. tures discourse tree structures by new hierarchi-
cal attention networks, inspired by the models for
We used an existing RST parser (Hernault et al.,
capturing sentence-level dependency structures,
2010) to add discourse dependency structure an-
e.g. machine translation (Hashimoto and Tsu-
notations to the DailyMail dataset (Hermann et al.,
ruoka, 2017), dependency parsing (Zhang et al.,
2015) and thereby obtained a large-scale annotated
2017), constituency parsing (Kamigaito et al.,
dataset to train the model. One of the advantages
2017) and sentence compression (Kamigaito et al.,
of our model is that we do not need the RST an-
2018). Note that these models were designed
notations in the inference phase because the model
for sentence-level tasks while we focus on the
automatically infers the latent discourse tree struc-
document-level summarization task.
ture of the source document and outputs the prob-
ability for each sentence as a salience score. Sentence selection modules that consider dis-
course structures of documents have been shown
We empirically compared our model with other to be useful in ILP-based summarizers. Hirao et al.
models. The results showed that discourse infor- (2013) attempted to incorporate discourse infor-
mation improves the performance, and also that mation in ILP-based sentence extractors. Kikuchi
our models perform competitively with or better et al. (2014) later proposed another ILP model that
than state-of-the-art neural network-based extrac- takes into account the discourse structure. Their
tive summarizers. model jointly selects and compresses sentences in
498
an ILP summarizer. Unlike the researches above, resent the discourse dependency tree of x. Specif-
our focus is on incorporating discourse informa- ically, element Ek,l equals 1 if the edge from xk to
tion into RNN-based summarizers. xl exists in the discourse tree; otherwise Ek,l = 0.
Penn Discourse TreeBank (PDTB) (Prasad Note that we use the discourse structure matri-
et al., 2008) and RST are the most commonly ces E only in the training phase. The model does
used framework to represent a discourse structure. not require the RST annotations of the source doc-
PDTB focuses on the relation between two sen- ument when calculating the probability distribu-
tences, and the annotated structure for a docu- tion p(yi |x, θ).
ment is not necessarily a tree. In contrast, RST
is forced to represent a document as a tree. Dis- 4 RNN-Based Extractive Summarizer
course parsers for both schema are available (Her- In this section, we first explain the base model and
nault et al., 2010; Feng and Hirst, 2014; Wang give the details of our proposed attention mod-
and Lan, 2015). There are at least two methods ule in the following section. The base model
to convert an RST-based tree structure to a depen- is composed of two main components: a neural
dency structure (Hirao et al., 2002; Li et al., 2014). network-based hierarchical document encoder and
Hayashi et al. (2016) compared these methods and a decoder-based sentence scorer. The document
mentioned that DEP-DT by Hirao et al. (2002) has encoder is further split into two components; a
an advantage for applying to summarization tasks. sentence reader and a document reader. The hi-
We use DEP-DT for this research since we focus erarchical architecture is commonly used in recent
on integrating the tree structure into a summarizer. neural network-based models (Cheng and Lapata,
We found only one model that jointly learns 2016; Nallapati et al., 2017; Cohan et al., 2018).
RST parsing and document summarization (Goyal
and Eisenstein, 2016). They used the SampleRank 4.1 Word Reader
algorithm (Wick et al., 2011), a stochastic struc- The goal of the Word reader is to convert sen-
ture prediction model, while our main focus is tence xi to a sentence embedding hi . For each
to take into account discourse structures in RNN- word wi,j in a sentence xi , the word reader first
based summarizers. convert every word embedding emb(wi,j ) to hid-
den states → −e i,j = LST M (→ −e i,j−1 , emb(wi,j ))
3 Problem Formulation ←
− ←
−
and e i,j = LST M ( e i,j+1 , emb(wi,j )) by using
bi-directional Long short-term memory (LSTM)
We formulate extractive document summarization
(Hochreiter and Schmidhuber, 1997). Then, ← −
e i,j
as a sentence tagging problem. We first briefly ex- →
−
and e i,j are concatenated into a hidden state
plain the notation in the paper and describe the de-
hi,j = [→
−e i,j ; ←
−
e i,j ], where [ ] represents a concate-
tails of the model in the following sections.
nation operation of a vector. After that, all hi,j in
The source document x is represented as a the sentence xi are averaged and represented as a
sequence of sentences x1 , ..., xN . Each sen- sentence embedding hi .
tence xi is composed of a sequence of words
wi,j (1 ≤ j ≤ Mi ), where Mi is the number of 4.2 Sentence Reader
words in xi . We also consider x0 as a dummy root Once we obtain sentence embeddings hi for each
node. The summarizer outputs a sequence of bi- sentence xi , the Sentence reader then reads sen-
nary decisions y = y1 , ..., yN , where yi = 1 for tence embeddings hi by another bi-directional
the i-th sentence xi to be included in the summary LSTM and generates context-aware sentence rep-
and yi = 0 for the sentence not to be included. resentation Hi for each xi . Specifically, two vec-
The binary decisions y are made by using a neu- tors generated by the forward recurrent neural net-
ral network-based probability distribution function →
− →
−
work H i = LST M ( H i−1 , hi ) and the backward
p(yi |x, θ), where θ is the set of learned parame- ←− ←
−
H i = LST M ( H i+1 , hi ) are concatenated into
ters. The model finds the best decisions y by a
sentence representation Hi for xi :
simple greedy search to maximize the sum of the −
→ ← −
probabilities within the length constraint. Hi = [Hi ; Hi ]. (1)
Thus, our goal is to construct a better function We now obtain the context-aware sentence repre-
p(yi |x, θ) given training data D. Each instance in sentations H = {H1 , ..., HN }. Finally, all Hi are
D is a triple (x, E, y), where E is a matrix to rep- averaged to make a document embedding K.
499
4.3 Decoder-Based Sentence Scorer
Sentence Scorer y1 …… y4
This module outputs the probability of including
softmax softmax
xi in the summary, p(yi = 1|x, θ), by using an
LSTM-based decoder. At each time step t (1 ≤ H1 Ω1 H4 Ω4
t ≤ N ), the previous state of the decoder st−1 … S1 … S4
and the sentence representation ht are fed into the Discourse-aware attention
LSTM, and the LSTM outputs a new state; st = Step2. Recursive Attention Step3.
LST M (st−1 , ht ). The initial state s0 is initialized d=2: α2,2,4 = α1,2,3 α1,3,4+α1,2,1 α1,1,4 Selective Attn.
by the last states of the backward LSTM in the
←− γ2 ,4
document reader H0 .
Extractive document summarization often d=1: α1,2,3 Σ
adopts a “hard attention”, which focuses only on
γ 1 ,4
the encoder hidden state Ht in the decoding time
step t (Cheng and Lapata, 2016). In addition, Step1. Parent Attention
document representation K is also important
to decode summaries (Nallapati et al., 2017).
Based on them, the output layer calculates the H1 H2 H3 H4
probability distribution of xt being included in the
summary as: Word/Sentence Readers
p(yt |x, θ) = sof tmax(Wo tanh(Wc [Ht ; st ; K])). (2) Figure 2: Overview of hierarchical attention mechanism
for generating attention vector Ω4 . Parent Attention
first calculates how likely the sentence xi is the parent of xj
5 Discourse-Aware Hierarchical for all combinations. Recursive Attention then gen-
erates weighted sum vectors γd,4 over encoder hidden states
Attention Network Hi considering how likely the sentence xi is the d-th order
parent of x4 . Selective Attention finally generates
We assume that taking into account the discourse another weighted sum vector Ω4 over γd,4 .
dependency structure is also useful in determining
whether the summary includes a target sentence or
not. Here, we make the model capable of account- taking into account the d-th-order parents of xi 1 .
ing for the information of the parent sentences on The module starts the calculation with the set-
the discourse dependency structure by incorporat- ting d = 1. Correspondingly, the module only
ing our proposed hierarchical attention mechanism considers the 1st-order parents of xi . The Parent
into the RNN-based extractive summarizer. Attention Module has already calculated the prob-
As shown in Figure 2, the goal of our attention ability of xk being the parents of xi . Thus, the
mechanism is to generate an attention vector Ωi Recursive Attention Module simply uses the prob-
containing the information from the parent sen- abilities as the weights α1,k,i for every Hk and out-
tences of xi through the three-step attention mod- puts the weighted sum vector γ1,i .
ules. Below, we first give an overview of each When d = 2, the weights α2,k,i for every Hk
step in the procedure and then formulate the com- are calculated on the basis of how likely xk be-
ponents after that. comes the 2nd-order parent of xi . Here, d-th-order
refers to the distance between xk and xi . For ex-
Step1: Parent Attention Module This module
ample, suppose there are two different paths con-
calculates the probability of xk being the parent of
necting two nodes, and that their distances are both
xi for all combinations of k and i where k 6= i. We
2, illustrated by the path colored blue and red in
denote this probability as p(k|i, H). In the figure,
Figure 2. The module multiplies the weights of
the starting point of an edge is the parent, and the
the edges on each path, α1,2,3 × α1,3,4 for the
end point is the child. The probability p(k|i, H)
red path and α1,2,1 × α1,1,4 for the blue path,
is used as the weight for the edge from Hk to Hi .
and then the module sums the multiplied values;
The edge weights are passed to the Recursive At-
tention Module. 1
We use the plural form “parents” here because how likely
a sentence becomes the parent of xi is represented as a prob-
Step2: Recursive Attention Module This mod- ability distribution in our model and multiple parents can be
ule outputs the weighted sum vectors γd,i over H, considered.
500
P
α2,2,4 = α1,2,3 × α1,3,4 + α1,2,1 × α1,1,4 . We con- Ωi = d βd,i γd,i . (8)
sider α2,2,4 to be the probability of x2 being the Finally, the output layer receives the concate-
2nd-order parent of x4 . Then, the module uses the nated vector of Hi and Ωi :
value as the weight and outputs the weighted sum
p(yi |x, θ) = sof tmax(Wo tanh(Wc0 [Hi ; st ; K; Ωi ])). (9)
vector γ2,4 .
When d > 2, the module recursively calculates 5.2 Objective
the weight αd,k,i by using the previously calcu- The training updates the parameters to maximize
lated weight αd−1,k,i as shown in the next section. both the label probability and 1st-order attention
Step3: Selective Attention Module Once distribution α1,k,l . Specifically, we use the follow-
weighted sum vectors γd,i have been obtained tak- ing loss function for optimization:
PN PN
ing into account the d-th-order parents of xi , this − log p(y|x) − λ · k=1 i=1 Ek,i log α1,k,i . (10)
module calculates the weights of each order d to In this equation, Ek,i is 1 if the edge from xk to
select a suitable order. The module again calcu- xi exists in the training instance. Thus, all the pa-
lates the weights for every order d and generates a rameters are updated to reproduce the correct la-
weighted sum vector Ωi . bels and edges appearing in the training data D. λ
5.1 Formulation of Attention Modules is a parameter to control the priority of the output
labels or the edges given by an RST parser.
Here, we describe the formulation of each atten-
tion module. The Parent Attention Module calcu- 6 Experiments
lates the probability of xk being the parents of xi
for all combinations of k and i where k 6= i: Data and Preprocessing: We used two different
datasets for the experiments; the DailyMail dataset
p(k|i, H) = sof tmax(g(k, i)),
(3) for training and evaluation, and the DUC2002 test
g(k, i) = vaT tanh(Ua · Hk + Wa Hi ),
set2 only for evaluation.
where va , Ua and Wa are weight matrices. The DailyMail dataset (Hermann et al., 2015)
The Recursive Attention Module recursively consists of news articles extracted from Daily Mail
calculates the probability of xk being the d-th- Online3 and their “story highlights” created by hu-
order parents of xi : man writers. Nallapati et al. (2016) regarded the
(
p(k|i, H) (d = 1), highlights as human-generated abstractive sum-
αd,k,i = PN (4)
l=0 αd−1,k,l × α1,l,i (d > 1). maries. For training extractive summarization
Furthermore, in a discourse dependency tree, models, we need to annotate sentences with binary
ROOT should not have any parent, and a sentence labels for sentence extraction. To do this, Cheng
should not depend on itself. To satisfy these con- and Lapata (2016) used a rule-based approach
straints, we impose the following on α1,k,i : considering the similarity between the original
( document and extracted sentences. On the other
1 (k = 0, i = 0),
α1,k,i = (5) hand, Nallapati et al. (2017) proposed a simple
0 (k = i, i 6= 0).
heuristic for labeling sentences to be included in
The first equation constrains the ROOT node not to the summary by maximizing the ROUGE scores,
have any parent sentence. The second constraint using the highlights as reference summaries. We
ensures that a sentence does not depend on itself. used the latter scheme to annotate the binary la-
The calculated probabilities αd,k,i are then used bels for sentence extraction.
to weigh the vectors in H, and the weighted sum As a preprocessing, we applied the HILDA
vector γd,i is generated as: parser (Hernault et al., 2010) to annotate RST-
PN
γd,i = k=0 αd,k,i Hk . (6) based discourse information for all the documents.
Once the weighted sum vector γd,i is obtained The RST trees were then converted into depen-
for each order d, the Selective Attention Module dency structures by using the method described
calculates the weights βd,i for each γd,i to find a in Hirao et al. (2013). The parser requires the
suitable order: features extracted from word surfaces and the in-
formation on paragraph boundaries. However,
βd,i = sof tmax(Wβ [Hi ; si ; K]), (7)
2
https://www-nlpir.nist.gov/projects/
where Wβ is a weight matrix. The attention vector duc/guidelines/2002.html
is obtained as a weighted sum of γd,t : 3
http://www.dailymail.co.uk/
501
the preprocessed DailyMail dataset4 provided by our models with previously reported performances
Cheng and Lapata (2016) and commonly used on the DUC2002 test set. LREG is a feature-rich
for summarization tasks was not suitable for our logistic regression based approach used as a base-
use. The dataset is anonymized; all named enti- line in Cheng and Lapata (2016). ILP is a phrase-
ties were replaced by the special token @entity, based extraction system proposed by Woodsend
and paragraph boundaries were deleted. There- and Lapata (2010). The approach extracts the
fore, we used the non-anonymized version of the phrases and recombines them subject to the con-
dataset provided by Hermann et al. (2015). We straints in the ILP such as length, coverage or
obtained 196,557 training documents, 12,147 vali- grammaticality. Both TGRAPH (Parveen et al.,
dation documents and 10,396 test documents from 2015) and URANK (Wan, 2010) are graph-based
the DailyMail dataset. sentence extraction approaches, that perform well
The DUC2002 test set consists of 116 pairs of on the DUC2002 corpus.
source documents and their extractive summaries, Evaluation Metrics: We conducted both auto-
and 567 pairs of source documents and their ab- matic evaluation and human evaluation. In au-
stractive summaries. We used the dataset for eval- tomatic evaluation, we adopted ROUGE scores
uation on out-of-domain data. (Lin, 2004). We specifically calculated ROUGE-
Compared Models: We compared our models 1, ROUGE-2 and ROUGE-L by using the Pyrouge
with various baseline models. DIS w/ PAR is library6 . The highlights in the Dailymail dataset
our model with the model parameter λ > 0. With were treated as reference summaries when we cal-
this setting, all the parameters are tuned to repro- culated the scores. We used three length con-
duce the correct labels and the edges given by the straints; 75 bytes, 275 bytes (Nallapati et al., 2017;
RST parser. DIS fixed is the discourse-aware Cheng and Lapata, 2016) and the bytes of ref-
model with the attention vector Ωt = Ht−1 . Thus, erence summaries. We truncated generated sum-
this model always treats the preceding sentence as maries in the middle to conform to the length con-
the parent. DIS w/o PAR is the model with the straints. We adopted the last constraint to evalu-
model parameter λ = 0. Note that the objective ate whether a model can include sufficient infor-
function in this model does not take into account mation within the ideal summary length. For the
the RST annotations given by the RST parser. evaluation on out-of-domain data, we report the
Thus, all the discourse structures are learned to ROUGE scores on the DUC2002 abstractive and
reproduce the correct sentence labels without the extractive test sets. Our models are trained on the
information from the parser. DailyMail dataset and tested on DUC2002.
We compared the above models with the model We additionally carried out human evaluation
without any discourse-aware attention mecha- because ROUGE scores cannot capture the co-
nisms (no-attn) to verify the effectiveness of herence, though our attention modules are de-
our attention mechanisms. Lead-3 is a com- signed to improve the coherence of summaries.
mon baseline to select the first three sentences. We used Amazon Mechanical Turk to conduct hu-
SummaRuNNer is a well-known RNN-based man evaluation. Specifically, randomly selected
summarizer by Nallapati et al. (2017). This model 100 documents and their four summaries gener-
uses some types of information that we do not ated by DIS w/ PAR, Lead-3, no-attn, and
use, such as the similarity between the source doc- SummaRuNNer were shown to the workers. Five
ument and the target sentence, and the novelty workers were asked to rate each summary on a 1-5
score of the target sentence, while our approach scale in terms of coherence and informativeness.
incorporates the information on the parent sen- The instruction shown to the workers follows the
tence of the target sentence. NeuralSum is DUC quality question7 .
also a neural network-based summarizer which Training Details: We used Adam (Kingma and
uses convolutional neural networks in the encoder. Ba, 2015) for the optimizer, where the learning
Refresh is a state-of-the-art method using rein- rate was set to 0.001. In accordance with the
forcement learning (Narayan et al., 2018) 5 . model parameters used in Nallapati et al. (2017),
In addition to the above methods, we compared 6
The options for the Rouge script were “-a -c 95 -m -n 2
4
http://homepages.inf.ed.ac.uk/mlap/ -b 75” and “-a -c 95 -m -n 2 -b 275”.
index.php?page=resources 7
https://duc.nist.gov/duc2007/
5
We used the implementation provided by the authors. quality-questions.txt
502
we limited the vocabulary size of the input to ROUGE scores on DUC2002 dataset: The re-
150,000 and replaced the out-of-vocabulary words sults are shown in Table 2. Neural network-based
with the token UNK. The size of the mini-batch approaches achieved similar scores on the abstrac-
was set to 8. We used the size of 100 for hid- tive test set because the length constraint is long;
den layers in the LSTMs and 300 for word em- specifically it was set to 100-words. However,
beddings, which were initialized with pre-trained graph-based approaches (TGRAPH and URANK)
embeddings, word2vec-slim. Note that the previ- performed better than the neural network-based
ous researches (Nallapati et al., 2017; Cheng and approaches. As reported in Nallapati et al. (2017),
Lapata, 2016) also used pre-trained embeddings. neural network-based approaches suffer the diffi-
We filtered the training instances consisting of 50 culties in achieving high performance on out-of-
or more sentences in the source document, follow- domain data due to its high capability to fit in-
ing Nallapati et al. (2017). The parameter λ of domain data. Another possible reason might be
all the discourse-aware models was tuned on the the method for creating the binary labels for the
validation set. We tried the following values for λ: training dataset. The binary decisions on the train-
0.01, 0.1, 1.0, 2.0, 5.0 and 10.0. ing dataset were made to maximize the ROUGE-
F scores. Thus, the labels are strongly affected
7 Results and Discussion by the length of reference summaries in the Dai-
lyMail dataset. Since the average length of the
ROUGE scores on DailyMail dataset: Table 1 reference summaries in the DUC test set is longer
shows the ROUGE scores evaluated on the Dai- than the average length in the DailyMail dataset,
lyMail test set. For fair comparison, we also the models trained on the DailyMail dataset might
re-trained the baseline models by using our non- face difficulties. Our proposed models achieved
anonymized dataset. the significantly better performances among the
DIS w/o PAR achieved better ROUGE neural network-based approaches on the extractive
scores than no-attn in all variations of length test sets, which are for the settings with shorter
constraint except for ROUGE-L score on the set- length constraints (50 and 100 words) .
ting with d = {1, 2}. Furthermore, we obtained Human Evaluation: Table 3 shows the re-
better scores for DIS fixed than DIS w/o sults. DIS w/ PAR were evaluated better than
PAR. Incorporating the simple discourse informa- no-attn and SummaRuNNer in terms of co-
tion which treats the preceding sentence as the herence in the settings with all the different length
parent in the objective function improved the per- constraints. These differences are statistically sig-
formance. Exploiting the discourse information nificant with the sign test (p < 0.05). Thus, human
given by the RST parser (DIS w/ PAR) further evaluation also supports the effectiveness of incor-
improved the scores in most settings. These obser- porating discourse information. Lead-3 is in-
vations suggest that discourse information is use- herently strong in terms of coherence because this
ful in RNN-based summarizers. model is constrained to extract consecutive sen-
In the setting with the length constraint of tences while other models possibly extract non-
75 bytes, we observed a statistically significant consecutive ones. It was evaluated better in the
difference between DIS w/ PAR and other setting with 75 bytes length constraint.
neural network-based models (SummaRuNNer , Analysis: Table 4 shows an example of the source
Refresh and NeuralSum) on the settings with document and outputs of two models; the sen-
d = {1, 2} and d = {1, 2, 3}. We also ob- tences selected by our model are colored red and
served the similar tendency in the setting with the those selected by SummaRuNNer are blue, and
length constraint of reference summaries. Fur- those selected by both are purple. In this example,
thermore, we did not observe a statistically sig- S7 elaborates S6. Our summarizer successfully
nificant difference between DIS w/ PAR and extracted S6, that made the output summary more
SummaRuNNer in the setting with the length similar to the gold summary.
constraint of 275 bytes. Those facts would sug-
gest that our models achieve a performance simi- 8 Conclusion
lar to the other baseline models in the setting with
longer length constraints, and can perform better We presented a hierarchical attention network that
with shorter length constraints. captures the discourse dependency structure of the
503
75 275 Ref.
R-1 R-2 R-L R-1 R-2 R-L R-1 R-2 R-L
DIS w/ PAR, d={1} 22.9 9.6 12.1 41.9 17.4 35.2 +
41.1 +
16.9 +
36.7
DIS w/ PAR, d={1,2} +
25.0 +
11.1 +
13.4 41.7 17.4 35.0 41.0 16.7 +
36.7
DIS w/ PAR, d={1,2,3}# +
24.6 +
11.2 +
13.5 41.8 17.2 35.2 +
41.1 +
16.8 +
36.8
DIS w/ PAR, d={1,2,3,4} +
24.1 10.8 +
13.2 41.1 17.5 35.1 +
41.2 +
16.9 +
37.0
DIS fixed, d={1} 23.4 10.3 12.7 39.9 16.1 33.5 39.4 15.7 35.4
DIS fixed, d={1,2} 23.5 10.4 12.8 40.3 15.9 33.6 39.7 16.1 35.8
DIS fixed, d={1,2,3} 22.9 9.8 12.3 40.3 16.4 33.8 39.6 15.9 35.6
DIS fixed, d={1,2,3,4} 22.6 9.2 11.7 39.8 15.6 33.4 39.3 16.1 35.9
DIS w/o PAR, d={1} 21.2 8.1 11.0 40.1 15.8 33.7 39.6 15.5 35.5
DIS w/o PAR, d={1,2} 21.1 7.5 10.6 40.0 15.8 33.0 39.6 15.6 35.5
DIS w/o PAR, d={1,2,3} 20.9 7.9 10.9 40.5 16.1 34.1 40.0 15.8 35.8
DIS w/o PAR, d={1,2,3,4} 21.1 8.0 10.9 40.2 15.7 33.6 39.6 15.5 35.6
Lead-3 23.0 9.4 11.8 41.9 17.0 32.5 40.4 16.3 36.1
no-attn 20.1 7.1 10.4 39.6 15.4 33.3 39.3 15.3 35.2
SummaRuNNer (re-run) 23.2 9.6 11.0 42.0 17.2 32.5 37.6 14.8 33.7
Refresh (re-run) 23.1 10.9 12.6 37.9 16.5 31.4 36.6 15.8 34.1
NeuralSum (re-run) 22.4 9.1 11.8 40.8 16.3 34.8 40.3 15.9 36.1
Table 1: ROUGE Scores on DailyMail dataset. The models are trained and tested on DailyMail dataset. The
length constraints are set to 75 bytes, 275 bytes and the reference length. The best scores among the models in
bold. The symbol + indicates statistical significance using 95% confidence interval with respect to the nearest
baseline, estimated by the ROUGE script. # indicates the model that achieved the best score in ROUGE-2 among
the same methods with different d in the development dataset.
504
References Sepp Hochreiter and Jurgen Schmidhuber. 1997.
Long short-term memory. Neural Computation,
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- 9(8):1735–1780.
gio. 2015. Neural machine translation by jointly
learning to align and translate. In Proceedings of Hidetaka Kamigaito, Katsuhiko Hayashi, Tsutomu Hi-
ICLR2015. rao, and Masaaki Nagata. 2018. Higher-order syn-
tactic attention network for longer sentence com-
Jianpeng Cheng and Mirella Lapata. 2016. Neural pression. In Proceedings of NAACL2018, pages
summarization by extracting sentences and words. 1716–1726.
In Proceedings of ACL2016, pages 484–494, Berlin,
Germany. Hidetaka Kamigaito, Katsuhiko Hayashi, Tsutomu
Hirao, Hiroya Takamura, Manabu Okumura, and
Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Masaaki Nagata. 2017. Supervised attention for
Trung Bui, Seokhwan Kim, Walter Chang, and Nazli sequence-to-sequence constituency parsing. In Pro-
Goharian. 2018. A discourse-aware attention model ceedings of IJCNLP2018 (Volume 2), pages 7–12.
for abstractive summarization of long documents. In
Yuta Kikuchi, Tsutomu Hirao, Hiroya Takamura, Man-
Proceedings of NAACL2018, pages 615–621.
abu Okumura, and Masaaki Nagata. 2014. Single
document summarization based on nested tree struc-
Vanessa Wei Feng and Graeme Hirst. 2014. A linear-
ture. In Proceedings of ACL2014, pages 315–320.
time bottom-up discourse parser with constraints
and post-editing. In Proceedings of ACL2014, pages Diederik P Kingma and Jimmy Ba. 2015. Adam: A
511–521. method for stochastic optimization. In Proceedings
of ICLR2015.
Naman Goyal and Jacob Eisenstein. 2016. A joint
model of rhetorical discourse structure and summa- Sujian Li, Liang Wang, Ziqiang Cao, and Wenjie Li.
rization. In Proceedings of the Workshop on Struc- 2014. Text-level discourse dependency parsing. In
tured Prediction for NLP, pages 25–34. Proceedings of ACL2014, pages 25–35.
Kazuma Hashimoto and Yoshimasa Tsuruoka. 2017. Chin-Yew Lin. 2004. Rouge: A package for auto-
Neural machine translation with source-side latent matic evaluation of summaries. In Proceedings of
graph parsing. In Proceedings of EMNLP2017, ACL2004 Workshop, pages 74–81.
pages 125–135.
H. P. Luhn. 1958. The automatic creation of literature
abstracts. IBM J. Res. Dev., (2):159–165.
Katsuhiko Hayashi, Tsutomu Hirao, and Masaaki Na-
gata. 2016. Empirical comparison of dependency Minh-Thang Luong, Hieu Pham, and Christopher D.
conversions for rst discourse trees. In Proceedings Manning. 2015. Effective approaches to attention-
of SIGDIAL2016, pages 128–136. based neural machine translation. In Proceedings of
EMNLP2015, pages 1412–1421.
Karl Moritz Hermann, Tomas Kocisky, Edward
Grefenstette, Lasse Espeholt, Will Kay, Mustafa Su- William C Mann and Sandra A Thompson. 1988.
leyman, and Phil Blunsom. 2015. Teaching ma- Rhetorical structure theory: Toward a functional the-
chines to read and comprehend. In Proceedings of ory of text organization. Text-Interdisciplinary Jour-
NIPS2015, pages 1693–1701. nal for the Study of Discourse, 8(3):243–281.
Hugo Hernault, Helmut Prendinger, Mitsuru Ishizuka, Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017.
et al. 2010. Hilda: A discourse parser using sup- Summarunner: A recurrent neural network based se-
port vector machine classification. Dialogue & Dis- quence model for extractive summarization of doc-
course, 1(3). uments. In Proceedings of AAAI2017, pages 3075–
3081.
Tsutomu Hirao, Hideki Isozaki, Eisaku Maeda, and
Ramesh Nallapati, Bowen Zhou, Cicero dos Santos,
Yuji Matsumoto. 2002. Extracting important sen-
Ça glar Gulçehre, and Bing Xiang. 2016. Abstrac-
tences with support vector machines. In Proceed-
tive text summarization using sequence-to-sequence
ings of COLING2002, pages 1–7.
rnns and beyond. In Proceedings of CoNLL2016,
pages 280–290.
Tsutomu Hirao, Masaaki Nishino, and Masaaki Na-
gata. 2017. Oracle summaries of compressive sum- Shashi Narayan, Shay B Cohen, and Mirella Lapata.
marization. In Proceedings of ACL2017, pages 275– 2018. Ranking sentences for extractive summariza-
280. tion with reinforcement learning. In Proceedings of
NAACL2018, pages 1747–1759.
Tsutomu Hirao, Yasuhisa Yoshida, Masaaki Nishino,
Norihito Yasuda, and Masaaki Nagata. 2013. Daraksha Parveen, Hans-Martin Ramsl, and Michael
Single-document summarization as a tree knapsack Strube. 2015. Topical coherence for graph-
problem. In Proceedings of EMNLP2013, pages based extractive summarization. In Proceedings of
1515–1520. EMNLP2015, pages 1949–1954.
505
Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt-
sakaki, Livio Robaldo, Aravind K Joshi, and Bon-
nie L Webber. 2008. The penn discourse treebank
2.0. In Proceedings of LREC2008.
Alexander M. Rush, Sumit Chopra, and Jason Weston.
2015. A neural attention model for sentence sum-
marization. In Proceedings of EMNLP2015, pages
379–389.
Xiaojun Wan. 2010. Towards a unified approach to
simultaneous single-document and multi-document
summarizations. In Proceedings of COLING2010,
pages 1137–1145.
Jianxiang Wang and Man Lan. 2015. A refined end-
to-end discourse parser. In Proceedings of the Nine-
teenth Conference on Computational Natural Lan-
guage Learning-Shared Task, pages 17–24.
Michael L Wick, Khashayar Rohanimanesh, Kedar
Bellare, Aron Culotta, and Andrew McCallum.
2011. Samplerank: Training factor graphs with
atomic gradients. In Proceedings of ICML2011,
pages 777–784.
Kristian Woodsend and Mirella Lapata. 2010. Auto-
matic generation of story highlights. In Proceedings
of ACL2010, pages 565–574.
Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He,
Alex Smola, and Eduard Hovy. 2016. Hierarchical
attention networks for document classification. In
Proceedings of NAACL2016, pages 1480–1489.
Xingxing Zhang, Jianpeng Cheng, and Mirella Lapata.
2017. Dependency parsing as head selection. In
Proceedings of EACL2017, volume 1, pages 665–
676.
506
Semi-Supervised Induction of POS-Tag Lexicons with Tree Models
Maciej Janicki
University of Leipzig
NLP Group, Department of Computer Science
Augustusplatz 10, 04109 Leipzig
macjan@o2.pl
507
Proceedings of Recent Advances in Natural Language Processing, pages 507–515,
Varna, Bulgaria, Sep 2–4 2019.
ing model with new vocabulary without complete relations
relatedness
re-training. Furthermore, the closely related tri-
gram HMMs used to be state-of-the-art for a long relation related relate
time and are still used in popular tools like HunPos
(Halácsy et al., 2007; Megyesi, 2009). Transfer- relational relates
ring this result to state-of-the-art tagging methods
correlation
remains a topic for further work.
508
generated independently from a distribution over ambiguous in German. The knowledge that Ger-
arbitrary strings, called ρ. Furthermore, every rule man nouns are always capitalized does not pro-
r has a fixed probability θr of being applied to gen- vide much of a clue, because words belong-
erate a new word. The probability of the whole ing to any other part-of-speech may also oc-
forest is defined as follows: cur capitalized. Even worse, many further sim-
Y ilar words are ambiguous in their meaning and
P r(V, E|R, θR ) ∝ ρ(v) part-of-speech, e.g. Dichten (‘density.N. PL’ or
v∈V0
( capitalized ‘dense.A DJ . NOM . PL . DEF’ or ‘com-
Y Y Y θr if hv, v 0 , ri ∈ E, pose (e.g. a poem).V. INF’), schlichten (‘sim-
×
1 − θr if hv, v 0 , ri ∈
/E ple.A DJ . NOM . PL . DEF’ or ‘mediate.V.I NF’).
v∈V r∈R v 0 ∈r(v)
It is much easier to reason about possible tags
(2)
for those words if we take into account morpho-
V denotes the set of vertices (words) and E the set logically related words. For example, we might
of (labeled) edges of the graph. R denotes the set observe words like richtet or richtete, which to-
of rules and θR the vector of probabilities θr for gether with richten look unambiguously like a
the rules from R. Furthermore, V0 ⊆ V denotes verb paradigm. Similarly, the occurrence of a form
the set of root nodes of the graph and r(v) the set like rechtes can convince us that rechten is an ad-
of words that can be derived from word v by ap- jective, because verbs do not take the suffix -es.
plication of rule r. Note that the latter might be For ambiguous forms, we will likely find parts of
an empty set if the context on the left-hand side of different paradigms, for example schlichtet (verb)
the rule is not matched. Each possible derivation and schlichtes (adjective), which will allow us to
from r(v) corresponds to a Bernoulli variable with notice the ambiguity. Of course, in order to con-
probability θr . duct such analysis, we have to know which affix
(Sumalvico, 2017) describes an unsupervised configurations are characteristic for which part of
fitting procedure for this model using Monte Carlo speech. This is the part that we are going to learn
Expectation Maximization algorithm. The com- from labeled data.
putation involves drawing large samples of graphs
from the conditional distribution P r(E|V, R, θR ) 4.1 Applying Tagged Rules to Untagged
using a Metropolis-Hastings sampler. Words
Let us assume that we have learned a morphol-
4 Computing POS-Tags
ogy model on tagged data. Now we are pre-
We now turn to applying the model introduced in sented with a new set of words, possibly contain-
the previous section to the task of semi-supervised ing many words not present in the original train-
POS tag lexicon induction. The idea is to train ing set. In this section, we will show how the
a model of morphology on a labeled vocabulary trained model can be applied to derive guesses for
(coming from a tagged corpus) and apply this tags in the new vocabulary. The approach follows
model to infer the tags for another vocabulary. the idea sketched in the previous section: the tag
The underlying intuition is that morphological of the word will be determined by the neighbor-
relationships between words can give hints about ing words, together with the knowledge about the
their POS tags and inflectional forms, which are morphology contained in tagged rules.
not visible in the isolated forms. For example, To illustrate the approach with a minimal ex-
consider the following German3 words: Fichten ample, let us assume that our tagset consists of
‘spruce.N. PL’, richten ‘judge.V. INF’, rechten only three tags: NN, VVINF and VVFIN, and
‘right.A DJ . NOM . PL . DEF’.4 Phonetically and or- that the untagged vocabulary consists of the Ger-
thographically they are very similar and all in- man words machen, mache, macht. We compute
clude an inflectional suffix -en, which is highly the edge probabilities for every edge that is possi-
3
It is very difficult to find plausible examples of this phe-
ble according to the model, under every tagging.
nomenon in English due to its small inflection and very pro- For example, the model might consist of the rules
ductive zero-affix derivation. and parameters listed in Table 1.
4
All cited words are ambiguous in their inflectional form
(but not part-of-speech). The glosses shown here are picked Using those values, we can reason about the
as examples. possible taggings based on an untagged graph.
509
r θr lows:
/Xen/NN → /Xe/NN 0.3
/Xen/VVINF → /Xe/VVFIN 0.01 τv,t = EE|V,R,θ ET |V,E,R,θ δTv ,t (3)
/Xen/VVINF → /Xt/VVFIN 0.2
δx,y denotes the Kronecker delta. Thus, we take
the expectation over all possible graphs for the
Table 1: An example model learnt from a tagged
given vocabulary, and then over all possible tag-
vocabulary.
gings for a fixed graph. The inner expecta-
tion can be computed exactly by a variant of
/Xen/ → /Xe/ Forward-Backward algorithm introduced in Sec.
(a) machen mache
4.2. In order to approximate the outer expecta-
tion, we use Markov Chain Monte Carlo sampling
macht
over untagged graphs as described by (Sumalvico,
/Xen/ → /Xe/ 2017). However, the sampling algorithm will need
(b) machen mache
/Xen/
some modifications, as contrary to the original ap-
→ /X proach, the edge probabilities are not independent
t/
macht of the graph structure. We describe those modifi-
cations in Sec. 4.3. Finally, the computed values
Figure 2: Two possible morphology graphs cor- of τv,t will be fed to an already pre-trained HMM
responding to the words machen, mache, macht. to provide it with guesses for the tags of unknown
What does each of them tell us about the possible words, before it is reestimated on untagged text.
tags of those words according to Table 1? This procedure is described in detail in Sec. 4.4.
510
v1 v2 v3 v4 v5 v6 v7 v8 v9 outG (v) denotes the set of outgoing edges of
node v. In order to compute the forward proba-
Figure 3: The Forward-Backward computation bility of a non-root node v, let us assume that it
for a linear sequence in an HMM. αv6 ,t = is derived by the edge (v 0 , v, r). In addition to the
P (v1 , . . . , v6 , T6 = t), whereas βv6 ,t = forward probability of v 0 (the parent of v) and the
P (v7 , v8 , v9 |T6 = t). edge deriving v, we also take into account the side
branches, i.e. all subtrees rooted in children of v 0
v7 v9 other than v. The resulting formula is as follows:
v2 v4 v6 Y 0 00 0 0
v1 v8 αv = αv0 ∗ T (v ,v ,r ) βv00 ·T (v ,v,r)
v3 v5 (v 0 ,v 00 ,r0 )∈outG (v 0 )
v 00 6=v
(7)
Figure 4: The Forward-Backward computation for The last remaining issue is the forward prob-
a tree. Also here, αv6 ,t = P (v1 , . . . , v6 , T6 = t) ability of root nodes. The generative model
and βv6 ,t = P (v7 , v8 , v9 |T6 = t). defined by (2) contains a distribution ρ(·) over
arbitrary strings, from which the string forms
from root to v, but also all side branches sprouting of the root nodes are chosen. In the tagged
from this path. case, we augment this distribution to ρ(v, t) =
We will now derive recursive formulas for for- ρstring (v)ρtag (t |v). As in (Sumalvico, 2017)’s ex-
ward and backward probabilities for the tree case. periments, we use a simple character-level uni-
For this purpose, we assign a transition matrix to gram model for ρstring (·). In order to model the
each graph edge. For each possible tagging of the distribution ρtag (t|v), which predicts a word’s tag
source and target node, the transition matrix con- from its string form, we use a character-level re-
θr
tains a value 1−θ , where r is the corresponding current neural network. Note that the forward
r
tagged rule. Continuing the example from 4.1, the probability of root nodes is equal to their probabil-
probabilities in Table 1 yield the following transi- ity according to the root model, i.e. αv,t = ρ(v, t).
tion matrix: 4.3 Modifications to the Sampling Algorithm
In order to approximate the expected value over
T (machen,mache,/Xen/→/Xe/) = possible graphs given a vocabulary, we use
NN VVINF VVFIN the Metropolis-Hastings sampler proposed by
NN
0.3
0 0 (Sumalvico, 2017). The algorithm computes each
1−0.3 (4)
VVINF 0 0 0.01 new graph by proposing a small change in the pre-
1−0.01
VVFIN 0 0 0 vious graph. The possible moves are: adding or
deleting a single edge, exchanging an edge for an-
other one with the same target node and the so-
Furthermore, let λv,t denote the probability that
called ‘flip’ move. The latter simultaneosly ex-
the node v with tag t is a leaf, i.e. it contains no
changes two edges for two others and is designed
outgoing edges. Then:
as a way to prevent the creation of a cycle while
Y Y
λv,t = (1 − θr ) (5) adding an edge. The algorithm subsequently com-
r∈R (v 0 ,t0 )∈r(v,t)
putes an acceptance probability from the probabil-
ities of the changed edges and decides whether to
It is trivial to see that for leaf nodes, βv = λv . accept the proposed graph as a new sample point.
For a non-leaf node v, we multiply λv by the terms As we have seen in the analysis of Fig. 2,
θr
1−θr for each outgoing edge, summed over every adding an edge typically has consequences for the
possible tagging. In the matrix and vector nota- whole subtree, in which the edge is added. The
tion, this corresponds to the following formula:7 values τv for all nodes in the subtree may change,
Y 0
which in turn changes the probability of all edges
βv = λv ∗ T (v,v ,r) βv0 (6) in the subtree. This behavior constitutes a signifi-
(v,v 0 ,r)∈outG (v) cant difference compared to the original sampling
7
Asterisk denotes element-wise multiplication, while dot algorithm, in which the edge probabilities were in-
or no symbol denotes the dot product. dependent of the graph structure and the cost of a
511
. r . .
v v0 v3 v2 v5 v4 v1
. . .
. . . . . . . . . . .
Figure 5: Adding or removing the edge (v, v 0 , r). Figure 6: In case of a ‘flip’ move, the smallest
subtree containing all changes is the one rooted
change could be easily computed from the cost of in v3 . The deleted edges are dashed, while the
added and removed edges. However, we can use newly added edges are dotted. In order to obtain
the forward and backward probabilities to score the new βv3 , we recompute the backward proba-
the changes. bilities in the whole subtree. αv3 is not affected by
the changes. (The node labels are consistent with
PObserve that for every node v, the value the definition in (Sumalvico, 2017).)
t αv,t βv,t is the probability of whole subtree, to
which v belongs. This property – being able to
compute the probability of a whole subgraph us- (added or deleted) according to the tree, to which
ing the values of a single node – is crucial in eval- they belong (more specifically, according to the
uating the sampler moves. root of the tree, to which the edge’s source node
currently belongs). In each tree, we look for a
Adding or removing a single edge. Consider
minimum subtree that contains all the changes
the graph in Fig. 5, to which the edge (v, v 0 , r)
(Fig. 6). We build a copy of this subtree with all
is supposed to be added. Without this edge, we
changes applied and recompute the forward prob-
have two separate
P trees with Pa total probability ability for its root and the backward probabilities
expressed by ( t αv,t βv,t )( t αv0 ,t βv0 ,t ). After
for the whole subtree. Finally, we use the (newly
adding this edge, we obtain a single tree. As v
computed) forward and backward probability of
obtains a new child node, βv will change. Let βv0
the subtree root to determine the probability of the
denote the new value, which can be computed as
whole tree after changes.
follows:
0
βv0 = βv ∗ T (v,v ,r) βv0 (8) 4.4 Extending an HMM with New
Note that neither αv nor βv0 is affected by adding Vocabulary
this The vectors τv obtained from the sampling ap-
P edge.0 The probability of the new tree is simply
t αv,t βv,t . If the move is accepted, the β values
proach sketched in the previous subsections pro-
of all nodes on the path from the root to v have vide us with a morphologically motivated POS-
to be updated, as well as the α values of all nodes tag distribution for words from the untagged cor-
except for this path. pus. We augment the HMM’s emission probabil-
Deleting an edge involves a very similar com- ity matrix with those values for previously unseen
putation. In this case, words.8
P the probability of the graph
before deletion is t αP v,t βv,t , whereas
P the proba-
bility after deletion is ( t αv,t βv,t 0 )( 5 Experiments
t ρv 0 ,t βv,t ).
0
Here, βv,t is the updated backward probability of 5.1 Experiment Setup
v excluding the deleted edge.
We conducted evaluation experiments for 9 lan-
Other moves. When exchanging a single edge guages: Ancient Greek (GRC), German (DEU),
to another one with the same target node, we al- Finnish (FIN), Gothic (GOT), Latin (LAT), Lat-
ready need to be careful, as two distinct cases vian (LAV), Polish (POL), Romanian (RON) and
arise: either the change takes place within one Russian (RUS), using the Universal Dependen-
tree, or it involves two separate trees. If we pro- cies9 corpora. Each corpus is randomly split into
ceeded as in the previous paragraph, those cases 8
The values τv are conditional probabilities of a tag given
would require different formulas. Instead of con- word, while the emission probabilities of an HMM are prob-
ducting such a detailed analysis of the changes, abilities of a word given tag. We use the Bayes’ formula to
we apply a more general approach that covers the convert the former to the latter. We obtain the marginal prob-
abilities of tags during the HMM pre-training and assume
‘flip’ moves as well. equal frequencies for unseen words.
First, we group all edges that are to be changed 9
http://universaldependencies.org
512
1 Fitting on the training set. 5.2 Evaluation Measures
2 1 + estimation on the development set. Two kinds of evaluation are performed: lexicon
3 2 + extension of the vocabulary with ran- and tagging evaluation. In the first case, the qual-
dom initialization. ity of tag guesses for unknown words, τv , is mea-
4 2 + extension of the vocabulary with sured directly. It is desirable for those values
tag guesses provided by a character-based to not only predict the correct tag for unambigu-
RNN. ous words, but also to handle ambiguity correctly,
5 2 + extension of the vocabulary with tag which means providing probabilities that corre-
guesses provided by the whole-word mor- spond to the expected frequency of a word with
phology model. the certain tag. We derive the gold standard data
6 2 + extension of the vocabulary with gold from the labels in the development set using the
standard tag guesses. following formula:
Table 2: Different setups of the HMM tagger used nv,t
τ̂v,t = P (9)
in the tagging experiment. t0 nv,t0
513
coarse-grained fine-grained Lang. 1 2 3 4 5 6
Lang.
RNN WWM RNN WWM GRC .669 .742 .732 .789 .791 .857
GRC .607 .660 .229 .300 FIN .606 .744 .728 .795 .823 .838
FIN .507 .643 .255 .411 DEU .771 .755 .831 .849 .851 .836
DEU .616 .676 .154 .210 GOT .768 .770 .803 .819 .828 .865
GOT .483 .661 .157 .308 LAT .743 .808 .816 .825 .866 .856
LAT .544 .704 .236 .448 LAV .670 .723 .731 .796 .803 .848
LAV .506 .646 .240 .368 POL .715 .787 .745 .781 .842 .829
POL .587 .695 .196 .315 RON .785 .846 .853 .880 .884 .870
RON .540 .705 .241 .301 RUS .809 .877 .877 .903 .905 .914
RUS .682 .769 .330 .522
Table 4: Tagging accuracy with coarse-grained
Table 3: Lexicon evaluation. tags.
Lang. 1 2 3 4 5 6
because of the absence of rules allowing for alter- GRC .566 .464 .569 .628 .638 .699
natives (a phenomenon illustrated in Fig. 2). Thus, FIN .535 .377 .567 .651 .675 .717
the latter achieves better scores especially on cor- DEU .594 .498 .587 .618 .620 .631
rectly tagged unambiguous words. GOT .636 .550 .659 .677 .678 .739
The comparison of tagging accuracies is shown LAT .590 .493 .617 .660 .687 .734
in Tables 4 and 5. The columns correspond to the LAV .585 .471 .594 .657 .668 .713
tagger configurations explained in Table 2. In gen- POL .554 .437 .575 .625 .627 .679
eral, a rise of the score from left to right is to be RON .712 .713 .759 .764 .761 .824
expected. RUS .731 .654 .736 .780 .789 .813
The difference between columns 1 and 2 il-
lustrates the influence of reestimating the trained Table 5: Tagging accuracy with fine-grained tags.
model on unlabeled data without adding the OOV
words to the vocabulary. Interestingly, this results as if taking into account some wrong taggings
in an improvement in case of the coarse-grained during Baum-Welch estimation could accidentally
tagset, but in a decline when using the fine-grained improve the estimation, because the wrong tag
tagset, both significant. However, adding the OOV might also have occurred in the given context.
words from the development set to the vocabulary, This seems especially plausible for cases like com-
even with randomly initialized probabilities (col- mon and proper nouns, which are often confused.
umn 3), further improves the accuracy (with a few
exceptions), so that the result is consistently better 5.4 Discussion
than column 1. Column 4 introduces intrinsic tag Although the results speak consistently in favor of
guessing as initial probabilities for newly added using morphology-based tag guessing, as well as
words, rather than random values. This results in using tag guessing at all, the benefits are some-
a further improvement, especially significant for what less clear than one could expect. Especially
Finnish, Ancient Greek and Latvian (both coarse- in the case of fine-grained tags, our expectation
grained and fine-grained). was that, due to the discrete nature of morpholog-
The most important comparison in this evalua- ical rules, at least the tags of unambiguous words
tion is between column 4 and 5. This illustrates would be identified mostly correctly. This was
the benefit of using extrinsic tag guessing (column supposed to greatly improve the Baum-Welch es-
5), rather than intrinsic. This results in a consis- timation, as instead of considering many hundred
tent improvement, ranging from very slight to sig- possible tags, the correct one is already known,
nificant. The most significant improvements are which turns the estimation into almost supervised
shown in bold. Finally, column 6 displays what learning. However, we had underestimated the im-
one might expect to be the upper bound on the pact of the small size of training corpus on the
accuracy: the one that would be achieved if tags morphology component. Most fine-grained tags
were guessed perfectly (i.e. as τ̂v ). Surprisingly, it are very rare, so many morphological rules related
is not always the highest value in a row. It looks to such forms are not learnt.
514
References Sylvain Neuvel and Sean A. Fulop. 2002. Unsuper-
vised learning of morphology without morphemes.
Malin Ahlberg, Markus Forsberg, and Mans Hulden. In Proceedings of the 6th Workshop of the ACL
2014. Semi-supervised learning of morphologi- Special Interest Group in Computational Phonology
cal paradigms and lexicons. In Proceedings of the (SIGPHON). pages 31–40.
EACL. pages 569–578.
Milan Straka and Jana Straková. 2017. Tokenizing,
Matthew S. Crouse, Robert D. Nowak, and Richard G. pos tagging, lemmatizing and parsing ud 2.0 with
Baraniuk. 1998. Wavelet-based statistical signal udpipe. In Proceedings of the CoNLL 2017 Shared
processing using hidden markov models. IEEE Task: Multilingual Parsing from Raw Text to Univer-
Transactions on Signal Processing 46(4):886–902. sal Dependencies. pages 88–99.
Jean-Baptiste Durand, Paulo Gonçalvès, and Yann Maciej Sumalvico. 2017. Unsupervised learning of
Guédon. 2004. Computational methods for hid- morphology with graph sampling. In Proceedings
den markov tree models – an application to wavelet to RANLP 2017. Varna, Bulgaria.
trees. IEEE Transactions on Signal Processing
52(9):2551–2560. Zdeněk Žabokrtský and Martin Popel. 2009. Hidden
markov tree model in dependency-based machine
Greg Durrett and John DeNero. 2013. Supervised translation. In Proceedings of the 47th Annual Meet-
learning of complete morphological paradigms. In ing of the Association for Computational Linguis-
Proceedings of NAACL-HLT. pages 1185–1195. tics. Singapore, pages 145–148.
515
Word Sense Disambiguation Based on Constrained Random Walks in
Linked Semantic Networks
516
Proceedings of Recent Advances in Natural Language Processing, pages 516–525,
Varna, Bulgaria, Sep 2–4 2019.
tially good compromise. Such methods can cover, tural properties of existing sense inventories e.g.
at least potentially, all word senses described in a wordnets.
lexical knowledge base. Very large wordnets de- One of the first solutions to WSD task for Pol-
scribing more than 100,000 words (lemmas) and ish (Baś et al., 2008) was mainly focused on su-
their senses by hundreds of lexico-semantic rela- pervised approaches. The authors trained the clas-
tion instances have been constructed for several sifiers on a relatively small dataset for a pre-
languages, including the English and Polish lan- selected, annotated vocabulary. Later works were
guage. A large and dense network of relations focused mainly on WSD methods close to sense
seems to be a good basis for mapping word oc- induction from text corpora, e.g. (Broda and Pi-
currences in texts onto their senses. asecki, 2011).
An enormous amount of semantic data appeared Weakly supervised WSD approaches are usu-
on the Web in recent years. With the growth of ally based on sense inventories and their seman-
resources joining the Linked Open Data collection tic structure describing senses. The main assump-
the available semantic information becomes a very tion is that the words being a part of a text can
important knowledge source for WSD tasks. be mapped onto their potential senses existing in
Semi-supervised methods are very often based a given sense inventory, mainly the synsets exist-
on the idea of mapping texts onto the wordnet ing in a wordnet. The senses activate local sub-
graph of lexico-semantic relations as the initial graphs of wordnet’s semantic structure to activate
activation of the graph nodes. Next a recur- meanings possibly being relevant to a given text
sive spreading activation method, mostly based on fragment. As the initial activation is sparse a kind
PageRank (Page et al., 1998) is applied. However, of spreading activation algorithm is next applied
as seen in the following paragraphs, such methods in order to recursively concentrate this informa-
raise problems with efficiency. tion in some “hot” areas and identify word senses
In this paper we explore Monte Carlo methods located in them or close to them. The identified
based on a random walk algorithm for the task of synsets should be the most likely senses for words
Word Sense Disambiguation. We show that the in the text. There are several parameters to set in
lexical knowledge base with its expansions and this general scheme: the initial activation (coming
the way they are exploited has a strong impact from the text words), spreading activation algo-
on WSD performance and the proposed method rithm (topology and relations) and identification of
allows for the efficient utilisation of large lexi- association between “hot” areas and senses to be
cal knowledge bases. The presented solutions are selected. Various methods following this scheme
evaluated on popular Polish and English bench- were proposed.
mark datasets. Weakly supervised WSD methods are mostly
based on the recursive PageRank algorithm (Page
2 Related Work et al., 1998) for spreading activation. Mihalcea
The first group of approaches is based on su- et al. (2004) proposed application of the original
pervised machine learning algorithms. The re- PageRank to WSD called Static PageRank.
searchers have adapted many different ML meth- PageRank algorithm (henceforth PR) is an it-
ods for WSD task. One can find the classi- erative method for ranking nodes in the graph G.
cal NLP approaches based on feature engineer- In WSD the nodes in G represent synsets and the
ing and popular classification algorithms e.g. de- edges of G correspond to wordnet relations (link-
cision trees, decision lists, Naive Bayes solutions, ing synsets, and in some wordnets also linking
kNN, adaptive boosting, as well as more modern specific word senses).
approaches using Deep Learning with LSTM, biL- (Agirre and Soroa, 2009; Agirre et al., 2014)
STM and more sophisticated variants of neural ar- proposed a modified version called Personalised
chitectures. The main issue is that most of the PageRank (PPR) in which the values in v, called
words are strongly imbalanced in terms of their personalised vector, depend on the textual context
sense distribution, thus, due to the lack of required of the disambiguated word. The non-zero score
training data, the supervised approaches present a values are assigned to those nodes which are con-
lower recall in all-words WSD setting. textually supported. In PPR all words from the
The knowledge-based solutions use the struc- context are disambiguated at once. The v values
517
are equal to: rithm for named entity disambiguation. The mo-
tivation for their work was the fact that PR-based
1 methods mainly rely on global coherence, but the
v[i] = CS
, i = 1, 2, ..., N (1)
N S(i) methods should utilise the local similarity more ef-
fectively.
where CS is the number of different lemmas in the
context, N S(i) – the number of synsets sharing 3 Knowledge Base
the same context lemma with the synset i.
In addition a modified version of PPR The general sense inventory for all-words WSD is
called Personalised PageRank Word-to-Word usually created on the basis of a wordnet. Word-
(PPR_W2W) was also presented by (Agirre and nets can be presented as graphs with nodes repre-
Soroa, 2009; Stevenson et al., 2012), in which senting word senses or synsets (but also with two
a word to be disambiguated is excluded from types of nodes) and edges expressing the struc-
the occurrence contexts, i.e. all synsets of this ture of the lexico-semantic relations described in
word have initial scores in v set to zero. Thus, a given wordnet. The methods based on lexi-
PPR_W2W cannot be run once for all ambiguous cal knowledge bases usually explore the lexico-
words in the context. The vector v must be ini- semantic relations represented by a wordnet to dis-
tialised individually for each ambiguous word in ambiguate the words given the context.
the context – this is a disadvantage of PPR_W2W. Most wordnet relations are paradigmatic, but
A potential advantage is the removal of the effect for WSD we also need syntagmatic relations,
of mutual amplification of the closely connected rarely covered by wordnets, because the plain
senses of the word being disambiguated. All wordnet structure might be insufficient to success-
PR-based WSD algorithms showed good perfor- fully disambiguate a text. As the large public
mance that is increasing with the enlargement and sources like Linked Open Data are available, we
enrichment of the knowledge base. However, with can try to apply them for the expansion of the
larger graphs the time of processing increases lexical knowledge base. They may also contain
non-linearly causing a significant drop in effi- Named Entities which can be very important for
ciency. The problem is especially visible in the WSD, e.g. helping to identify the narrow semantic
case of PPR_W2W in which the algorithm must context.
be restarted several times per context. More formally, a lexical knowledge base is a
In (K˛edzia et al., 2014), PR-based WSD algo- graph G(V, E) consisting of nodes and edges,
rithm for Polish was presented and run with the where V = {v1 , v2 , ..., vN } represents a set of
help of plWordNet 2.1. The graph consisted of concepts (modelling lexical meanings) and E =
synsets linked by edges representing a selected {e1 , e2 , ..., eM } a set of edges corresponding to
subset of the synset relations. Next several ver- lexico-semantic associations linking these con-
sions of the PR-based algorithms, namely Static cepts.
PR, PPR and PPR_W2W was applied to Pol-
3.1 Existing Knowledge Bases
ish texts and plWordNet 2.2 in (K˛edzia et al.,
2015). The achieved precision (on KPWr) was in In literature, we can find many attempts to com-
the range 42.79%-50.73% for nouns and 29.79%- bine Princeton WordNet with resources of many
32.94% for verbs. PPR_W2W produced the best types to obtain a better knowledge base for WSD.
results. Different variants of combining plWord- UKB lexical knowledge base, e.g. (Agirre and
Net with the Suggested Upper Merged Ontology Soroa, 2009), consists of Princeton WordNet 3.0
(SUMO) (Pease, 2011) on the basis of the map- or eXtended WordNet (Harabagiu et al., 1999)
ping constructed in (K˛edzia and Piasecki, 2014). which was expanded by introducing links ex-
All three PR-based algorithm were evaluated. A tracted from SemCor, manually disambiguated
slight improvement of the precision for nouns up glosses from Princeton WordNet Gloss Corpus,
to 50.89% for PPR_W2W could be observed when and Wikipedia.
the two joined graphs were treated as one large BabelNet (Navigli and Ponzetto, 2012) was
graph. initially created by linking the largest multilin-
In (Pershina et al., 2015) the authors presented gual Web encyclopedia, i.e. Wikipedia, with the
a graph-based algorithm using random walks algo- most popular computational semantic lexicon, i.e.,
518
WordNet. Later it was expanded with a number of 3.3 Expansions
resources: The initial performance of the algorithms can be
• OmegaWiki, a large collaborative multilin- moderate when the knowledge-base is limited only
gual dictionary (January 2017 dump); to the plain wordnet structure. We can signifi-
cantly improve the overall performance by intro-
• Wiktionary, a collaborative project to pro- ducing new semantic links to the basis knowledge
duce a free-content multilingual dictionary graph. In this work we expanded the ideas pre-
(February 2018 dump); sented in (Agirre and Soroa, 2009), (Agirre et al.,
2018) and (Moro et al., 2014). Following the pro-
• Wikidata, a free knowledge base that can cedure presented in (Agirre et al., 2018) we ex-
be read and edited by humans and machines tended the structure of our knowledge graph by in-
alike (February 2018 dump); cluding the links extracted from the the Princeton
• Wikiquote, a free online compendium of WordNet Gloss Corpus2 including manually dis-
sourced quotations from notable people and ambiguated glosses.
creative works in every language (March The Suggested Upper Merged Ontology
2015 dump); (SUMO) (Pease, 2011) is a formal representation
of concepts, organised into hierarchies of classes
• VerbNet (Kippera et al., 2006), a Class-Based and subclasses, which is widely used for semantic
Verb Lexicon (version 3.2); analysis in NLP. The lexical senses of PWN 3.0
and plWordNet 3.2 have been mapped onto their
• Microsoft Terminology, a collection of termi- equivalent concepts of SUMO. The mapping
nologies that can be used to develop localised procedure for plWordNet was based on interlin-
versions of applications (July 2015 dumps); gual links existing between plWordNet and PWN
• GeoNames, a free geographical database (Maziarz et al., 2016) and the initial mapping
covering all countries and containing over of PWN senses to SUMO ontology (K˛edzia
eight million place names (April 2015 and Piasecki, 2014). We used the structure of
dump); SUMO ontology as a more general semantic
description for lexical senses existing in wordnet.
• FrameNet (Ruppenhofer et al., 2016), a lex- The semantic structure of our knowledge base
ical database of English that is both human- was extended with concepts and links existing in
and machine-readable (version 1.6); the SUMO ontology by attaching the concepts
to corresponding synsets and linking them with
• Open Multilingual WordNet (Bond and Fos- SUMO relations.
ter, 2013), a collection of wordnets available Wikipedia has opened many new opportunities
in different languages. for semantic analysis. The structure of Wikipedia
The mappings provided by BabelNet (espe- has been used as a knowledge-base for the task
cially the links between synsets and Wikipedia of named entity disambiguation (NED), but also
pages) were built semi-automatically. We chose to adapted for WSD. Graph-based approaches for
manually map synsets of plWordNet by lexicogra- computing semantic relatedness and disambigua-
phers instead of use the semi-automatic mappings tion can be improved by using the semantic in-
provided by BabelNet to have more control over formation contained in Wikipedia and expand the
the accuracy of our results. underlying knowledge base e.g. a wordnet. For
this work we decided to add the links extracted
3.2 Presented Knowledge Base from Wikipedia by using the mapping of lexical
For the work presented here, two knowledge senses to equivalent Wikipedia articles. For ev-
graphs were built on the basis of the two largest ery mapped synset we added new semantic links
wordnets, namely plWordNet 3.2 (Maziarz et al., by analysing the content of the page and extract-
2016) for Polish, and Princeton WordNet 3.1 (Fell- ing monosemous words. The lexical senses of
baum, 1998) for English. They are mutually monosemous words were linked to mapped synset.
mapped on each other as a result of the laborious 2
http://wordnetcode.princeton.edu/
work of bilingual lexicographers. glosstag.shtml
519
4 Methods tribution P instead of performing a random jump
(usually c = 0.85). The π = π Pe represents a vec-
The main drawback of PageRank-based methods tor of PageRank scores for nodes of a given graph.
is that to compute the score of a sense given the We can also show, that the final PageRank dis-
context they have to compute the features with tribution π can be then defined as shown in (3),
respect to the global structure of a given knowl- which directly follows from (2).
edge graph. This means the PageRank values
have to be computed for every single node in the 1−cX k k
∞
graph. To avoid this issue we use only the local π= c P (3)
n
approximation of PageRank as it was presented in k=0
(Avrachenkov et al., 2007). One of the first attempts to compute a local ap-
proximation of PageRank scores, namely (Fogaras
4.1 PageRank Based Methods
et al., 2005), was based on the property presented
The computational complexity of Personalised in (3) which leads us to Monte Carlo methods. The
PageRank (PPR) is still a limiting factor for fast, authors in (Avrachenkov et al., 2007) proved that
real time WSD of large textual data. The naive al- we can easily approximate PageRank values using
gorithm for computing PageRank values requires random walks with restarts. Let the random walk
to iterate over the entire graph. The most popular start from a randomly chosen page and terminate
approach to compute PPR scores is the Power Iter- with the probability (1 − c). The random walk
ation method (Berkhin, 2005), where the score is runs over the graph and makes the transitions ac-
defined in a recursive way taking into account the cording to transition matrix P with probability c.
global information. Recently the methods of local Let πj be the final PageRank score for node j in
approximation of PageRank scores have received the graph G. So, the initial PageRank formula (2)
a lot of attention, especially in the case of dynamic can be replaced with its rough estimate. We can
graphs where the structure of the graph changes in show that using the properties of equation (3) we
time. can also transform the Personalised PageRank for-
Let G = (V, E) be a graph where V is a set mula and compute a rough estimate of its scores
of nodes and E represents a set of edges. The (Avrachenkov et al., 2010) with equation (4).
overall number of nodes and edges is N and M ,
m
respectively. We can define the adjacency ma- 1 X
πˆj (s) = (1 − c) Nj (s, r) (4)
trix AN ×N of the graph G as AN ×N = [aij ], m
r=1
i ∈ {1, 2, ..., N }, j ∈ {1, 2, ..., N }, where the val-
ues aij are representing links existing in the graph, A seed of initial nodes s is used to perform ran-
where aij = 1 if there is an edge pointing from dom walks over the graph, where c is the prob-
node vi to vj , otherwise aij = 0. A Markov tran- ability that the random walks terminates, m rep-
sition matrix P can be defined as PN ×N = [pij ] resents the overall number of required random
with pij values being normalised by the number of walks, Nj (s, r) is the number of random walks
outgoing links (from node vi ) pij = d1i if aij = 1, ending for node j, starting from seed node s in
otherwise pij = 0. The graph might also contain r-th random walk. The nodes representing a seed
dangling nodes without any outgoing links. To are usually randomly sampled from the graph.
handle these cases pij values for dangling nodes This leads us to the following algorithm of PPR
are usually replaced by a constant N1 , which means computation:
adding a link to every node in the graph. The static
1. Simulate m runs of the random walks initi-
PageRank can be interpreted then as a stationary
ated at a node s.
distribution π of a Markov chain with final transi-
tion matrix Pe (Google’s matrix): 2. Evaluate πj as a fraction of m random walks
which end at node j ∈ {1, 2, ..., n}.
1
Pe = cP + (1 − c) R (2)
N The Monte Carlo approaches are based on ran-
The matrix R consists of entries being equal to dom sampling, thus, we may obtain slightly dif-
one. The value c ∈ (0, 1) is the probability that ferent results for every run, especially when the
the random walk follows a link according to dis- number of the required iterations is too small to
520
obtain faster progress of convergence giving a nar- Geom(p) starting from each seed node s. The
row confidence interval for estimated parameter. seed consists of the nodes (synsets) representing
The unnecessary randomness can be avoided by all of available senses for the words from a given
initiating the random walk from each node in the disambiguation context. The importance score γj
same way, rather than jumping to random nodes for a given node j is the total number of visits
over the graph. The accuracy of our estimate can to node j divided by the total number of visited
be also improved by using variance reduction tech- nodes.
niques e.g. by using Common Random Numbers
approach (Clark, 1990). 1 X
m
γˆj (s) = (1 − c)[Nj (s, r) + Rj (s, r)] (5)
m
4.2 Our Method r=1
PageRank value using random walk methods over γˆj (s) = (1 − c)[Nj (s, r) + Rj (s, r)]
m s∈s
r=1
semantic graphs, thus, we adopted the idea pre- (6)
sented in (Fogaras et al., 2005) and (Pershina et al., The parameter m denotes the overall number of
2015). The methods are usually based on random performed random walks. The expected length of
walk algorithm where the main assumption is that a single random walk is expressed as a geometric
the local properties of senses are sufficient to ac- distribution Geom(c), usually initialized with the
curately disambiguate the texts. This assumption value of parameter c (eq. (4)). To reduce the ran-
allows us to reduce processing time of the algo- domness effects we generate the lengths of walks
rithm when working with large lexical knowledge only once, for all of our seed nodes, which is sim-
bases. ilar to the variance reduction with Common Ran-
The underlying lexical knowledge base is a cru- dom Numbers (Clark, 1990).
cial element of a WSD method and it has a great Rj (s, r) represents the overall number of ran-
impact on its performance. The algorithms and dom resets, where the decision of a random jump
their properties presented in Sec. 4.1 were an in- to the starting node s is distributed according to
spiration for further work on disambiguation algo- Bernoulli(p). The parameter p is dependent on
rithms using large semantic networks. the smoothed similarity score between usage ex-
The properties of personalised PageRank al- ample and a context.
gorithm can be a limiting factor for WSD per- We can compute a more accurate score estimate
formance. The main issue is that in some spe- by using the recursive property of PageRank for-
cific cases a large node degree does not indicate mula and averaging the scores of the neighbours.
a high significance, especially when the underly- The recomputed final score for a single node j can
ing knowledge base is a heterogenous graph con- be defined as:
structed from different semantic resources (on-
tologies, dictionaries, wordnets). Adding or re- 1 X
moving certain links can change the PageRank γˆj (s) = (1 − c) γˆv (s) (7)
|O(j)|
scores of the target nodes. The main problem is v∈O(j)
that not all of the links are correct, and to pro- In (7), the set O(j) represents the neighbour-
tect PageRank scores against incorrect links we hood of a single personalisation node j and γˆv (s)
can manipulate the importance of the links by us- is a scoring function computed for neighbour v in
ing a heuristic weighting schema. We can also O(j).
manipulate our seed to make the PPR algorithm In (Agirre and Soroa, 2009) and (Agirre et al.,
more robust to textual noise. Since we are inter- 2018) the authors proved that the sense frequency
ested only in the importance of our personalisation is a strong signal for accurate WSD. To make our
nodes (representing senses of words for a given methods comparable we decided to use the same
context), the initial seed for PPR computation is source of sense frequency, namely SemCor cor-
limited to a set of personalised nodes (usually a set pus. Following the idea presented in (Agirre et al.,
of senses representing the words in a small textual 2018), the final score of a synset is computed as
window). a linear combination of its normalised sense fre-
Simulate m random walks of length L ∼ quency and graph-based scores.
521
5 Evaluation dataset consisted of 74 words in total, including 45
nouns and 29 verbs. The overall number of anno-
The proposed method incorporates several expan- tated word occurrences was 5,148 with 3,219 noun
sions to improve the overall performance of WSD. occurrences and 1,929 verb ones. Since there is a
For the evaluation of WSD in English we utilised small inconsistency between senses used in the an-
available semantic resources: i) manually disam- notation and produced by WSD methods, i.e. com-
biguated glosses, ii) synsets linked to Wikipedia, ing from plWordNet 3.2, we decided to exclude
iii) the knowledge extracted from SemCor corpus. from it 473 word occurrences for which the appro-
In the case of Polish language the available se- priate sense does not exist any more in our WSD
mantic resources are limited. The glosses and model.
usage examples for Polish senses were not dis-
ambiguated, thus, the synsets were linked only 5.2 Experimental Setting
with the senses of monosemous words appearing
To accomplish a satisfactory convergence and ob-
in their glosses or usage examples. In the case
tain a small variance of final accuracy we adapted
of Wikipedia-based expansion we used a different
a following set of parameters for our experimental
mapping – the Polish synsets were partially linked
part: the transition probability c = 0.3, the over-
to Wikipedia in a manual process (around 50,000
all number of random walks per node rw_iter =
of synsets were mapped to Polish Wikipedia). We
1000, the importance of sense frequencies on the
also translated the links by using interlingual syn-
final score α = 0.5, and (1−α) for the importance
onymy links between Polish and English (enWord-
of random walk-based scores. For a Polish dataset
Net) parts of plWordNet.
we did not use sense frequencies since there are no
5.1 Datasets sense-tagged corpora available to compute the fre-
quencies. The resultant performance (tables tab. 1
In the experimental part we evaluate our meth-
and tab. 2) was computed 10 times and averaged.
ods on the English dataset described in (Raganato
et al., 2017) and the Polish dataset presented in 6 Results
(K˛edzia et al., 2015). The former dataset con-
sists of the five standard English texts prepared for Table 1. presents the final peformance for En-
Senseval and SemEval competitions for all-words glish dataset. The proposed method and intro-
WSD task. A sense inventory for the gold stan- duced knowledge base expansions mostly outper-
dard annotations was built on a basis of Princeton formed a very strong most frequent sense (MSF)
WordNet 3.0 which makes it approximately com- baseline. The method achieved the results being
patible with our knowledge-base, i.e. built on ex- on the comparable level with other weakly su-
tended Princeton WordNet 3.1 in the case of En- pervised baseline methods, namely UKB (Agirre
glish and plWordNet 3.2 (mapped to WordNet 3.1 et al., 2018), Babelfy (Moro et al., 2014), and
and via it to SUMO) in the case of Polish. As we WoSeDon (K˛edzia et al., 2015). Table 2. shows
use WordNet 3.1 some small discrepancies can in- the average disambiguation time per context. As
fluence the results of the comparison with the test it was expected, the Monte Carlo approach pro-
datasets, but they should not have significant im- vides better time efficiency (PPR vs PPRMC)
pact on the outcome of the comparison. and a comparable performance to power iteration
Regarding the dataset for Polish only partially method. The best results were achieved by mix-
sense-annotated corpora exist. Składnica is a man- ing all available sources of semantic knowledge:
ually annotated dependency treebank with 13,035 the links extracted from disambiguated glosses,
sentences written Polish. The updated version of Wikipedia pages, SemCor texts and the links pro-
the semantic annotation in Składnica was based on vided by SUMO ontology. The same tendency
plWordNet 3.2. was observed for Polish dataset (table 3) – the
The Polish Corpus of Wrocław University of best performance was noted for a mixed setting.
Technology (henceforth KPWr) consists of 1,127 The performance obtained for this dataset was also
documents manually annotated with the plWord- on a comparable level with the approaches based
Net 2.1 senses. The annotation was limited to on power iteration method proposed in (K˛edzia
a pre-selected set of words representing different et al., 2015). PPRMC-1 uses a knowledge graph
cases of homonymy and polysemy. The original of synsets only. PPRMC-2 expands the model of
522
Method Sens-2 Sens-3 Sem-07 Sem-13 Sem-15
MFS 66.80 66.20 55.20 63.00 67.80
Babelfy 67.00 63.50 51.60 66.40 70.30
UKB-nf 61.30 54.90 42.20 60.90 62.90
UKB-sf 67.50 66.40 54.10 64.00 67.80
UKB-nf-w2w 64.20 54.80 40.00 64.50 64.50
UKB-sf-w2w 68.80 66.10 53.00 68.80 70.30
PPRMC-1 66.26 64.28 54.06 65.08 67.12
PPRMC-2 66.35 65.13 55.60 65.56 66.63
PPRMC-3 66.47 65.94 56.04 65.26 67.71
PPRMC-4 66.78 66.28 56.48 65.90 68.10
Table 1: Averaged F1-scores of PPRMC for different knowledge base expansions: PPRMC-1: graph of
synsets only, PPRMC-2: graph of synsets extended with links extracted from manually disambiguated
glosses, PPRMC-3: PPRMC-2 extended with links extracted from SemCor and Wikipedia, PPRMC-4:
PPRMC-3 with additional links extracted from SUMO ontology.
Table 2: Average disambiguation time [s] per context for SemEval’15 dataset with respect to the size of
underlying knowledge base. PPR settings: damping_factor=0.85, max_iterations=25, PPRMC settings:
c=0.3, rw_count=1000. The disambiguation context was limited to a small window of three sentences.
Method Skład.-N Skład.-V KPWr-N KPWr-V knowledge base used to define word senses and
PPRMC-1 63.19 44.75 52.92 33.42 provide a basis for their disambiguation. How-
PPRMC-2 64.27 46.01 53.24 33.73
PPRMC-3 64.88 46.22 53.31 33.66 ever, very good results are obtained only when
PPRMC-4 65.28 46.51 53.66 33.09 a rich and large knowledge base is utilised. In
WoSeDon 63.92 46.43 53.61 33.71 such a case, PPR-based WSD methods become
WoSeDon 64.85 47.29 53.80 34.08 slow that limits their applicability. We have shown
WoSeDon 65.27 47.55 54.02 34.00
WoSeDon 66.18 48.74 54.90 33.89
that an estimation of the PPR algorithm on the ba-
sis of the Monte Carlo scheme can preserve most
Table 3: Averaged precision of PPRMC for Polish of the quality of the method while gaining a lot
datasets computed for nouns (N) and verbs (V) in terms of the efficiency of computation. We
separately. A comparison with knowledge-based proposed a WSD tool3 that achieves the perfor-
solution presented in (K˛edzia et al., 2015) adapted mance comparable to the state-of-the-art among
to the structure of plWordNet 3.2. the weakly supervised methods, but it is 4-5 times
faster. In addition, the efficiency of the proposed
tool does not deteriorate so quickly with the in-
PPRMC-1 with links to monosemous words ex- creasing complexity of the knowledge base. The
tracted from glosses – in case of Polish glosses, proposed method also offers an opportunity to bal-
or to disambiguated senses – in case of English ance between accuracy and processing speed by
glosses. PPRMC-3 adds the links to monosemous selecting the number of random walks.
words extracted from Wikipedia, and PPRMC-4
introduces additional semantic links from SUMO
ontology.
7 Conclusions
Weakly supervised Word Sense Disambiguation
(WSD) methods express very good coverage that
is only limited by the coverage of the underlying 3
gitlab.clarin-pl.eu/ajanz/wsd-mc
523
References Sanda M. Harabagiu, George A. Miller, and Dan I.
Moldovan. 1999. Wordnet 2 - a morphologically
Eneko Agirre, Oier Lopez de Lacalle, and Aitor Soroa. and semantically enhanced resource. In Proceed-
2014. Random walks for Knowledge-Based Word ings of SIGLEX 1999.
Sense Disambiguation. Computational Linguistics
40(1):57–84. Paweł K˛edzia and Maciej Piasecki. 2014. Ruled-based,
interlingual motivated mapping of plwordnet onto
Eneko Agirre, Oier López de Lacalle, and Aitor sumo ontology. In Nicoletta Calzolari et al., edi-
Soroa. 2018. The risk of sub-optimal use of open tor, Proc. Ninth International Conference on Lan-
source nlp software: Ukb is inadvertently state-of- guage Resources and Evaluation (LREC’14). Euro-
the-art in knowledge-based wsd. arXiv preprint pean Language Resources Association.
arXiv:1805.04277 .
Paweł K˛edzia, Maciej Piasecki, Jan Kocoń, and
Eneko Agirre and Aitor Soroa. 2009. Personalizing Agnieszka Indyka-Piasecka. 2014. Distribu-
pagerank for word sense disambiguation. In Pro- tionally extended network-based Word Sense
ceedings of the 12th Conference of the European Disambiguation in semantic clustering of Pol-
Chapter of the Association for Computational Lin- ish texts. IERI Procedia 10(Complete):38–44.
guistics. Association for Computational Linguistics, https://doi.org/10.1016/j.ieri.2014.09.073.
pages 33–41.
Paweł K˛edzia, Maciej Piasecki, and Marlena Orlińska.
Konstantin Avrachenkov, Nelly Litvak, Danil Ne-
2015. Word sense disambiguation based on large
mirovsky, and Natalia Osipova. 2007. Monte carlo
scale polish clarin heterogeneous lexical resources.
methods in PageRank computation: When one itera-
Cognitive Studies (15).
tion is sufficient. SIAM Journal on Numerical Anal-
ysis 45(2):890–904.
Karin Kippera, Anna Korhonen, Neville Ryant, and
Konstantin Avrachenkov, Nelly Litvak, Danil A Ne- Martha Palmer. 2006. Extensive classifications of
mirovsky, Elena Smirnova, and Marina Sokol. 2010. english verbs. In Proceedings of the 12th EURALEX
Monte carlo methods for top-k personalized pager- International Congress.
ank lists and name disambiguation. arXiv preprint
arXiv:1008.3775 . Marek Maziarz, Maciej Piasecki, Ewa Rudnicka, Stan
Szpakowicz, and Paweł K˛edzia. 2016. plword-
Dominik Baś, Bartosz Broda, and Maciej Piasecki. net 3.0 – a comprehensive lexical-semantic re-
2008. Towards word sense disambiguation source. In Nicoletta Calzolari, Yuji Matsumoto,
of Polish. In Proceedings of the Interna- and Rashmi Prasad, editors, COLING 2016, 26th
tional Multiconference on Computer Science International Conference on Computational Lin-
and Information Technology — 3rd Interna- guistics, Proceedings of the Conference: Techni-
tional Symposium Advances in Artificial Intelli- cal Papers, December 11-16, 2016, Osaka, Japan.
gence and Applications (AAIA’08). pages 65–71. ACL, ACL, pages 2259–2268. ACL Anthology.
http://www.proceedings2008.imcsit.org/pliks/162.pdf. http://www.aclweb.org/anthology/C16-1213.
Pavel Berkhin. 2005. A survey on pagerank comput- Rada Mihalcea, Paul Tarau, and Elizabeth Figa. 2004.
ing. Internet Mathematics 2(1):73–120. PageRank on semantic networks, with application
to Word Sense Disambiguation. In Proceedings
Francis Bond and Ryan Foster. 2013. Linking and ex- of the 20th International Conference on Compu-
tending an Open Multilingual Wordnet. In 51st An- tational Linguistics. Association for Computational
nual Meeting of the Association for Computational Linguistics, Stroudsburg, PA, USA, COLING ’04.
Linguistics. ACL, pages 1352–1362. https://doi.org/10.3115/1220355.1220517.
Bartosz Broda and Maciej Piasecki. 2011. Evaluating Andrea Moro, Alessandro Raganato, and Roberto Nav-
LexCSD in a large scale experiment. Control and igli. 2014. Entity linking meets word sense disam-
Cybernetics 40(2):419–436. biguation: a unified approach. Transactions of the
Association for Computational Linguistics 2:231–
Gordon M Clark. 1990. Use of common random num- 244.
bers in comparing alternatives. In Proceedings of
the 22nd conference on Winter simulation. IEEE Roberto Navigli and Simone Paolo Ponzetto. 2012.
Press, pages 367–371. BabelNet: The automatic construction, evaluation
and application of a wide-coverage multilingual se-
Christiane Fellbaum, editor. 1998. WordNet – An Elec- mantic network. Artificial Intelligence 193:217–
tronic Lexical Database. The MIT Press. 250.
Dániel Fogaras, Balázs Rácz, Károly Csalogány, and Lawrence Page, Sergey Brin, Rajeev Motwani, and
Tamás Sarlós. 2005. Towards scaling fully person- Terry Winograd. 1998. The PageRank citation rank-
alized pagerank: Algorithms, lower bounds, and ex- ing: Bringing order to the web. Technical report,
periments. Internet Mathematics 2(3):333–358. Stanford Digital Libraries Working Paper.
524
Adam Pease. 2011. Ontology - A Practical Guide. Ar-
ticulate Software Press.
Maria Pershina, Yifan He, and Ralph Grishman. 2015.
Personalized page rank for named entity disam-
biguation. In Proceedings of the 2015 Conference of
the North American Chapter of the Association for
Computational Linguistics: Human Language Tech-
nologies. pages 238–243.
Alessandro Raganato, Jose Camacho-Collados, and
Roberto Navigli. 2017. Word sense disambiguation:
A unified evaluation framework and empirical com-
parison. In Proceedings of the 15th Conference of
the European Chapter of the Association for Compu-
tational Linguistics: Volume 1, Long Papers. pages
99–110.
Josef Ruppenhofer, Michael Ellsworth, Miriam R. L
Petruck, Christopher R. Johnson, Collin F. Baker,
and Jan Scheffczyk. 2016. FrameNet II: Extended
Theory and Practice.
Mark Stevenson, Eneko Agirre, and Aitor Soroa. 2012.
Exploiting domain information for Word Sense
Disambiguation of medical documents. JAMIA
19(2):235–240.
525
Classification of Micro-Texts Using Sub-Word Embeddings
526
Proceedings of Recent Advances in Natural Language Processing, pages 526–533,
Varna, Bulgaria, Sep 2–4 2019.
Figure 1: Overview of the proposed system
One of the first deep learning approaches by (Lewis et al., 2004), Enron dataset (Klimt and
Rhodes (2015) used word embeddings and a con- Yang, 2004), and Twitter dataset (Yilu et al.,
volutional neural network. Skip-gram method 2016). The RCVI is used to train the embed-
with negative sampling is used for word vector- dings which in turn generated the high-level fea-
ization from the Google news dataset while using tures from the other two corpora by concatenat-
random initialization for unseen words. Further- ing the vector of means and standard deviations.
more, the convolutional neural network was based They then used the feed forward neural network to
on a similar model (Collobert et al., 2011) for lan- identify the authors.Therefor, in this work, we also
guage processing, where the activation function is used the same five authors (Ashley Nunn75, brad-
changed to rectified linear units and introducing shaw1984, shawnevans81 , terrymarvin63, and
dropouts. Finally, it demonstrated that the system WhieRose65) from the twitter dataset that they
performs well on longer text sequences. used to test our proposed system.
Shrestha et al. (2017) makes use of a similar ar-
chitecture; a convolutional neural network (CNN) 3 Methodology
using character embeddings instead of word em- In this section, we describe our proposed approach
beddings for short texts. The CNN model takes and the features we used to train our system. Fig-
character unigram or bigram as an input that ure 1 shows an overview of our system in which
passes through the character embedding layer be- we worked on two datasets: one by Schwartz et al.
fore feeding it to the convolutional layer. Similar (2013) and the other one by Yilu et al. (2016). Sec-
datasets and evaluation criteria as Schwartz et al. tion 3.1 explains more about the data.
(2013) are used to determine the performance of We carried our pre-processing on the data be-
the system. This approach increases the overall fore preforming feature extraction. There are a to-
accuracy to approximately 76%. Even though the tal of four features that are obtained from the text
results are better than all previous researches there namely word n-grams, character n-grams, flexi-
is a scope for further improvement. ble patterns and word embeddings. An n-gram
A more recent study on this field (Phan and (section 3.2) is a sequence of n words where n is
Zincir-Heywood, 2018) used word embeddings a positive integer. For example, in the phrase -
and the neural network to identify the authors of ”This is a sentence” -, a word unigram would be
short text messages. They used the three dif- - ”This”, ”is”, ”a” - and a word bigram would be
ferent datasets namely Reuters Corpora (RCVI) ”this is”, ”is a”, ”a sentence” - and so on. On
527
the other hand, character n-grams are similar to the MLP, we concatenated sets of two adjacent
word n-grams except, they are a sequence of char- records. For example, the first text is combined
acters. For instance, if we use the previous ex- with the second, the third with the fourth and so
ample, ”This is a sentence”, a character unigram forth. Though the initial number of records re-
would be - ”T”, ”h”, ”i”, ”s” - and a character mains the same, combining the original texts en-
bigram would be - ”th”, ”hi”, ”is”. ables us to have a larger sequence, which achieves
In section 3.3 we define Flexible patterns, ex- better accuracy even with the earlier approaches of
plain the existing method and our new approach to feature extraction (Schwartz et al., 2013; Shrestha
create them. The idea behind flexible patterns is et al., 2017). The larger sequence also helps us
that some users tend to use the same sequence of to achieve more meaningful patterns which is not
words in their writing style and only change a few otherwise possible.
keywords called content words (CW). For exam- We also applied the above approach on a dif-
ple, the flexible pattern of the following phrases; ferent dataset (Yilu et al., 2016) used in Phan and
”I read the paper today” and ”I drove the car yes- Zincir-Heywood (2018). We used the same 5 au-
terday” is ”I CW the CW CW”. The words read, thors (users) as they did, with 2000 tweets each.
paper, today, drove, car and yesterday are replaced In that paper (Phan and Zincir-Heywood, 2018),
by the word CW based on the pre-defined condi- the names of the tagged users in a tweet were not
tion. Therefore, masking some words, would cre- masked. However, we masked them, i.e. whenever
ate a pattern and separate the texts of some users we encounter a tagged user - @user - we change
from other users. We modified the existing ap- it to a specific word preventing our system from
proach (Schwartz et al., 2013) to make it suitable overfitting.
for smaller datasets and also to make it easier to
implement. 3.2 Word and Character N-Grams
Next, we talked about the word embeddings Both word and character n-grams were extracted
(Mikolov et al., 2013) feature set in section 3.4. from the datasets used. For the value of n in n-
We used the skip-gram technique to create the grams and the minimum occurrence of each n-
300-dimensional vector representation of our data gram, we used the same parameters as used by
and used that to create the embeddings by com- Schwartz et al. (2013) for comparison purposes.
bining them using the weights obtained by using We also took into consideration the maximum oc-
TF-IDF. currence of the n-grams. For example, a pattern
Later we analyzed (section 3.5) all the models including prepositions (as shown in 1 below) or a
we used and the reason for choosing a Multi-Layer masked username (as shown in 2 below) are very
Perceptron (MLP) architecture (Gillian, 2014). common in tweets.
We also explained the architecture of our MLP 1. for a
model and the parameters we used to further op- 2. <User> I
timize. Lastly, in section 4 we will compare the
This is even more common in character n-
results of our approach to both datasets.
grams. For this reason, we kept the upper limit of
the n-grams to 0.9, which means we do not con-
3.1 Datasets
sider any word or character n-grams that appear in
To compare our results with the previously men- more than 90 percent of the documents. This fur-
tioned approaches, we used the same dataset that ther helps us identify the unique writing style of
Schwartz et al. (2013) used. The dataset contains an author.
a total of 7000 authors, out of which we selected We restricted each feature to a maximum of
50 authors at random, where each author has 1000 50,000 in our experiments where each author had
tweets. We masked the username (@user), num- 1000 tweets. To assign weights to the n-grams,
bers, links, date and time from the texts to min- we used the term frequency-inverse document fre-
imize the bias and noise in the data. We also quency (TF-IDF) weighted scheme.
changed all of the letter characters to lowercase TF-IDF (Manning et al., 2008) calculates the
and employ word stemming to reduce the size of importance of the word based on how frequently
the vocabulary used. it appeared in a text, which was then balanced by
Before extracting features and feeding data into the frequency of the word in the entire corpus.
528
Figure 2: Skip-Gram word embeddings
In a text document d, the frequency of term t is: lary. This method is based on the various experi-
tft ,d = ft ,d . This shows the total number of times ments we carried out to choose the optimum num-
(frequency count) the term occurs in the docu- ber.
ment, whereas the inverse document frequency is:
N is the total number CW 6 2 × log10 n (2)
idf (t ,d ) = log , where N
t∈D
of documents D in the corpus. Moreover, we em- After replacing all the CWs in the corpus, we then
ployed sub-linear scaling to the TF-IDF by taking used the same approach as we did for word n-
the log of the term frequency, Eq.1. grams. Then we applied the TF-IDF weighted
scheme to those flexible n-grams before inputting
( them into the model.
1 + logtf td , if tftd = 1
wft ,d = (1) 3.4 TF-IDF Weighted Word Embeddings
0, otherwise
Word embedding (Mikolov et al., 2013) is a se-
3.3 Flexible Patterns mantic parsing technique used to create the vec-
Flexible patterns are a branch of word n-grams, tor representation of a text in a smaller dimen-
where each word is either a high-frequency word sional space compared to the classic Bag of words
or a content word and some words can be both approach (Zhang et al., 2010). The idea be-
(Schwartz et al., 2013). For a corpus size s if hind word embedding is that semantically similar
a word appears more than 10−4 × s times it is words should be close to each other. That is, in an
a High-Frequency Word (HFW) and if it appears n-dimensional space, the angle between the simi-
less than 10−3 × s times it is a Common Word lar words should be close to zero. There are two
(CW). Also, the previous method takes into con- types of methods to achieve this; namely, Continu-
sideration that flexible patterns start/end with an ous Bag of Words (CBOW) and Skip-Gram meth-
HFW and there can be no consecutive HFWs. ods (Mikolov et al., 2013). While CBOW predicts
The problem with this approach is that it does the target words by taking the context word as in-
not work with a smaller corpus. For a corpus put, Skip-gram predicts the probability of the con-
where the value of s is 1000, no word is a CW. text words using the target words.
Thus, to overcome this limitation, Eq.2 is used to We used the word embedding approach from
calculate the CW for a bigger corpus. Therefore Bojanowski et al. (2017), which is based on the
in a corpus with a vocabulary size n, the CW is assumption that tweets contain several words, in-
calculated as the common log of the threshold for cluding, but not limited to, hashtags that are rare
CW is selected as twice of log of s which are total and sparsely occur in a corpus. To address this
number of words in the training dataset vocabu- sparsity issue, each word is represented by the sum
529
of the vector representations of its n-grams. Hav-
ing a dictionary as size D and a word w belongs
to this dictionary having Dw ⊂ {1, ..., D}, the set
of n-grams appearing in w. Associating a vector
representation of zd to each n-gram d, the scoring
function for w can be represented by Eq.3:
X
s(w, c) = zdT vc (3)
d∈Dw
530
authors. Figure 4 presents the overview of the pro-
posed MLP model. The number of nodes in the
input layer depend on the size of the input feature.
None represents that the dimension is variable, and
in this case, it is 86121. Then, there is a dense
layer, which is a fully connected layer, where each
input node is connected to each output node. In
the proposed model, there are two Dense layers.
One is a hidden layer of 1000 nodes and the other
is the output layer of 50 nodes, which corresponds
to 50 authors. This can be changed depending on
the number of authors (output classes). Inbetween
Figure 5: Epochs vs Loss for 50 epochs
the two dense layers, there is a dropout layer for
regularization.
Figure 4: Overview of the proposed model using a Multi-Layer Perceptron as TFIDF Emb. Flex Joining rec. CNN-C SCH Char LSTM-2
the classifier with one hidden layer and a dropout layer
.852 .829 .81 .761 .712 .703 .645
The input layer of the MLP model depends on Table 3: Accuracy for 50 users with 1000 tweets each
the number of features, which increases as the
number of tweets increases to train the model.
There is only one hidden layer with 1000 nodes. applied all three feature extraction techniques us-
It should be noted, that increasing the hidden lay- ing the MLP classifier and observed the improve-
ers did not improve the validation accuracy in our ments compared to the previous research results
experiments. The Tanh activation function (Weis- (see Section 2). To this end, we first combined
stein, 2002) is used for the hidden layer. The last the subsequent tweets and applied the approach
layer consists of the same number of nodes as the presented by Schwartz et al. (2013), then we im-
number of authors where the SoftMax activation proved that method to calculate flexible patterns
function (Nwankpa et al., 2018) is used. Further- and their effect on the accuracy of the system. Last
more, we have a 30% dropout after the hidden but not the least, we built weighted TF-IDF em-
layer to prevent overfitting. In the proposed sys- beddings and combined them with the improved
tem, ADAM (Kingma and Ba, 2014) is adapted flexible patterns to form the combined set of fea-
as the optimization algorithm with a learning rate tures as input into the proposed MLP model. Ten-
of 0.001. We divided the data into batches of fold cross-validation approach was used to eval-
64 and trained our model for a maximum of 40 uate the performance of the proposed system. We
epochs. Figure 5 and Figure 6 show the number evaluated the proposed system on the same dataset
of epochs vs loss, and the number of epochs vs as used by Schwartz et al. (2013); Shrestha et al.
accuracy graphs for training and validation data, (2017), where the dataset consisted of 50 authors,
respectively. These results show that both the loss each having 1000 tweets.
and accuracy are optimal at around 40 epochs. Columns 1-3, in Table 3, shows the results of
In the following experiments, we incrementally all three feature extraction techniques used with
531
# TFIDF Emb. Mod. flex Joining rec. CNN-C SCH Char LSTM-2
500 .791 .76 .748 .724 .672 .655 .597
200 .730 .694 .679 .665 .614 .585 .528
100 .648 .619 .608 .617 .565 .517 .438
50 .589 .563 .53 .562 .505 .466 .364
the MLP, and columns 4-7 represents the results vious approaches. Then applying the flexible pat-
from the previous works on the same dataset. Our terns, further improved the accuracy by approxi-
results are mutually inclusive, and the results are mately 2%. Finally, implementing the weighted
built upon combining all the feature extraction TF-IDF word embedding, and combining it with
techniques. For example, we used the weighted all the other features increases the accuracy of the
TF-IDF embeddings in combination with the flex- proposed system to approximately 85%, Table 3.
ible patterns. In Table 4, we also compared the proposed
CNN-C, shown in Table 3, represents the best system’s accuracy to other approaches as we re-
result obtained by Shrestha et al. (2017) where a duced the number of tweets from 500 to 50 for
convolutional neural network architecture is pro- each author (50). After reducing the number of
posed using character n-grams, specifically uni- tweets, the proposed system still outperformed
grams and bigrams as input. The convolutional all the other previous approaches and performed
model used was a three-level architecture with the well even when the dataset became as small as 50
input layer as the character embedding layer, a tweets per author.
convolutional module, and finally, a dense layer
with a Softmax activation function for classifica- 5 Conclusion and Future Work
tion. The unigram model performed well on the
In this paper, we explored a feature extraction
smaller dataset. Alternatively, the bigram model
technique that was based on a word embedding
had better accuracy on the bigger dataset.
model weighted by TF-IDF. We also worked on
SCH, shown in Table 3, represents the best re-
modifying and improving the existing implemen-
sult obtained by Schwartz et al. (2013). They used
tation of flexible patterns and proposed a neural
a linear SVM for classification and their model
network architecture that makes use of a combina-
was a combination of word and character n-grams
tion of these features to perform authorship attri-
along with a new feature set called flexible pat-
bution. Our model outperformed all the existing
terns (see section 3.3), which is modified and used
systems based on similar testing criteria and using
in our system as well.
the same datasets. Since we trained our embed-
Char, shown in Table 3, represents one of the
dings from scratch our system performs equally
systems used by Shrestha et al. (2017) in which
well, irrespective of the language.
they compared the performance of their system
With the success of word embeddings, we look
based on the earlier character n-gram approaches
forward to working upon new word embedding
(Layton et al., 2010; Schwartz et al., 2013) and
techniques such as Elmo (Peters et al., 2018) and
proposed a logistic regression model that em-
Bert (Devlin et al., 2018), which are based on the
ployed character n-grams of sizes two to four.
context of a word in a corpus. Alongside that, we
LSTM-2, shown in Table 3, represents the state-
are also interested to improve our neural network
of-the-art LSTM model based on the success of
architecture using transfer learning models such as
previous implementations (Tai et al., 2015; Tang
ULMFit (Howard and Ruder, 2018).
et al., 2015). This model was also used with
bigrams as input to evaluate the performance of 6 Acknowledgments
those systems, with respect to other models on the
same dataset. This research is supported by the Natural Science
Introducing the concatenation of the consecu- and Engineering Research Council of Canada.
tive records enables the accuracy of the proposed This research is conducted as part of the Dalhousie
system to be improved by 5% compared to the pre- NIMS Lab at: https://projects.cs.dal.ca/projectx/.
532
References Chigozie Nwankpa, Winifred Ijomah, Anthony Gacha-
gan, and Stephen Marshall. 2018. Activation func-
Piotr Bojanowski, Edouard Grave, Armand Joulin, and tions: Comparison of trends in practice and research
Tomas Mikolov. 2017. Enriching word vectors with for deep learning. arXiv preprint arXiv:1811.03378
subword information. Transactions of the Associa- .
tion for Computational Linguistics 5:135–146.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
Ronan Collobert, Jason Weston, Léon Bottou, Michael B. Thirion, O. Grisel, M. Blondel, P. Pretten-
Karlen, Koray Kavukcuoglu, and Pavel Kuksa. hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-
2011. Natural language processing (almost) from sos, D. Cournapeau, M. Brucher, M. Perrot, and
scratch. Journal of machine learning research E. Duchesnay. 2011. Scikit-learn: Machine Learn-
12(Aug):2493–2537. ing in Python . Journal of Machine Learning Re-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and search 12:2825–2830.
Kristina Toutanova. 2018. Bert: Pre-training of deep Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt
bidirectional transformers for language understand- Gardner, Christopher Clark, Kenton Lee, and Luke
ing. arXiv preprint arXiv:1810.04805 . Zettlemoyer. 2018. Deep contextualized word rep-
resentations. arXiv preprint arXiv:1802.05365 .
Nick Gillian. 2014. Mlp nickgillianwiki.
http://www.nickgillian.com/wiki/ Tien D Phan and Nur Zincir-Heywood. 2018. User
pmwiki.php/GRT/MLP. (Accessed on identification via neural network based language
04/25/2019). models. International Journal of Network Manage-
ment page https://doi.org/10.1002/nem.2049.
Marti A Hearst. 1992. Automatic acquisition of hy-
ponyms from large text corpora. In Proceedings of Tie-Yun Qian, Bing Liu, Qing Li, and Jianfeng Si.
the 14th conference on Computational linguistics- 2015. Review authorship attribution in a similarity
Volume 2. Association for Computational Linguis- space. Journal of Computer Science and Technology
tics, pages 539–545. 30(1):200–213.
Jeremy Howard and Sebastian Ruder. 2018. Universal Dylan Rhodes. 2015. Author attribution with cnns
language model fine-tuning for text classification. cs224.
arXiv preprint arXiv:1801.06146 .
Roy Schwartz, Oren Tsur, Ari Rappoport, and Moshe
Diederik P Kingma and Jimmy Ba. 2014. Adam: A Koppel. 2013. Authorship attribution of micro-
method for stochastic optimization. arXiv preprint messages. In Proceedings of the 2013 Conference
arXiv:1412.6980 . on Empirical Methods in Natural Language Pro-
cessing. pages 1880–1891.
Bryan Klimt and Yiming Yang. 2004. The enron
corpus: A new dataset for email classification re- Prasha Shrestha, Sebastian Sierra, Fabio Gonzalez,
search. In European Conference on Machine Learn- Manuel Montes, Paolo Rosso, and Thamar Solorio.
ing. Springer, pages 217–226. 2017. Convolutional neural networks for authorship
attribution of short texts. In Proceedings of the 15th
Robert Layton, Paul Watters, and Richard Dazeley. Conference of the European Chapter of the Associa-
2010. Authorship attribution for twitter in 140 char- tion for Computational Linguistics: Volume 2, Short
acters or less. In 2010 Second Cybercrime and Papers. volume 2, pages 669–674.
Trustworthy Computing Workshop. IEEE, pages 1–
8. Kai Sheng Tai, Richard Socher, and Christopher D
Manning. 2015. Improved semantic representations
David D Lewis, Yiming Yang, Tony G Rose, and Fan from tree-structured long short-term memory net-
Li. 2004. Rcv1: A new benchmark collection for works. arXiv preprint arXiv:1503.00075 .
text categorization research. Journal of machine
learning research 5(Apr):361–397. Duyu Tang, Bing Qin, and Ting Liu. 2015. Docu-
ment modeling with gated recurrent neural network
Laurens van der Maaten and Geoffrey Hinton. 2008. for sentiment classification. In Proceedings of the
Visualizing data using t-sne. Journal of machine 2015 conference on empirical methods in natural
learning research 9(Nov):2579–2605. language processing. pages 1422–1432.
Christopher D. Manning, Prabhakar Raghavan, and Eric W Weisstein. 2002. Inverse hyperbolic tangent .
Hinrich Schtze. 2008. Introduction to Information
Zhou Yilu, Alsarkal Yaqoub, and Zhang Nan. 2016.
Retrieval. Cambridge University Press, Cambridge,
Linking virtual and real-world identities twitter
UK. http://nlp.stanford.edu/IR-book/information-
dataset.
retrieval-book.html.
Yin Zhang, Rong Jin, and Zhi-Hua Zhou. 2010. Un-
Tomas Mikolov, Kai Chen, Greg Corrado, and Jef- derstanding bag-of-words model: a statistical frame-
frey Dean. 2013. Efficient estimation of word work. International Journal of Machine Learning
representations in vector space. arXiv preprint and Cybernetics 1(1-4):43–52.
arXiv:1301.3781 .
533
Using Syntax to Resolve NPE in English
534
Proceedings of Recent Advances in Natural Language Processing, pages 534–540,
Varna, Bulgaria, Sep 2–4 2019.
of VPE using machine learning (Nielsen, 2003) present paper, we do not deal with one-anaphora,
and parsed text (Nielsen, 2004), using linguistic zero anaphora or pronominals, and restrict our fo-
principles (McShane and Babkin, 2016), with sen- cus to NPE.
tence trimming methods (McShane et al., 2015),
reconstruction of sentences with gapping using 3 Task Description
improved parsing techniques that encode elided Resolution of ellipsis comprises two tasks - detec-
material dependencies (Schuster et al., 2018), etc. tion of the elided material and antecedent selec-
There are no known systems that handle NPE tion. In some cases, reference resolution might
detection and resolution in English. However, also be necessary (Liu et al., 2016; Nielsen, 2003).
on a related linguistic phenomenon called one- For example, in (3), the common sense interpreta-
anaphora or one-substitution, in which the elided tion is that Sam loves his girlfriend. But it could
noun is replaced by an overt pro-form, there is also lead to a sloppy reading where it means Sam
a thorough data-driven investigation (Gardiner, loves John’s girlfriend.
2003) and machine-learning methods that use
heuristics proposed in this study (Ng et al., 2005). 3. John loves his girlfriend. Sam does [e] too.
Another phenomenon similar to NPE is zero-
anaphora, which has been thoroughly studied in Note that (3) presents an example of VPE as the
some pro drop languages such as Chinese (Yeh verb along with its predicate are elided. In this
and Chen, 2019a,b) and Japanese (Iida et al., 2007; paper, we focus only on the first two tasks, i.e. de-
Asao et al., 2018; Chen, 2016). Zero-anaphora tection of NPE and antecedent selection.
does not occur in English, although there is some
4 Dataset Preparation
evidence of the phenomenon being used to achieve
certain interactional functions in ordinary conver- There are no dedicated linguistic resources or
sational settings by English speakers (Oh, 2005). datasets for the analysis of NPE in English. How-
There are also proposed heuristics for determining ever, there are many well-known corpora that con-
antecedents of pronominal words (Lappin and Le- tain annotated instances of NPE. One such re-
ass, 1994; Kennedy and Boguraev, 1996). In the source is the Universal Dependency (UD) tree-
Table 1: Syntactic categories that can and cannot license NPE in English, with examples for each category.
535
bank (Silveira et al., 2014) for English that con- sentences without NPE. Hence, our testset has 76
tains example sentences for different types of el- positive and 132 negative samples.
lipsis such as VPE, NPE, etc. The UD tree-
bank marks NPE by raising the dependents of the 5 System Overview
elided noun to the position of head in cases where
the dependents are overtly marked. Through Our system is divided into two parts. An input
a simple search for noun dependents that are sentence is fed into the NPE detection system that
given the status of noun heads, we get a to- decides whether an NPE is present or not. If an
tal of 146 cases of NPE in 120 sentences from NPE is detected, it sends the sentence to the Res-
the UD treebank. There is another compara- olution system where a potential antecedent is se-
tively small corpus called the ParCorFull: a Par- lected. The output of the complete system is ei-
allel Corpus Annotated with Full Coreference ther a decision that there is no ellipsis present in
(Lapshinova-Koltunski et al., 2018) that is dedi- the sentence or an ellipsis site marked along with
cated to anaphora. This corpus targets anaphora, its antecedent.
but deals partly with NPE cases as well, mark-
5.1 NPE Detection
ing them with a nom-ellipsis tag. A simple search
for this tag gives us 5 sentences containing 5 NPE For the task of NPE detection, we exploit a very
cases. We also pick a total of 80 sentences contain- useful syntactic feature, which is the presence of
ing 83 cases of NPE from linguistic textbooks on overt remnants of the noun phrase at the ellipsis
ellipsis (Lobeck, 1995; Saito et al., 2008; Menzel, site. In case of NPE in English, these trigger words
2017; Kim et al., 2019; Corver and van Koppen, are often determiners and modifiers of the elided
2011) to cover even the infrequently occurring noun. These are also known as licensors of ellip-
cases. Finally, we randomly pick 132 sentences sis. In the examples presented in (1) and (2), the li-
that do not contain NPE from UD treebank, Par- censors of elided noun are the cardinal number two
COrFull and the same linguistic textbooks. Some and possessive proper noun Mary’s respectively.
of these negative samples of NPE contain sen- We use these remnants or noun modifiers present
tences with ellipsis other than NPE, such as VPE. at the ellipsis site as cues to locate the elided noun.
An interesting feature about these license of
In total, we curate a small dataset of 337 sen-
NPE in English is that they can only belong to cer-
tences, of which 205 sentences have 234 instances
tain syntactic categories. These include cardinal
of NPE (some sentences contain more than one in-
and ordinal numbers, plural demonstrative deter-
stance of NPE) and the remaining 132 sentences
miners, possessives, adjectives, quantifiers, and a
without NPE. This dataset, albeit small, covers
certain types of determiners. Table 1 provides ex-
a wide variety of the cases of NPE discussed in
amples from each of these categories, along with
the ellipsis literature and will be useful for future
examples of syntactic categories that cannot li-
work. Since the focus of this paper is on present-
cense NPEs in English. Hence, the idea is to use
ing a system for detection and resolution of ellip-
the syntactic environment of the nominal ellipsis
sis, we do not undertake the formidable task of
site to perform detection. Linguistically, our ap-
preparation of annotation guidelines and perform-
proach is similar to detection of VPE by using
ing annotation to prepare a Gold dataset in this pa-
auxiliary and modal verbs as cues (McShane and
per. However, this could be an important future
Babkin, 2016), as VPE in English are licensed by
work considering the limited available resources
auxiliary and modal verbs.
for the analysis of NPE.
To fine tune our rules, we only need positive 5.1.1 Look for Pre-Modifiers and
samples of NPE. However, for testing, it is im- Determiners
portant to use both positive and negative samples. NPE detection is carried out in two steps. In the
We split our dataset into two parts. For fine tun- first step, the input sentence is parsed using the
ing the rules, we randomly pick 140 out of the 205 state-of-the-art spaCy parser (Honnibal and John-
positive sentences (roughly 70%), containing a to- son, 2015) and using the Part-of-Speech (POS)
tal of 158 NPE instances. The remaining 65 sen- tags, we check for the presence of nominal mod-
tences (roughly 30%) contain 76 NPE instances ifiers and determiners from the aforementioned
and are included in the testset along with the 132 syntactic categories that can potentially license an
536
(c) Check for Punctuation
Figure 1: F1-score corresponding to different win-
dow sizes for searching nouns. We first check for a simple feature that
checks for noun modifiers close to punctua-
82 81 tion marks. Since a punctuation can indicate
80 80 a sentential or phrasal break, this could indi-
80 78 78 78 78
F1-Score (%)
537
reference phenomenon (Hobbs and Kehler, 1997). increase in F1-score by 13.73%. Finally, addition
This becomes our motivation to resolve NPE. of only the feature that checks for verbs and aux-
iliaries that immediately follow the noun modifier
(a) Match POS tags of the Licensor with other gives an increase in F1-score by 11.46%. These
Noun Modifiers features are independent of each other and do not
The syntactic environment of the an NPE follow any hierarchy.
comprises the remnants left in the noun The final system together with only the signif-
phrase. One simple way to see structural par- icantly important features is tested on the testset
allelism between the syntactic environment containing 76 positive and 132 negative samples of
of antecedent and that of ellipsis to locate NPE. The detection system is able to correctly de-
antecedents is through matching the syntac- tect 65 instances of NPE out of 76. It also rightly
tic category information. From the detec- predicts 113 out of 132 negative samples as not
tion task, we already have the POS tag in- containing any NPE. It fails to detect 11 positive
formation of the licensor of the NPE. In the cases and falsely detects 29 others. This gives us
first step of antecedent selection, the system a final precision of 69.15%, a recall of 85.53%
checks for other noun phrases in the sentence and an F1-score 76.47%. Out of the 65 NPE
that contain the modifiers with the same POS cases detected by the system, 41 have a textual an-
tag as that of the licensor of the NPE. tecedent and the remaining 24 are exophoric and
need extra-linguistic context to resolve. Our sys-
(b) Select Antecedent tem is able to select a potential antecedent for 37
If a POS tag matching the licensor of the of these from the text, of which 32 are correct pre-
NPE (detected in the NPE detection task) dictions. The system fails to select any antecedent
is found in the sentence, the system outputs for the 9 cases. This gives us a final precision
the noun that the modifier with the same of 78.79%, a recall of 63.41% and an F1-score
tag modifies as the antecedent of the NPE. 70.27%. See table 2 for precision, recall and F1-
If there are more than one such modifiers score values for the NPE detection and resolution
found, the system selects the one nearest to tasks on the testset.
the NPE as distance generally has a role to One of the main reasons of the low accuracy of
play in anaphora and coreference resolution our system is wrong POS tags generated for sen-
tasks (Lappin, 1996). tences with missing or incomplete information as
in the case of ellipses (Menzel, 2017). Secondly,
6 Results although, licensors of NPE and modifiers of the
Simply looking for nouns in the context of the antecedent indeed show similarity in terms of syn-
noun modifiers gives a poor F1-score of 64.20%. tactic category information, this might not always
Addition of only the feature that checks for noun be the case.
modifiers raised as verbal arguments results into 5. The books were new, and all six [e] were on
an increased in F1-score by 10.25%. Addition syntax.
of only the punctuation feature after auxiliary and
modal verbs resulted into an increase in accuracy For example in (5), the NPE licensor is a cardi-
by 3% for VPE detection task (Nielsen, 2004). nal number, but antecedent books has the definitive
However, in our task, this resulted into a drop article in its Noun Phrase.
in accuracy by 0.05%. Hence, we excluded this
7 Conclusion & Future Work
feature from the final system. Addition of only
the feature that checks for prepositions immedi- This paper described a syntax-based system for
ately following the noun modifier results into an automatic detection of NPE in English. The sys-
538
tem takes in free English text and exploits syntac- Tilbe Goksun, Tom W. Roeper, Kathy Hirsh-Pasek, and
tic constraints on the licensors of NPE to mark the Roberta Michnick Golinkoff. 2010. From noun-
phrase ellipsis to verbphrase ellipsis: The acquisi-
site of ellipsis and syntactic parallelism between
tion path from context to abstract reconstruction.
antecedent-ellipsis syntactic environments to se-
lect potential antecedents. Evaluated on a testset Christine Gunther. 2011. Noun ellipsis in english: ad-
containing both positive and negative NPE sam- jectival modifiers and the role of context. The struc-
ture of the noun phrase in English: synchronic and
ples, the system achieves an F1-score of 76.47% diachronic explorations, 15(2):279–301.
on the detection task and 70.27% on the resolu-
tion task. Although these numbers are not high, Michael Alexander Kirkwood Halliday and Ruqaiya
they can be useful as baselines for future work Hasan. 1976. Cohesion in english. page 76.
in this direction. NLP research on ellipsis and Daniel Hardt. 1997. An empirical approach to vp ellip-
NPE in particular suffers from a scarcity of re- sis. Computational Linguistics, 23(4):525541.
sources. While a rule-based system such as ours
does not need sizable data for training, with more Daniel Hardt. 1999. Dynamic interpretation of
verb phrase ellipsis. Linguistics and Philosophy,
language resources available in future, machine 22(2):187–221.
learning methods can also be used.
Jerry R. Hobbs and Andrew Kehler. 1997. A theory
of parallelism and the case of vp ellipsis. In Pro-
ceedings of the 35th Annual Meeting of the Associa-
References tion for Computational Linguistics and Eighth Con-
Yoshihiko Asao, Ryu Iida, and Kentaro Torisawa. ference of the European Chapter of the Association
2018. Annotating zero anaphora for question an- for Computational Linguistics, ACL ’98/EACL ’98,
swering. In LREC. pages 394–401, Stroudsburg, PA, USA. Association
for Computational Linguistics.
Johan Bos and Jennifer Spenader. 2011. An annotated Matthew Honnibal and Mark Johnson. 2015. An im-
corpus for the analysis of vp ellipsis. Language Re- proved non-monotonic transition system for depen-
sources and Evaluation, 45(4):463494. dency parsing. In Proceedings of the 2015 Con-
ference on Empirical Methods in Natural Language
Wei Chen. 2016. The motivation of ellipsis. The- Processing, pages 1373–1378, Lisbon, Portugal. As-
ory and Practice in Language Studies, 6(11):2134– sociation for Computational Linguistics.
2139.
Nina Hyams, Victoria Mateu, and Lauren Winans.
Sandra Chung, William Ladusaw, and James Mc- 2017. Ellipsis meets wh-movement: sluicing in
Closkey. 2010. Sluicing (:) between structure and early grammar.
inference. In Representing language: Essays in
honor of Judith Aissen. Ryu Iida, Kentaro Inui, and Yuji Matsumoto. 2007.
Zero-anaphora resolution by learning rich syntactic
Norbert Corver and Marjo van Koppen. 2011. pattern features. ACM Trans. Asian Lang. Inf. Pro-
Np-ellipsis with adjectival remnants: A micro- cess., 6.
comparative perspective. Natural Language and
Linguistic Theory, 29. Kyle Johnson. 2001. What vp ellipsis can do, and what
it cant, but not why. pages 439–479.
Jeroen van Craenenbroeck and Jason Merchant. 2013.
Ellipsis phenomena. In The Cambridge Handbook Christopher Kennedy and Branimir Boguraev. 1996.
of Generative Syntax, pages 701–745. Cambridge Anaphora for everyone: pronominal anaphora res-
University Press. olution without a parser. In In Proceedings of COL-
ING, page 113118.
Mary Dalrymple, Stuart M. Shieber, and Fernando Nayoun Kim, Laurel Brehm, and Masaya Yoshida.
C. N. Pereira. 1991. Ellipsis and higher-order unifi- 2019. The online processing of noun phrase ellipsis
cation. Linguistics and Philosophy, 14(4):399–452. and mechanisms of antecedent retrieval. Language,
Cognition and Neuroscience, 34(2):190–213.
Lyn Frazier. 2008. Processing ellipsis: A processing
solution to the undergeneration problem? In Pro- Ronald W Langacker. 1999. Grammar and Concep-
ceedings of the 26th West Coast Conference on For- tualization (Cognitive Linguistics Research), vol-
mal Linguistics. ume 14. Mouton de Gruyter, New York.
Mary Gardiner. 2003. Identifying and resolving one- Shalom Lappin. 1996. The interpretatin of ellipsis. In
anaphora. Department of Computing, Division of Shalom Lappin, editor, The Handbook of Contempo-
ICS, Macquarie University. rary Semantic Theory, pages 145–176. Blackwell.
539
Shalom Lappin and Herbert Leass. 1994. A syntacti- Sebastian Schuster, Joakim Nivre, and Christopher D.
cally based algorithm for pronominal anaphora res- Manning. 2018. Sentences with gapping: Parsing
olution. Computational Linguistics, 20:535561. and reconstructing elided predicates. ArXiv e-prints.
Ekaterina Lapshinova-Koltunski, Christian Hardmeier, Natalia Silveira, Timothy Dozat, Marie-Catherine
and Pauline Krielke. 2018. Parcorfull: a parallel de Marneffe, Samuel Bowman, Miriam Connor,
corpus annotated with full coreference. In Proceed- John Bauer, and Christopher D. Manning. 2014. A
ings of the Eleventh International Conference on gold standard dependency corpus for English. In
Language Resources and Evaluation, LREC 2018, Proceedings of the Ninth International Conference
Miyazaki, Japan, May 7-12, 2018. on Language Resources and Evaluation (LREC-
2014).
Charlotte Lindenbergh, Angeliek van Hout, and Bart
Hollebrandse. 2015. Extending ellipsis research: Frank Wijnen, Tom W. Roeper, and Hiske van der
The acquisition of sluicing in dutch. BUCLD 39 On- Meulen. 2003. Discourse binding: Does it begin
line Proceedings Supplement, 39. with nominal ellipsis?
Zhengzhong Liu, Edgar Gonzalez, and Dan Gillick. Ming Xiang, Julian Grove, and Jason Merchant. 2014.
2016. Verb phrase ellipsis detection using automati- Ellipsis sites induce structural priming effects.
cally parsed text. pages 32–40.
Ching-Long Yeh and Yi-Chun Chen. 2019a. Using
Anne Lobeck. 1995. Functional Heads, Licensing, and zero anaphora resolution to improve text categoriza-
Identification. Oxford University Press. tion.
Marjorie McShane and Petr Babkin. 2016. Detection Ching-Long Yeh and Yi-Jun Chen. 2019b. An em-
and resolution of verb phrase ellipsis. Linguistic Is- pirical study of zero anaphora resolution in chinese
sues in Language Technology, 13(1). based on centering model.
Marjorie McShane, Sergei Nirenburg, and Petr Babkin.
2015. Sentence trimming in service of verb phrase
ellipsis resolution. In EAPCogSci.
Katrin Menzel. 2017. Understanding English-German
contrasts: a corpus-based comparative analysis of
ellipses as cohesive devices. Ph.D. thesis, Universi-
tat des Saar- landes, Saarbrucken.
Jason Merchant. 2004. Fragments and ellipsis. Lin-
guistics and Philosophy, 27(6):661–738.
Jason Merchant. 2010. Three Kinds of Ellipsis: Syn-
tactic, Semantic, Pragmatic?
Hwee Tou Ng, Yu Zhou, Robert Dale, and Mary Gar-
diner. 2005. A machine learning approach to iden-
tification and resolution of one-anaphora. pages
1105–1110.
Leif Arda Nielsen. 2003. Using machine learning tech-
niques for vpe detection.
Leif Arda Nielsen. 2004. Verb phrase ellipsis detection
using automatically parsed text.
Sun-Young Oh. 2005. English zero anaphora as an in-
teractional resource. Research on Language and So-
cial Interaction, 38(3):267–302.
Dongwoo Park. 2017. When does ellipsis occur, and
what is elided? PhD dissertation, University of
Maryland.
Alain Rouveret. 2012. Vp ellipsis, phases and the syn-
tax of morphology. Natural Language & Linguistic
Theory, 30(3):897963.
Mamoru Saito, Jonah Lin, and Keiko Murasugi. 2008.
Nominal-ellipsis and the structure of noun phrases
in chinese and japanese.
540
Is Similarity Visually Grounded? Computational Model of Similarity for
the Estonian language
541
Proceedings of Recent Advances in Natural Language Processing, pages 541–549,
Varna, Bulgaria, Sep 2–4 2019.
based on Estonian language resources: corpora, is surface similarity. By evaluating the similarity
taxonomies and lexical ontologies. using computational vision models we contribute
The rest of the paper is organized as follows. to a better understanding of the notion of similarity
The next section puts our research in context and itself. Second, we ask how the traditional models
then the EstSimLex-999 set, the SimLex-999 set derived from Estonian corpora and lexical ontolo-
translated into Estonian, is presented 3. Section 4 gies correlate with the judgments of native Esto-
discusses the families of models for computing the nian speakers. In this way we extend the similar-
similarity between the word pairs in EstSimLex- ity study to other language, a less explored one,
999. Section 5 presents and discusses the results. yet an interesting one. Third, we study if our com-
In particular, we answer how similarity is influ- putational models trained on Estonian data predict
enced by language and quantify the power of the better the EstSimLex-999 scores or the SimLex-
computer vision models to capture the similarity 999 scores. More precisely we want to know if the
for concrete concepts. The paper ends with the language influences the similarity judgments.
conclusions.
3 EstSimLex-999
2 Related Work
To translate SimLex-999 the Google Translation
Felix Hill and coauthors (2015) undertook an ex- API and a bilingual English-Estonian dictionary
tensive discussion of the concept of similarity in containing 87665 entries have been used, obtain-
Computational Linguistics, introduced the gen- ing rough Estonian equivalents. A native Estonian
uine similarity set SimLex-999 and computed the speaker has chosen the correct translations. If an
correlations between the similarity measures of English word in a similarity pair is ambiguous, the
corpus based computational models and the hu- sense that makes the pair more similar is preferred.
man judge scores. Closely following this line Finally, after correction and the discussion with an
of research Ira Leviant and Roi Reichart (2015) Estonian linguist we have produced the similar-
translated SimLex-999 in Italian, German and ity set referred from now on as EstSimLex-999.
Russian and collected similarity scores from na- When translating, we have been careful to pre-
tive speakers. They compute correlations between serve the part of speech of the English concepts.
the human judgments and Vector Space Models This makes the comparison between the computa-
(VSM) in a multilingual setting. tional models of similarity for English and Esto-
In the subsequent research the authors improve nian easier. Nevertheless, due to cultural and lin-
the similarity computational models and boost guistic differences some English similarity pairs
the correlation coefficient with the human judg- were hard to translate. For example, the English
ments. For example, Schwartz et al. (2015) pair (taxi, cab) was translated as (taksi, takso)
learn a word level representation based on sym- even if the second term of the Estonian pair is not
metric patterns that achieves a Spearman corre- widely used. Another example is the pair (sup-
lation of 0.517 with SimLex-999. An interest- per, dinner). The Estonian culinary tradition does
ing work belongs to Faruqui and Dyer (2015), not distinguish between the two concepts, there-
who used non-distributional word representations fore we have translated the pair with the synony-
derived form Princeton WordNet, FrameNet and mous words (õhtusöök, õhtueine). Please, notice,
Penn Treebank to reach a Spearman correlation of that for many non-British native English speakers
0.58 with SimLex-999. Hybrid models (Recski the words supper and dinner are also synonymous.
et al., 2016),combining features from lexical on- Some translations would have been more accu-
tologies and word embeddings, seem to be even rate using multiwords, but we abide by the orig-
better (Spearman Correlation 0.76). inal requirement that the similarity pairs should
In this work we were not interested in obtain- contain single words only. Overall, we have pro-
ing the best correlation between the computational duce an accurate translation of the English original
models and the human judgments. That will be SimLex-999 set preserving the distribution of the
the topic of a future work. Instead, we were con- part of speeches and satisfying the demand that the
cerned with three problems. First, we are inter- word pairs should not contain multiple words.
ested in how much of similarity is grounded in the Four native Estonian speakers have rated the de-
visual features, that is, how much of the similarity gree of similarity between each of the 999 pairs.
542
The rating instructions are the same as in the orig- ilarity the computer vision models work best for
inal study (Hill et al., 2015). These instructions do concepts representing concrete objects.
not attempt to define what similarity is, but rather In what follows we will briefly describe the
clarify the concept contrasting it with associa- models tested.
tion, and comparing it with synonymy. The inter-
annotator agreement was computed as the aver- 1. Word2Vec. Word2Vec is a distributional
age of pairwise Spearman correlations between model (Mikolov et al., 2013) implemented as
the scores of all raters. The overall agreement is a two layer neural network. If two words ap-
0.766. A direct comparison with the correlation pear in similar contexts in a corpus the net-
coefficient computed in the English study (0.67) work will output embedding vectors, known
is not possible because the number of annotators as neural vector embeddings, which are close
is different. At this stage we were not interested in the embedding space. Word2Vec computes
in recruiting many annotators through platforms the neural vector embeddings either predict-
like Mechanical Turk, but rather in gaining in- ing the target word from the context (this
sights into human similarity judgments by direct method is known as continuous bag of word
discussion with the annotators. In any case, re- (CBOW)) or as the target context from the
cruiting a comparable number of Estonian speak- word (this method is know as Skip Gram).
ers is unlikely, as this language is natively spoken
2. SenseGram. SenseGram (Pelevina et al.,
by less than 1 million people. The main thing no-
2016) is not a distributional model per se,
ticed is that there are few pairs of adjectives and
but a method to obtain word senses from
verbs highly rated in English but with a low score
word embeddings. This word discrimination
in Estonian. For example, the English verb pair
method takes as input word embeddings (like
(appear, attend) has a score of 6.28 in SimLex-999
those generated by Word2Vec or any other
and its Estonian translation (ilmuma, osalema) has
distributional model) and clusters them. The
a score equal to 0.5.
induced word senses correspond to the clus-
ters of word embeddings.
4 Models for Similarity Computations
Three families of similarity models are evaluated 3. Path Similarity Measures. The path sim-
: distributional models, semantic network models ilarity measures exploit the graph struc-
and computer vision models. ture of semantic networks to find simi-
larities between concept pairs. We have
The distributional models are an implementa-
explored various similarity measures like
tion of John Rupert Firth’s hypothesis “You shall
Leacock-Chodorow similarity (Leacock and
know a word by the company it keeps” (Firth,
Chodorow, 1998).
1961) which basically states that words that have
similar meanings appear in comparable syntag- 4. Autoencoders. Autoencoders are deep neu-
matic contexts. Nowadays, the most advanced ral networks which learn to reconstruct the
distributional models are the neural word embed- input. In the reconstruction process one of
dings. the autoencoder layers contains less nodes
The second family of models derive the seman- than the input layer, thus forcing the net-
tic similarity from the taxonomic structure of se- work to learn a lower level representation of
mantic networks. The IS-A relation induces the the input. The idea behind using the au-
inheritance of the properties.The above mentioned toencoders is that the sparse representations
concepts, apple and pear, are similar because they learned when encoding similar concepts will
inherit all the properties from their superordinate be close in the embedding space.
concept (fruit). Unlike the distributional models,
the semantic networks tells us also why the con- 5. Pretrained Convolutional Neural Net-
cepts are similar. works. Convolutional Neural Networks
The third family of models are the computer (CNN) are deep neural networks architec-
vision models. The similarity between two con- tures suitable for extracting patterns from
cepts is the distance between their image represen- images. Inspired by experiments in neuro-
tations. Because of the visual nature of this sim- science (Hubel and Wiesel, 1959), CNN’s
543
first layers train convolution filters to detect SL-999 ESL-999
low-level features of an image like lines and Model r ρ τ r ρ τ
corners. Higher network levels combine the cbow 1 .42 .42 .29 .46 .47 .33
low level-features to find high-level image sg 2 .37 .36 .24 .41 .42 .3
features roughly corresponding to the human cbow 3 .33 .33 .23 .33 .34 .24
language semantic descriptions of the ob-
jects. For example, they might detect parts, Table 1: The results for the best three distribu-
like the wheels or the hood of a car. When tional models
CNN are trained on big databases of classi-
fied images the semantic representations of senses/concept, with about 300 word pairs
the concrete concepts can be “read” from the having more than one sense. For the am-
deeper network levels. These representations biguous word pairs (where at least one of the
are then used to compute concept similarity. words in the pair has more than one sense) the
word sense that maximizes the cosine simi-
5 Results
larity score of a word pair is evaluated.
First the results for the distributional models are
• Estnltk pretrained word embeddings. Es-
shown, then the results for the semantic network
tnltk (Orasmaa et al., 2016) contains 8 word
models will be presented. Finally, the results for
pretrained embeddings. 4 of them are trained
the neural computer vision models will be shown.
with the CBOW method and the other 4 were
The correlation coefficients between the scores as-
trained with the Skip-Gram method, on the
signed by the computational similarity models for
raw and lemmatized versions of the Estonian
interesting subsets (e.g. abstract and concrete con-
Reference Corpus (Kaalep et al., 2010). The
cepts) and the two similarity sets are also com-
Estonian Reference Corpus is a 1.3 billion
puted. In the tables in this section the SimLex-999
word corpus, crawled from the web, contain-
is abbreviated as SL-999 and the EstSimLex-999
ing mainly newspaper text.
as ESL-999.
• Facebook pretrained word embeddings.
5.1 Distributional Models
The Facebook word embeddings (Bo-
When evaluating the Word2Vec and SenseGram janowski et al., 2017) have been trained with
models, if the embedding vectors corresponding to CBOW method on 294 language versions of
the words in the SimLex-999 or EstSimlex-999 are Wikipedia.
missing, the word-pair is eliminated. The model
word similarity score is computed as cosine simi- The distributional models evaluate on average
larity between the vector embeddings correspond- 985 word pairs. Each CBOW and Skip Gram
ing to the words in the each word pair. Pearson (r), model has 4 meta-parameters : the number of di-
Spearman (ρ ), and Kendall (τ ) correlations are mensions, the window size, the minimum count
calculated between EstSimLex-999 and Simlex- threshold and the number of iterations. Due to
999 human judge scores and the model word sim- consideration related to space we only present the
ilarity scores. The word embeddings were trained best three results in the table 5.1. The whole set
on Estonian monolingual corpora and the Estonian of results for the 67 distributional models trained
Wikipedia. The following word embeddings have and all the figures and the tables in this paper are
been used: available online linked from our github repository.
2 . The best model on the first row in the table
• EA word embeddings. 9 Skip-Gram and 5.1, for example, has been trained with the 300 di-
20 CBOW models, with different parame- mensions, a window size equal to 1, the minimum
ter settings, were trained on the lemmatized count threshold being 10, and 20 iterations.
version of etTenTen corpus of Estonian Web In the first place one can notice that CBOW
1 by Eleri Aedma. Word senses were in-
trained word embedding perform better than Skip-
duced from the traditional word embeddings Gram trained word embeddings. Moreover, the
using SenseGram. SenseGram finds 1.6 correlation coefficients between EstSimLex-999
1 2
DOI: 10.15155/1-00-0000-0000-0000-0012EL https://github.com/estsl/EstSimLex-999
544
Figure 1: The average performance for POS-based
subsets Figure 2: The average performance for the most
concrete and abstract subsets
545
SL-999 ESL-999 SL-999 ESL-999
r ρ τ r ρ τ r ρ τ r ρ τ
PS .47 .47 .35 .54 .52 .39 PS .32 .31 .22 .37 .34 .24
LC .36 .36 .26 .41 .43 .31 LC .31 .3 .21 .35 .34 .28
WuP .41 .45 .32 .49 .53 .39 WuP .39 .37 .28 .4 .37 .27
Table 2: The results for the Estonian Wordnet Table 3: The results for the Wikipedia Page Tax-
onomy
Palmer similarity (WuP) (Wu and Palmer, 1994).
Pearson (r), Spearman (ρ ), and Kendall (τ ) cor- much lower than those computed before (with the
relations were calculated between EstSimLex-999 Estonian Wordnet). Also, surprisingly and for the
and SimLex-999 human judge scores and the net- first time in this study, there is no statistically sig-
work similarity scores for the disambiguated word nificant difference between the correlation coef-
pairs. ficients computed with SimLex-999 human judg-
The Estonian Wordnet is an ongoing effort, ments scores and the EstSimLex-999 human judg-
it pursues roughly the same organization princi- ment scores.
ples as Princeton WordNet (Miller et al., 1990),
and it is manually built by a group of linguists. 5.3 Computer Vision Models
The version used in this study contains approxi- Because the visual similarity is correlated with the
mately 85.000 synsets. The above described dis- level of concreteness of an object, the computer
ambiguation procedure maps approximately 770 vision models are fed with word pairs where both
word pairs onto the Estonian Wordnet. The results words have a degree of concreteness higher than
corresponding to the Estonian Wordnet are in the the a threshold equal to 4.8 . This criterion gives
table 2. us 136 word pairs. We have downloaded, using
The best results are obtained for Wu & Palmer Yandex image search engine 200 images for each
similarity measure and EstSimLex-999. This sim- word in the selected word pairs.
ilarity measure considers the depth of the concepts The first architecture trains a Convolutional Au-
in the semantic network hierarchy along with the toencoder (CAE) on the downloaded images. The
depth of their Lowest Common Subsumer. Unlike encoder consists of 3 convolutional layers, each
the path similarity measure it favour the concepts followed by a max-pooling layer. The decoder
that are deeper in the hierarchy. As in the case consists of 3 convolutional followed by upsam-
of the distributional models the EstSimLex-999 pling layers. The similarity between two images is
correlation scores are better than the SimLex-999 calculated as the cosine similarity between the cor-
ones. A fact worth noticing is that the difference responding encoder vectors. The similarity score
between the correlation scores for the two similar- for a word pair is the average score between all the
ity sets is greater when we compute the similar- images corresponding to the words in the pair.
ity score based on the Estonian Wordnet structure The second architecture is the winner of the Im-
instead of estimating the similarity using distribu- ageNet 2015 competition. It is a CNN network ar-
tional models. chitecture invented by Microsoft Research, called
The taxonomy was extracted from the (Esto- ResNet(He et al., 2016) (abbreviation for Resid-
nian) Wikipedia page text (Wikipedia Page Tax- ual Network). DNNs with many layers are diffi-
onomy) by the language technology research cult to train due to vanishing and exploding gra-
group at Università Roma Tre (Flati et al., 2016). dient problems. ResNet solves these problems
The Wikipedia Page Taxonomy contains approx- with residual learning. The ResNet architecture
imately 87000 concepts. We could map around comes in many variants, depending on the number
200 word pairs onto the Wikipedia Page Taxon- of layers the network has. The widely employed
omy. The results for this taxonomy are in Table ResNet-18 variant with 18 layers is used in this
3. study. The network is pretrained on the ImageNet
The correlation coefficients between the simi- database (Deng et al., 2009) which contains over
larity measures computed for the Wikipedia Page 1 million images classified under 1000 Princeton
taxonomy and the human judgments scores are WordNet categories. Being trained on such a big
546
SL-999 ESL-999
Model
r ρ τ r ρ τ
CAE .25 .28 .19 .17 .22 .15
RN18 .37 .38 .26 .34 .38 .27
547
semantic features when these features are com- address the same problem for different languages.
puted from huge databases of images like Ima-
geNet. Although the empirical evidence heavily Reproducibility
depends on the quality of the ImageNet database The EstSimLex-999 set annotated with the hu-
this finding shows that other factors have sig- man judge similarity scores, the code used to
nificant weight in human similarity judgments. compute the results in this paper, the complete
Maybe a case can be that these factors are the deep set of tables and the figures and most of the
semantic features we have briefly mentioned in the resources used in this paper can be referenced
introduction section. However, a definitive answer from the Github repository https://github.
to what these factors are and how to account for com/estsl/EstSimLex-999 .
them goes beyond this research.
Acknowledgments
6 Conclusions
A number of persons have helped us with this
In this study have addressed some aspects of the study. We like to thank Aliis Aalmann and
computational models of similarity applied to the Kadri Muischnek for the aid in translation of the
Estonian language. SimLex-999. We also like to thank Eleri Aedmaa
In the first place we have found that the neu- for making available the word and sense embed-
ral visual models of similarity can explain a part dings and the anonymous annotators for rating the
of the similarity between the words representing Estonian word pairs. This study was supported by
concrete concepts. This invites the conclusion that the Estonian Ministry of Education and Research
deep similarity models might be involved in ac- (IUT20-56).
counting for the unexplained part of similarity.
In the second place and differently from the References
finding in (Leviant and Reichart, 2015) an effect of
Eduard Barbu, Heili Orav, and Kadri Vare. 2018. Topic
the language on the similarity has been found. Un- interpretation using wordnet. In Kadri Muischnek
like in that study the Estonian computational mod- and Kaili Müürisep, editors, Baltic HLT. IOS Press,
els of similarity better correlate with the word pair volume 307 of Frontiers in Artificial Intelligence
similarity score assigned by the Estonian subjects and Applications, pages 9–17.
that with the scores assigned by English speaking Piotr Bojanowski, Edouard Grave, Armand Joulin, and
subjects. Tomas Mikolov. 2017. Enriching word vectors with
In the third place the best computational mod- subword information. Transactions of the Associa-
tion for Computational Linguistics 5:135–146.
els are those derived from human built semantic
networks. They are better than the neural distri- J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-
butional models but still they correlate moderately Fei. 2009. ImageNet: A Large-Scale Hierarchical
Image Database. In CVPR09.
with the human judgments. This means that there
is more to similarity than taxonomic similarity. On Jrme Euzenat and Pavel Shvaiko. 2013. Ontology
the other hand the Estonian Wordnet is still work Matching. Springer Publishing Company, Incorpo-
rated, 2nd edition.
in progress, therefore we cannot rule out that a
more complete wordnet can boost the similarity Manaal Faruqui and Chris Dyer. 2015. Non-
scores. Unlike the original study (Hill et al., 2015) distributional word vector representations. CoRR
abs/1506.05230. http://arxiv.org/abs/1506.05230.
we have found that the word embedding compu-
tational models better correlate to the scores for J.R. Firth. 1961. Papers in Linguistics 1934-
nouns and not the adjectives. Intuitively, this re- 1951: Repr. Oxford University Press.
https://books.google.ee/books?id=VxiWHAAACAAJ.
sult makes sense as nouns have richer mental rep-
resentations than other morphological categories, Tiziano Flati, Daniele Vannella, Tommaso Pasini, and
therefore one expects that the similarity is better Roberto Navigli. 2016. Multiwibi: The multilingual
wikipedia bitaxonomy project. Artif. Intell. 241:66–
defined for nouns than for other parts of speech. 102.
In the future we will work to better understand
Sebastien Harispe, Sylvie Ranwez, Stefan Janaqi, and
the other components of similarity for the concrete Jacky Montmain. 2015. Semantic Similarity from
concepts, improve and refine the computational Natural Language and Ontology Analysis. Morgan
models of similarity for the Estonian language and & Claypool Publishers.
548
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Representation Learning for NLP. Association for
Sun. 2016. Deep residual learning for image recog- Computational Linguistics, Berlin, Germany, pages
nition. 2016 IEEE Conference on Computer Vision 174–183. http://anthology.aclweb.org/W16-1620.
and Pattern Recognition (CVPR) pages 770–778.
Gábor Recski, Eszter Iklódi, Katalin Pajkossy, and
Felix Hill, Roi Reichart, and Anna Korhonen. 2015. Andras Kornai. 2016. Measuring semantic sim-
Simlex-999: Evaluating semantic models with (gen- ilarity of words using concept networks. In
uine) similarity estimation. Computational Linguis- Proceedings of the 1st Workshop on Representa-
tics . tion Learning for NLP. Association for Computa-
tional Linguistics, Berlin, Germany, pages 193–200.
David H. Hubel and Torsten N. Wiesel. 1959. Recep- https://doi.org/10.18653/v1/W16-1622.
tive fields of single neurons in the cat’s striate cortex.
Journal of Physiology 148:574–591. Roy Schwartz, Roi Reichart, and Ari Rappoport. 2015.
Symmetric pattern based word embeddings for im-
Heiki-Jaan Kaalep, Kadri Muischnek, Kristel Ui- proved word similarity prediction. In Proceed-
boaed, and Kaarel Veskis. 2010. The esto- ings of the Nineteenth Conference on Computational
nian reference corpus: Its composition and Natural Language Learning. Association for Com-
morphology-aware user interface. In Proceedings putational Linguistics, Beijing, China, pages 258–
of the 2010 Conference on Human Language 267. https://doi.org/10.18653/v1/K15-1026.
Technologies – The Baltic Perspective: Pro-
ceedings of the Fourth International Conference Stella Vosniadou and Andrew Ortony, editors. 1989.
Baltic HLT 2010. IOS Press, Amsterdam, The Similarity and Analogical Reasoning. Cambridge
Netherlands, The Netherlands, pages 143–146. University Press, New York, NY, USA.
http://dl.acm.org/citation.cfm?id=1860924.1860949.
Max Wertheimer. 1938. Laws of organization in per-
Claudia Leacock and Martin Chodorow. 1998. Com- ceptual forms. In W. Ellis, editor, A Source Book
bining Local Context and WordNet Similarity for of Gestalt Psychology, Routledge and Kegan Paul,
Word Sense Identification, volume 49, pages 265– London, pages 71–88.
283.
Zhibiao Wu and Martha Palmer. 1994. Verbs semantics
Ira Leviant and Roi Reichart. 2015. Judgment lan- and lexical selection. In Proceedings of the 32Nd
guage matters: Multilingual vector space models for Annual Meeting on Association for Computational
judgment language aware lexical semantics. CoRR Linguistics. Association for Computational Linguis-
abs/1508.00106. http://arxiv.org/abs/1508.00106. tics, Stroudsburg, PA, USA, ACL ’94, pages 133–
138. https://doi.org/10.3115/981732.981751.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg
Corrado, and Jeffrey Dean. 2013. Distributed
representations of words and phrases and their
compositionality. In Proceedings of the 26th
International Conference on Neural Information
Processing Systems - Volume 2. Curran As-
sociates Inc., USA, NIPS’13, pages 3111–3119.
http://dl.acm.org/citation.cfm?id=2999792.2999959.
549
Language-Agnostic Twitter Bot Detection
Jürgen Knauth
Institute for Computer Science
University of Göttingen
jknauth@uni-goettingen.de
550
Proceedings of Recent Advances in Natural Language Processing, pages 550–558,
Varna, Bulgaria, Sep 2–4 2019.
theoretic danger that democratic processes might count data records and conducted an even more
be in influenced by bots. Badawy et al., 2018 an- in depth analysis of the phenomenon of Twitter
alyzes a set of Russian bots and how their infor- bots. Building a classifier they achieved an accu-
mation is picked up by American twitter users re- racy of 95% (Cresci et al., 2015). In their 2017 pa-
lated to these bot accounts. Their research shows per they discovered a new generation of social bots
that information is introduced selectively into so- being more sophisticated compared to bots already
cial media such as Twitter. Badawy et al., 2018 known. They improved their performance by us-
show that this information gets picked up and is ing advanced clustering algorithms and a large set
spread - by users and other bots. The extent to of 126 features and were able to detect up to 97.6%
which such influence can actually influence pub- of Twitter bots on one of their data sets.
lic opinion is not yet clear and requires further re- Newer work uses sophisticated machine learn-
search. However, the very fact that such informa- ing techniques. Cai et al., 2017 detect bots by
tion is disseminated makes detecting Twitter bots modeling users and their behavior using deep
an important issue to address. learning by focusing on topic, latent sentiment and
temporal aspects. Their CNN-LSTM model learns
2 Related Work from user content as well as user behavior. With
this approach they achieve an F1 score of 88.30%
Detecting Twitter bots has been addressed as a re- percent for their deep learning model outperform-
search topic for quite some time. Wang, 2010 re- ing reference models based on content or behavior
constructed a graph model based on crawled Twit- alone.
ter accounts and tweets exploiting follower and Efthimion et al., 2018 use an approach with sup-
friend relationships. They use their data to im- port vector machines. They make use of the Lev-
plement a spam detection approach using classic enshtein distance to detect similar posts for their
Bayesian classication to distinguish between nor- classification approach. This is noteworthy, as we
mal behavior from suspicious behavior on Twitter. make use of Levenshtein distance as well, but in
In 2012 a large scale classification study was contrast to Efthimion et al., 2018 use it to detect
conducted by Chu et al., 2012 to detect bots, hu- similarities not among tweet content but on ac-
mans and ”cyborgs” in Twitter data. The term ”cy- count level. (See below.) Efthimion et al., 2018
borgs” is used here for accounts that are a symbio- use the same data as we use in our work and
sis of humans and bots, making this a multinomial achieve and accuracy of 95.77% (which we out-
problem. Chu et al., 2012 used a dataset of 2000 perform in our work).
manually classified Twitter accounts and achieved Lundberg et al., 2018 provide two classifiers
an accuracy of about 96% with a random forest that are used to not only build an offline classi-
approach. fier, but a system that performs online analysis of
Clark et al., 2015 uses a different approach. tweets as they emerge from the Twitter commu-
They not only crawl Twitter directly to extract nity. They achieve an accuracy of about 98%.
tweets but additionally use data from a honeypot.
They provided a setup with Twitter accounts be- 3 Contribution and Outline
ing bots themselves. Twitter users following these 3.1 Research Questions
accounts for no real reason are considered as be-
ing bots. Clark et al., 2015 focus on content of According to the importance of social media and
the tweets and use three different features to mea- Twitter bots outlined in the introduction, the fol-
sure tweet similarity and hyperlinks. Based on this lowing scientific questions arise regarding the
they propose a classifier to distingusish between recognition of Twitter bots:
”organic” and ”robotic” texts achieving an accu- • Is it possible to detect bots at account level
racy of 90.32% of the ”organic” and 95.2% of the only, without taking into account the content
”robotic” accounts. They additionally identify dif- provided by these accounts?
ferent behavior among these accounts indicating
• Does bot detection benefit from content anal-
that there are several different classes of spammers
ysis?
and spam bots.
Cresci et al., 2015 and Cresci et al., 2017 • Is it possible to reliably detect bots in a way
manually analyzed and annotated over 8000 ac- that avoids language-specific features?
551
• Creating training data for bot detection is dif- 4.2 Dataset
ficult. How large does such a data set have to Our work is based on the MIB dataset (Cresci
be in order to achieve useful training results et al., 2017), which contains 8375 annotated Twit-
for machine learning? Can a small data set be ter accounts.
sufficient?
• 3473 accounts - humans
4 Methodology
• 991 accounts - political candidate retweeters
4.1 Considerations about Bots and Humans
• 3457 accounts - paid apps spammers
Our approach to identifying Twitter bots is based
on the assumption that bots are fundamentally dif- • 464 accounts - amazon.com spammers
ferent from humans in some aspects. Our con-
• Total number of accounts: 8375 accounts
sideration is that two categories should be distin-
guished here: This data set contains data records about the ac-
counts themselves as well as tweets created.
• Technical differences
4.3 Baseline
• Purpose-related differences In the next sections we lay out our feature ex-
traction and machine learning process. In order
4.1.1 Technical Differences to compare our results with a baseline we reim-
Due to the fact that bots are computer programs, plement one of the machine learning classifica-
they are not subject to certain human limitations. tion systems described in Kudugunta and Ferrara,
Computer programs can act instantly in contrast 2018. This classifier is quite a high baseline as it
to humans who need time to reflect and are often performs very well and is - to our knowledge - the
occupied with other tasks of daily life and work. It current state of the art.
can therefore be assumed that human behaviour is
4.4 Feature Extraction
different from bot behaviour with regard to timing
and the orientation of published content. For building a machine learning classifyer we re-
Our considerations are also based on the as- quire a matrix of feature values for training and
sumption that it is difficult for computer programs - later on - classifying unseen test data. So in a
to imitate human behaviour. While it is possible first first step we extract a variety of features from
to simulate human inadequacies, for example by the Cresci et al., 2017 datasets. These features are
delaying reactions, it would be difficult for bots to discussed in the next subsections.
accurately mimic human behavior: This would re- 4.4.1 Account Based Features
quire extensive statistical analysis of the behaviour
The first group of features is derived from account
of Twitter users. We can therefore assume that
metadata. Our simple user profile features di-
bots are created using simpler methods, such as
rectly reflect values the Twitter API provides about
random temporal behaviour, which only resemble
users. We additionally derive features with some
human behaviour at first glance.
processing from the screen and user names.
4.1.2 Purpose-Related Differences Some of the features are self explanatory or ex-
plained by the Twitter API documentation. 2 Nev-
Bots have clear objectives, for example spread-
ertheless some of these features require additional
ing political messages or references to products.
discussion.
Bots bring specific content to attention, hashtags,
URLs. So they have some kind of ”agenda” which • Simple user profile features: We hypothe-
should be able to be exploited to some extent in size that metadata from the user profile pro-
general. vides valuable information about the user ac-
It should be noted here that the fact that we and count. Some of this data is generated by
other researchers are able to identify bots quite re- 2
https://developer.twitter.com/en/
liably clearly shows that these assumptions are not docs/tweets/data-dictionary/overview/
unfounded. user-object.html
552
Twitter itself and sometimes difficult to con- particular, we determine which of the 105 Uni-
trol directly by users. This data contains char- code code groups an account uses in the screen
acteristics about a user’s account we can ex- and user names. This is reflected in the cat-
ploit for machine learning. These account egorical features user name unicode group and
features include: screen name unicode group where we have one
feature for every Unicode code group. The ratio-
– default profile: Has the user altered the
nale behind this feature is that humans tend to be
profile?
quite creative in their choice of names and some-
– geo enabled: This feature reflects if
times tend to pick characters completely unrelated
users enable adding geographic infor-
to the alphabet of their own language. By compar-
mation if they publish a tweet.
ing occurrences of characters in various Unicode
– protected: When true, indicates that this code groups we take this behaviour into account.
user has chosen to protect their Tweets. Furthermore we want to make use of possi-
– is verified: This is some kind of quality ble differences between an account’s screen name
marker provided by Twitter: Accounts and user name. We model this by calculating
that are run by people of public inter- the respective Levenshtein distance (Levenshtein,
est can be verified as being authentic by 1966). We observed that bot accounts tend to
Twitter itself. choose user names and screen names that are sim-
– friends count: The number of users this ilar, while humans can be more creative in this re-
account is following. spect.
– followers count: The number of follow-
4.4.2 Content Based Features
ers this account has.
We derive additional features from the content a
– favourites count: The number of tweets
user provides.
this user has liked.
For that purpose we tokenize the tweets. We to-
– listed count: The number of public lists kenize by breaking the tweet text at a variety of
this account is a member of. spaces and punctuation such as commas, colons,
– statuses count: The number of tweets exclamation marks, brackets and similar charac-
issued by this account. ters that are commonly used in sentences. This to-
– profile use background image: Has the kenization respects characters occuring in emojis
user provided a background image? as well. While tokenization in itself is not entirely
language agnostic, these heuristics should never-
• User profile name features: User and screen
theless work for a fairly large set of (Latin script)
names are very much subject to a user’s
languages.
choice. Therefore we hypothesize that it pro-
Then we can extract features based on these to-
vides valuable information that helps to dis-
kens. Other features are directly derived from the
tinguish bots from humans. These features
metadata of tweets.
include:
– screen name length: The length of the • Behavioural features: We hypothesize that
screen name provided by a user. the tweeting behaviour of bots and humans
should exhibit differences. We model this
– user name length: The length of the ac-
behaviour by calculating statistical properties
count name provided by a user.
from the data such as minimum, maximum,
– screen name digits: Number of digits in
average, mean, median, standard deviation,
the screen name.
skewness, kurtosis, and others.
– user name unicode group: See below.
– screen name unicode group: See be- – time between retweets (distributional
low. feature): This set of features models
time between retweeting activities of an
– levenshtein user name screen name:
account.
See below.
– time between tweets (distributional):
We use some features that are closely related This set of features models time
to the screen and user names of accounts. In between tweeting activities.
553
– tweet rate (average tweet rate): The av- periments.
erage number of tweets per day For reasons of getting a more detailed under-
standing about the performance of various feature
• Core content features: We hypothesize that
sets we group all behavioral features to a feature
some aspects of intention and emotions can
set named tweet-behav, all tweet content re-
be derived from a tweet’s content. The fol-
lated features to a set named tweet-cont and
lowing features honor this in a language in-
all account related features to account. We use
dependent way.
account-lev as a feature set that only includes
– emojis classic (distributional feature): the Levenshtein distance between user and screen
See below. names to test this feature specifically.
– emojis kaomji faces (distributional): As we want to compare our work to Kudugunta
See below. and Ferrara, 2018 we reimplented all their account
– emojis line art (distributional): See be- features mentioned in their paper. We later on refer
low. to this feature group with k f reimpl.
– emojis other (distributional): See be- Our own feature group all contains almost all
low. features described above leaving out only four mi-
– number of tokens (distributional): This nor features as they don’t seem to improve the
feature models the size of a tweet to quality of our classifier: Information about the
some extent. default profile, location, the content of the ”pro-
– number of hashtags (distributional): tected” field and information about user back-
The number of hashtag based references ground image.
contained in tweets. 5 Results and Discussion
– n of tokens wo hashtags urls symbols
(distributional): A distribution based We generate feature matrices based on different
on the number of plaintext tokens in the sets of features as described in section 4.5. This
tweets excluding hashtags, URLs and data is normalized by scaling each feature to the
non-alphanumeric tokens. unit interval and then used to train and evaluate
– number of urls (distributional): The different machine learning models.
number of URL based references con- To evaluate the performance of our machine
tained in tweets. learning classifiers we separate our dataset into a
– special char repeats rate: This feature training and validation set by an 80:20 ratio using
detects sequences of question marks and (deterministic) random selection provided by the
exclamation marks. These are also of- scikit-learn framework (Pedregosa et al., 2011).
ten used for affirmation to give special We have experimented with classical methods
weight to what has been expressed in a such as Logistic Regression3 , Support Vector Ma-
tweet. We want to model this aspect. chines4 , Random Forests5 and Multi-layer Percep-
trons6 . The best results were achieved with Ad-
We assume that a precise sentiment analysis of aBoost7 (Freund and Schapire, 1999). Since the
the twitter text is not available for all languages training labels are not fully balanced, we follow
tweets could be written in. Nevertheless, we want Kudugunta and Ferrara, 2018 and resampled the
to extract and use emotional aspects from the con- training data using SMOTE-ENN (Lemaı̂tre et al.,
tent. In order to model some basic aspects of emo- 2017).
tion we detect various different sets of emoticons By using AdaBoost with SMOTE-ENN we
here. The extraction is pattern based and derived achieve the results displayed in table 1. Figure 1
from public emoticon collections as provided by contains some of the corresponding ROC curves.
Wikipedia. The four emoji features above are The feature set tweet-behav contains all fea-
distributions derived on individual occurrences of tures related to the tweeting behavior of an account
emojis in single tweets. 3
sklearn.linear model.LogisticRegression
4
4.5 Feature Groups sklearn.svm.LinearSVC
5
sklearn.ensemble.RandomForestClassifier
Based on these features we build sets of different 6
sklearn.neural network.MLPClassifier
features for use in different machine learning ex- 7
sklearn.ensemble.AdaBoostClassifier
554
Feature-Set P R F1 ACC AUC ROC
tweet-behav 0.9920 0.8888 0.9376 0.9314 0.9475
tweet-cont 0.6413 0.9701 0.7721 0.6685 0.9295
account 0.9907 0.9835 0.9871 0.9851 0.9929
account-lev 0.9481 0.9041 0.9838 0.9159 0.9604
all 0.9958 0.9835 0.9896 0.9881 0.9959
k f reimpl 0.9886 0.9804 0.9845 0.9821 0.9935
k f-AdaBoost-SMOTEENN 1.00 1.00 1.00 0.9981 0.9981
Table 1: Classification performance of models using different feature sets. k f reimpl is our reimple-
mentation of our reference classifier. Though we reimplemented all features we were not able to confirm
their original results. These are provided in k f-AdaBoost-SMOTEENN for reference.
Table 2: The 20 most important features for the system ”all”, sorted by ANOVA F-value.
and tweet-cont contains all content features. Our this features alone (the account-lev feature set) it
experiments show that neither feature set is able to is possible to achieve an accuracy of 0.8611.
achieve sufficiently good results, as there is a large In order to compare our system we reimple-
number of false positives. mented a classification system as described by
It turned out that using the Levenshtein distance Kudugunta and Ferrara, 2018, where AdaBoost
as a feature provides a considerable benefit. It is with SMOTE-ENN is used to get excellent results.
the most informative feature in the account fea- However, although we use the same dataset and
ture set, followed by geo enabled, statuses count, the same features, we were not able to achieve
user name length, screen name length and the same results. We assume the reason for this
features related to account metadata, such is that the exact selection of their training data
as default profile, favourites count and pro- is unknown to us. For our own analysis we di-
file use background image. vide the gold standard data using deterministic
In order to get an idea of how well this feature random selection. Depending on the exact nature
performs we conducted an experiment using it as of this random selection, it is understandable that
the only classifier. We found that by using only each trained system based on this selection will
555
Receiver operating characteristic Receiver operating characteristic
1.0 1.0
0.8 0.8
True Positive Rate
0.4 0.4
0.2 0.2
(a) ROC curve for ”k f reimpl” (b) ROC curve for ”all”
556
Learning Curves (AdaBoostClassifier) Learning Curves (AdaBoostClassifier)
1.00 1.00 Training score
Cross-validation score
0.95 0.95
0.90 0.90
Score
Score
0.85 0.85
0.80 0.80
0.75 0.75
Training score
Cross-validation score
0.70 0.70
0 1000 2000 3000 4000 5000 6000 7000 0 1000 2000 3000 4000 5000 6000 7000
Training examples Training examples
(a) Learning curve for ”k f reimpl” (b) Learning curve for ”account-lev”
0.90 0.90
Score
Score
0.85 0.85
0.80 0.80
0.75 0.75
Training score
Cross-validation score
0.70 0.70
0 1000 2000 3000 4000 5000 6000 7000 0 25 50 75 100 125 150 175 200
Training examples Training samples
(c) Learning curve for ”all” (d) Learning curve for ”tweet-cont”
0.90 0.90
Score
Score
0.85 0.85
0.80 0.80
0.75 0.75
Training score
Cross-validation score
0.70 0.70
0 25 50 75 100 125 150 175 200 0 25 50 75 100 125 150 175 200
Training samples Training samples
(e) Learning curve for ”all” (f) Learning curve for ”tweet-behav”
557
References F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Pretten-
Adam Badawy, Emilio Ferrara, and Kristina Lerman. hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-
2018. Analyzing the Digital Traces of Political Ma- sos, D. Cournapeau, M. Brucher, M. Perrot, and
nipulation: The 2016 Russian Interference Twitter E. Duchesnay. 2011. scikit-learn: Machine learning
Campaign. arXiv e-prints page arXiv:1802.04291. in Python. Journal of Machine Learning Research
12:2825–2830.
Chiyu Cai, Linjing Li, and Daniel Zeng. 2017. Detect-
ing social bots by jointly modeling deep behavior A.H. Wang. 2010. Don’t follow me: Spam detection
and content information. In Proceedings of the 2017 in Twitter. In Proceedings of the 2010 International
ACM on Conference on Information and Knowledge Conference on Security and Cryptography, Athens,
Management, Singapore. pages 1995–1998. Greece. pages 1–10.
Zi Chu, Steven Gianvecchio, Haining Wang, and Sushil
Jajodia. 2012. Detecting Automation of Twitter Ac-
counts: Are You a Human, Bot, or Cyborg? IEEE
Transactions on Dependable and Secure Computing
9(6):811–824.
558
Multi-level analysis and recognition of the text sentiment on the example
of consumer opinions
Jan Kocoń Monika Zaśko-Zielińska Piotr Miłkowski
Wrocław University University of Wrocław Wrocław University
of Science and Technology Institute of Polish Studies of Science and Technology
Wrocław, Poland Wrocław, Poland Wrocław, Poland
jan.kocon monika.zasko-zielinska piotr.milkowski
@pwr.edu.pl @uwr.edu.pl @pwr.edu.pl
559
Proceedings of Recent Advances in Natural Language Processing, pages 559–567,
Varna, Bulgaria, Sep 2–4 2019.
100
ferently. It is advantageous in opinion mining to 90
20 -01
20 -09
20 -05
20 -01
20 -09
20 -05
20 -01
20 -09
20 -05
20 -01
20 -09
20 -01
20 -09
20 -05
20 -01
20 -09
20 -05
20 -01
20 -09
20 -05
20 -01
20 -09
5
20 -05
-0
project were published, which aimed to create
04
04
05
06
06
07
08
08
09
10
10
12
12
13
14
14
15
16
16
17
18
18
19
11
20
methods of analysing the content written on the customer feedback sentiment analysis
560
authentic material provided online by students as polarity (20% of texts). We observed that the value
a final assessment of the course in each subject. of inter-annotator agreement in aspect annotation
Each text was manually annotated by two anno- task was very low, below 15% of Positive Spe-
tators: a psychologist and a linguist, who worked cific Agreement, PSA (Hripcsak and Rothschild,
accordingly to the general guidelines. In the pilot 2005).
project we decided to deal with sentiment annota-
tion of the whole text. In the sentiment annotation 3.2 New annotation guidelines
we used the same set of tags which is applied in In the main stage of the project we decided to an-
plWordNet 3.0 emo (Zaśko-Zielińska et al., 2015; notate the sentiment for the whole text (a meta
Janz et al., 2017) to lexical units: [+m] – strong level) and the sentence level. We assumed that this
positive, [+s] – weak positive, [-m] – strong nega- strategy allows to establish the acceptable value
tive, [-s] – weak negative, [amb] – ambiguous, [0] of PSA. We followed the rule that the meta an-
– neutral. notation results partially from sentence annota-
We used amb tag but we understood it differ- tions, however the frame polarity is the main fac-
ently. In annotation of lexical units in WordNet tor for the final meta annotation. We have pre-
with sentiment amb indicated the possibility that pared the following rules of annotation, regardless
units can be positive or negative in various con- of whether the entire text or sentence is annotated:
texts. Hence, in text sentiment analysis we as-
sumed that amb denotes ambiguous polarity, thus • SP – strong positive – entirely positive;
the entire text cannot be clearly described by using
• WP – weak positive – generally positive, but
neither positive nor negative annotation.
there are some negative aspects;
In the annotation we focused primarily on the
strategic places in the text. In a customer review • 0 – neutral;
these places are the opening and closing sentences,
namely the text frame. The beginning consists of • WN – weak negative – generally negative, but
the general opinion of the author on the subject of there are some positive aspects;
the evaluation and the end includes the author’s
• SN – strong negative – entirely negative;
recommendation to the recipients. The annota-
tors created their first general evaluation based on • AMB – ambiguous – there are both positive
these two segments. In the body of the review, the and negative aspects in the text that are bal-
authors have only subtly changed these opinions. anced in terms of relevance.
Regardless of the modification of the main opinion
in the text, we did not use the amb tag when the Table 1 shows the value of Positive Specific
text frame was unambiguously positive. The text Agreement (Hripcsak and Rothschild, 2005) ob-
frame polarity was influenced not only by lexical tained for a random sample of 111 documents
content but also nonverbal elements, e.g. emoti- from Medicine category.
cons or multiplication of punctuation marks.
4 Multi-level sentiment recognition
3.1 An attempt to annotate aspects
We selected three different classifiers for the
The analysis of the content of customer reviews in recognition tasks:
our pilot project consisted of two stages: the selec-
tion of text blocks describing separate aspects and • logistic regression (fastText) providing a
their annotation. Some parts of the text were not of baseline for text classification (Joulin et al.,
an argumentative nature that could justify the au- 2017)
thor’s decision to polarise the text. They included:
• bidirectional long short-term memory recur-
advice (e.g. how to sign up for lectures) or general
rent network in two variants:
information on lectures, duration of classes, etc.
The main stage of our project was conducted – using word vector representations only
based on text corpus consisting of consumer re- – using the same vectors extended with
views (80% of texts) and texts from the corre- general polarity information from senti-
sponding domain with high probability of neutral ment dictionary described in Section 4.1
561
L Type Only A A&B Only B PSA (consisting of 1,024 hidden units using hy-
SN 1 33 4 93% perbolic tangent activation method) merged
WN 2 2 2 50% with concatenation
0 0 24 0 100%
M AMB 1 2 3 50% • Dropout layer with dropout ratio equal to 0.2
WP 4 0 0 0%
• Dense layer with number of outputs repre-
SP 0 31 2 97%
sum 8 92 11 91%
senting number of all possible labels (6 in our
task) using normalised exponential function
SN 10 217 36 90%
(softmax) activation
WN 11 1 0 15%
0 36 273 17 91% BERT was designed to provide pre-trained
S AMB 2 7 14 47% deep bidirectional representations conditioning
WP 12 0 1 0% left and right context (Devlin et al., 2018), there-
SP 6 194 8 97% fore it achieves best performance on text frag-
sum 77 692 76 90% ments instead of single sentences. It’s ar-
Table 1: Annotation agreement between two experts (A and chitecture allows to fine-tune these represen-
B) at the level (L) of text (meta – M) and sentence (S) for a tations by adding one additional output layer
sample of 111 documents using Positive Specific Agreement which suits needs of specified task. For
metric, PSA (Hripcsak and Rothschild, 2005).
our task as a pre-trained model BERT-Base,
Multilingual Cased6 was selected, which
• bidirectional encoder representations from consists of 104 languages and 110M parameters,
transformers (BERT) with addition of se- and BertForSequenceClassification7
quence classification layer as a BERT classifier extended for multi-class clas-
sification.
We trained logistic regression model using pre-
trained vectors for Polish language (Kocoń and 4.1 Embedding vector extension
Gawor, 2018). This approach is much faster in Basing on the data accommodated in plWord-
both training and testing than deep learning classi- Net emo (Zaśko-Zielińska et al., 2015) we pre-
fiers (Joulin et al., 2017), however, it has disadvan- pared the dictionary for all annotated lexical units
tage which comes from not sharing parameters by and all possible levels of sentiment. Due to the
features and classes, therefore overall result can be lack of word sense disambiguation method, we
highly influenced by keywords with bigger class grouped the sentiment annotations by lemmas.
relativity. The final dictionary consists of a set of lemmas
BiLSTM on the other side takes into considera- with assigned numbers representing the propor-
tion not just words but full text fragment and bas- tions of individual sentiment annotations, sum-
ing on learnt patterns predicts potential outcome. ming up to 1, e.g. for a lemma akademicki (Eng.
Texts are divided into tokens and converted to cor- academic) there were 11 annotations: 3 neutral, 4
responding word embedding vectors generated by generally negative, 3 generally positive and 1 en-
fastText (Bojanowski et al., 2017), in this form it tirely positive. Therefore arbitrary values for word
is possible to use it as input for neural network. "akademicki" are:
Dimension of used vectors is equal to 300, there-
fore it must be reflected in the input shape. As • entirely positive = 0.0909
a loss function for training a categorical crossen-
• generally positive = 0.2727
tropy was chosen. Model prepared for the task
consists of the following layers: • neutral = 0.2727
• Gaussian noise layer with standard devia- • generally negative = 0.3636
tion of 0.01 accepting as input shape up to
128 words with vector matrix for each word • entirely negative = 0.0000
of size 300, therefore overall input shape is 6
https://github.com/google-research/
(128, 300) bert
7
https://github.com/huggingface/
• Bidirectional layer with LSTM instances pytorch-pretrained-BERT
562
• ambivalent = 0.0000 sentence evaluation types. We use also prefixes
of domains (Hotels, Medicine, School, Products)
Using the described dictionary we have pro- as suffixes for SD* and DO* variants, e.g. SDS-
posed additional variant of BiLSTM classifier with H is a single domain evaluation type performed
a word embedding vector extended with the val- on sentences within hotels domain, whereas DOT-
ues of sentiment for the lemma of the word from M is a domain-out evaluation type performed on
a prepared sentiment dictionary. Lemmas were re- texts trained on texts outside medicine domain and
trieved during a preparation of the input data us- tested on texts from that domain.
ing WCRFT part-of-speech tagger (Radziszewski,
2013). Therefore, in this approach the input word Table 2 shows the number of texts and sentences
vector dimension was extended with 6 values rep- annotated by linguists for all evaluation types,
resenting sentiment of the word. The final dimen- with division into the number of elements within
sion of the word embedding increased from 300 to training, validation and test sets. Linguists anno-
306. tated a total of 8,450 texts from four domains (ho-
tels, medicine, products, school) and 35,789 sen-
5 Evaluation tences from two domains (hotels, medicine). The
distribution of labels within each domain for texts
As in article (Kocoń et al., 2019), three variants of and sentences is presented in Table 3. Average an-
evaluation of the sentiment classification methods notated text length in each domain are as follows:
were prepared. The basic variant is a single do- 788 characters in hotels, 802 in medicine, 781 in
main in which the classifier is trained, tuned and products and 442 in school.
tested on a set of texts from one domain. The next
variant includes an analysis of the ability of the
classifier to model the sentiment of the text on a Type Domain Train Dev Test SUM
level independent of the domain of the text. For Hotels 2534 316 316 3166
this purpose, we take all available texts except the Medicine 2650 330 330 3310
SDT
texts from the selected domain. Then the texts are Products 790 98 98 986
divided into a training and a validation set. Test- School 792 98 98 988
ing of the model takes place on a test set from !Hotels 4756 528 - 5284
!Medicine 4635 514 - 5149
a selected domain, not taken into account at the DOT
!Products 6727 746 - 7473
stage of preparing the training and validation set. !School 6725 746 - 7471
The third test variant allows to examine the clas- MDT All 6771 846 845 8462
sifiers in order to generalise the task of sentiment Hotels 12434 1554 1553 15541
analysis in all available domains. For this purpose, SDS
Medicine 16200 2024 2024 20248
texts from all domains are treated as one set, which !Hotels 16200 2024 - 18224
is randomly divided into train, validation and test DOS
!Medicine 12434 1554 - 13988
sets. Summary of the different types of evaluation: MDS All 28581 3572 3571 35724
• SD – Single Domain – evaluation sets created Table 2: The number of texts/sentences for each evaluation
using elements from the same domain, type in train/dev/test sets.
563
6 Results
Classifier
Table 4 presents the values of F1-score for each
macro
micro
AMB
Type
WN
WP
SN
SP
F1
label (columns 3-8), global F1-score (column 9),
0
C1 80.00 30.51 93.98 00.00 83.33 36.84 73.50 90.87 70.20
SDT-H
micro-AUC and macro-AUC (columns 10-11) for C2
C3
71.11
72.82
25.00
22.95
00.00
94.12
04.76
14.81
72.44
81.98
00.00
27.27
53.00
68.45
77.14
89.83
66.13
72.44
all evaluation types related to the texts. In case C4 71.22 10.26 96.39 00.00 78.16 00.00 68.45 91.30 72.41
C1 81.05 15.38 96.39 00.00 80.63 00.00 78.55 93.44 66.39
of evaluation for a single domain for each label,
SDT-M
C2 78.69 11.11 95.71 14.29 80.31 06.67 77.04 94.02 71.81
fastText (using Logistic Regression) outperformed C3
C4
81.93
00.00
13.33
00.00
95.71
95.65
13.33
00.00
80.43
62.33
07.41
00.00
78.55
58.01
91.81
89.42
69.33
59.73
other classifiers in 13 out of 21 distinguishable C1 62.86 27.59 00.00 36.36 84.68 16.67 65.66 86.76 63.58
SDT-P
C2 28.57 30.30 00.00 00.00 69.16 26.67 49.49 78.37 53.46
cases. There are 12 cases for which the best score C3 25.00 00.00 00.00 00.00 67.26 00.00 43.43 77.64 51.46
C4 00.00 00.00 00.00 00.00 69.74 00.00 53.54 80.40 42.60
is not higher than F1=0.4. These are highly un- C1 79.61 52.63 00.00 00.00 50.00 00.00 61.62 83.80 62.33
derrepresented labels, for which the part of the to-
SDT-S
C2 72.22 33.33 00.00 00.00 27.27 36.36 52.53 80.39 54.40
C3 75.68 34.04 00.00 00.00 11.76 00.00 51.52 79.71 54.67
tal annotations within the domain is less than 10% C4 68.87 00.00 00.00 00.00 00.00 00.00 52.53 83.88 49.55
C1 70.91 23.08 95.24 00.00 83.49 21.05 69.09 88.21 66.34
(see Table 3). The best results are obtained for
DOT-H
C2 72.73 17.02 91.76 15.38 78.76 16.00 65.30 88.31 71.21
strong positive and strong negative cases. Inter- C2
C4
73.94
75.53
19.67
34.09
88.89
90.67
10.00
00.00
75.11
82.76
16.00
00.00
62.46
68.14
87.41
91.47
70.59
72.70
mediate labels (weak and ambiguous variants) are C1 72.51 08.70 86.67 17.39 75.29 00.00 69.18 86.13 68.37
DOT-M
C2 73.17 22.22 85.14 28.57 76.79 16.33 68.28 89.99 71.46
much more difficult to be recognised correctly. In C3 46.01 03.48 00.00 00.00 00.00 00.00 22.36 59.76 51.52
these cases deep neural networks outperform lo- C4
C1
72.11
60.61
10.71
47.06
90.32
00.00
16.67
50.00
85.06
76.00
12.90
00.00
72.51
59.60
90.88
89.71
73.14
70.29
gisitic regression in 6 out of 11 cases. BERT clas- DOT-P
C2 63.16 26.09 00.00 22.22 84.91 11.76 62.63 83.19 61.73
C3 54.55 28.57 00.00 00.00 85.71 00.00 59.18 83.81 63.25
sifier performs much better (13 out of 23 cases) C4 70.97 62.07 00.00 00.00 92.73 28.57 73.74 90.77 71.38
C1 68.00 12.50 00.00 13.33 36.36 00.00 43.43 90.12 68.67
in cross-domain knowledge transfer (DOT and
DOT-S
Table 5 presents results corresponding to those Table 4: F1-scores for text-oriented evaluation. Training sets
presented in Table 4, but this time for sentence- for evaluation types are the same as in Table 2 rows 1-9. Clas-
level annotations. Looking at Table 3, the number sifiers: C1 - fastText, C2 - BiLSTM, C3 - BiLSTM with word
embeddings extended using polarity dictionary, C4 - BERT.
of sentences marked as weakly positive or weakly Evaluation types are explained in Section 5.
negative is close to zero. These labels are not be-
ing recognised by any classifier. For other labels,
regardless of the type of evaluation, the best results
are mainly obtained using deep learning methods
Classifier
WN
WP
SN
SP
F1
0
rics: 12 out of 15 cases). C1 85.64 00.00 77.54 00.00 83.59 16.44 81.60 94.39 65.19
SDS-H
monitoring feedback, reactions and emotions are Table 5: F1-scores for sentence-oriented evaluation. Training
of great value as they fuel decisions and behav- sets for evaluation types are the same as in Table 2 rows 10-
iors (Tversky and Kahneman, 1989). However, 14. Classifiers: C1 - fastText, C2 - BiLSTM, C3 - BiLSTM
with word embeddings extended using polarity dictionary, C4
most of the existing solutions are still limited to - BERT. Evaluation types are explained in Section 5.
manual annotations and simplified analysis meth-
ods.
564
BERT’s performance is below the expectations
of this advanced method. Looking at both tables (4
and 5), BERT’s results are the best in 19 out of 69
label-specific cases, which is exactly as many as
fastText was. BiLSTM outperformed other meth-
ods in 31 cases. Adding an external sentiment
dictionary helped only in 14 label-specific cases.
BERT dominance is observed in DOT and MDT
cases, especially when analysing general metric
values, where the predominance of the method is
visible in 11 out of 15 cases. The advantage is
repeated for MDS but not for DOS. MDT case
is the most promising in terms of the further use
of the recognition method in applications such as
brand monitoring or early crisis detection. Fig-
ure 2 shows the ROC curves (Meistrell, 1990) for
this case. The values of the general F1, micro
AUC and macro AUC are the highest for the BERT
method (see Table 2).
We plan to publish the data created as part of the
presented works on an open license soon. We also
intend to test the contextualized embedding that
we are currently building using the ELMo deep
word representations method (Peters et al., 2018),
with the use of the large KGR10 corpus presented
in work (Kocoń et al., 2019). We also want to train
the basic BERT model with the use of KGR10 to
investigate whether it will improve the quality of
sentiment recognition. It is also very interesting
to use the propagation of sentiment annotation in
WordNet (Kocoń et al., 2018a,b), to increase the
coverage of the sentiment dictionary and to po-
tentially improve the recognition quality as well.
This objective can be achieved by other complex
methods such as OpenAI GPT-2 (Radford et al.,
2019) and domain dictionaries construction meth-
ods utilising WordNet (Kocoń and Marcińczuk,
2016).
Acknowledgements
Co-financed by the Polish Ministry of Educa-
tion and Science, CLARIN-PL Project and by the
National Centre for Research and Development,
Poland, grant no POIR.01.01.01-00-0472/16 –
Sentimenti8 .
References
Stefano Baccianella, Andrea Esuli, and Fabrizio Sebas-
tiani. 2010. Sentiwordnet 3.0: an enhanced lexical Figure 2: Receiver Operating Characteristic curves of all used
classifiers for Mixed Domain Text setting.
8
http://w3a.pl/projekty/
565
resource for sentiment analysis and opinion mining. Jan Kocoń, Arkadiusz Janz, and Maciej Piasecki.
In LREC. volume 10, pages 2200–2204. 2018b. Context-sensitive Sentiment Propagation in
WordNet. In Proceedings of the 9th International
Piotr Bojanowski, Edouard Grave, Armand Joulin, and Global Wordnet Conference (GWC’18).
Tomas Mikolov. 2017. Enriching word vectors with
subword information. Transactions of the Associa- Jan Kocoń, Arkadiusz Janz, Miłkowski Piotr, Monika
tion for Computational Linguistics 5:135–146. Riegel, Małgorzata Wierzba, Artur Marchewka, Ag-
nieszka Czoska, Damian Grimling, Barbara Konat,
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Konrad Juszczyk, Katarzyna Klessa, and Maciej Pi-
Kristina Toutanova. 2018. Bert: Pre-training of deep asecki. 2019. Recognition of emotions, polarity
bidirectional transformers for language understand- and arousal in large-scale multi-domain text reviews.
ing. arXiv preprint arXiv:1810.04805 . In Zygmunt Vetulani and Patrick Paroubek, editors,
Human Language Technologies as a Challenge for
Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Computer Science and Linguistics, Wydawnictwo
2011. Domain adaptation for large-scale sentiment Nauka i Innowacje, Poznań, Poland, pages 274–280.
classification: A deep learning approach. In Pro-
ceedings of the 28th international conference on ma- Jan Kocoń and Michał Marcińczuk. 2016. Generating
chine learning (ICML-11). pages 513–520. of Events Dictionaries from Polish WordNet for the
Recognition of Events in Polish Documents. In Text,
Emitza Guzman and Walid Maalej. 2014. How do Speech and Dialogue, Proceedings of the 19th In-
users like this feature? a fine grained sentiment ternational Conference TSD 2016. Springer, Brno,
analysis of app reviews. In 2014 IEEE 22nd inter- Czech Republic, volume 9924 of Lecture Notes in
national requirements engineering conference (RE). Artificial Intelligence.
IEEE, pages 153–162.
William C Mann and Sandra A Thompson. 1988.
George Hripcsak and Adam S. Rothschild. 2005. Tech- Rhetorical structure theory: Toward a functional the-
nical Brief: Agreement, the F-Measure, and Relia- ory of text organization. Text-Interdisciplinary Jour-
bility in Information Retrieval. JAMIA 12(3):296– nal for the Study of Discourse 8(3):243–281.
298. https://doi.org/10.1197/jamia.M1733.
Michael L Meistrell. 1990. Evaluation of neural net-
Arkadiusz Janz, Jan Kocoń, Maciej Piasecki, and work performance by receiver operating character-
Monika Zaśko-Zielińska. 2017. plWordNet as a Ba- istic (roc) analysis: examples from the biotechnol-
sis for Large Emotive Lexicons of Polish. In LTC’17 ogy domain. Computer Methods and Programs in
8th Language and Technology Conference. Fundacja Biomedicine 32(1):73–80.
Uniwersytetu im. Adama Mickiewicza w Poznaniu,
Poznań, Poland. Alexander Pak and Patrick Paroubek. 2010. Twitter as
a corpus for sentiment analysis and opinion mining.
Armand Joulin, Edouard Grave, Piotr Bojanowski, and In LREc. volume 10, pages 1320–1326.
Tomas Mikolov. 2017. Bag of tricks for efficient
text classification. In Proceedings of the 15th Con- Bo Pang and Lillian Lee. 2004. A sentimental educa-
ference of the European Chapter of the Associa- tion: Sentiment analysis using subjectivity summa-
tion for Computational Linguistics: Volume 2, Short rization based on minimum cuts. In Proceedings of
Papers. Association for Computational Linguistics, the 42nd annual meeting on Association for Compu-
pages 427–431. tational Linguistics. Association for Computational
Linguistics, page 271.
Jaap Kamps, Maarten Marx, Robert J Mokken,
Maarten De Rijke, et al. 2004. Using wordnet Matthew Peters, Mark Neumann, Mohit Iyyer, Matt
to measure semantic orientations of adjectives. In Gardner, Christopher Clark, Kenton Lee, and Luke
LREC. Citeseer, volume 4, pages 1115–1118. Zettlemoyer. 2018. Deep contextualized word rep-
resentations. In Proceedings of the 2018 Con-
Rob Kitchin. 2014. The data revolution: Big data, ference of the North American Chapter of the
open data, data infrastructures and their conse- Association for Computational Linguistics: Hu-
quences. Sage. man Language Technologies, Volume 1 (Long
Papers). Association for Computational Linguis-
Jan Kocoń and Michał Gawor. 2018. Eval- tics, New Orleans, Louisiana, pages 2227–2237.
uating KGR10 Polish word embeddings in https://doi.org/10.18653/v1/N18-1202.
the recognition of temporal expressions us-
ing BiLSTM-CRF. Schedae Informaticae 27. Alec Radford, Jeffrey Wu, Rewon Child, David Luan,
https://doi.org/10.4467/20838476SI.18.008.10413. Dario Amodei, and Ilya Sutskever. 2019. Language
models are unsupervised multitask learners. OpenAI
Jan Kocoń, Arkadiusz Janz, and Maciej Piasecki. Blog page 8.
2018a. Classifier-based Polarity Propagation in a
Wordnet. In Proceedings of the 11th International Adam Radziszewski. 2013. A tiered CRF tagger for
Conference on Language Resources and Evaluation Polish. In Intelligent tools for building a scientific
(LREC’18). information platform, Springer, pages 215–230.
566
Maite Taboada, Kimberly Voll, and Julian Brooke.
2008. Extracting sentiment as a function of dis-
course structure and topicality. Simon Fraser Uni-
veristy School of Computing Science Technical Re-
port .
Hiroya Takamura, Takashi Inui, and Manabu Okumura.
2005. Extracting semantic orientations of words us-
ing spin model. In Proceedings of the 43rd An-
nual Meeting on Association for Computational Lin-
guistics. Association for Computational Linguistics,
pages 133–140.
Amos Tversky and Daniel Kahneman. 1989. Ratio-
nal choice and the framing of decisions. In Multiple
Criteria Decision Making and Risk Analysis Using
Microcomputers, Springer, pages 81–126.
Byron C Wallace, Eugene Charniak, et al. 2015.
Sparse, contextually informed models for irony de-
tection: Exploiting user communities, entities and
sentiment. In Proceedings of the 53rd Annual Meet-
ing of the Association for Computational Linguistics
and the 7th International Joint Conference on Natu-
ral Language Processing (Volume 1: Long Papers).
volume 1, pages 1035–1044.
567
A Qualitative Evaluation Framework for Paraphrase Identification
568
Proceedings of Recent Advances in Natural Language Processing, pages 568–577,
Varna, Bulgaria, Sep 2–4 2019.
et al., 2012; Ji and Eisenstein, 2013) to end-to-end an approach is that it only requires modification
Deep Learning models (He et al., 2015; He and and additional annotation of the corpus. It does
Lin, 2016; Wang et al., 2016; Lan and Xu, 2018a; not place any additional burden on the developers
Kiros et al., 2015; Conneau et al., 2017). The PI of the systems and can be applied to multiple sys-
systems are typically divided in two groups: Su- tems without additional cost.
pervised PI systems and Unsupervised PI systems. We follow a similar line of research and propose
“Supervised PI systems” (He et al., 2015; He a new evaluation that uses ETPC (Kovatchev et al.,
and Lin, 2016; Wang et al., 2016; Lan and Xu, 2018a): a PI corpus with a multi-layer annotation
2018a) are explicitly trained for the PI task on a of various linguistic phenomena. Our methodol-
PI corpora. “Unsupervised PI systems” in the PI ogy uses the corpus annotation to provide much
field is a term used for systems that use a general more feedback to the competing systems and to
purpose sentence representations such as Mikolov evaluate and compare them qualitatively.
et al. (2013); Pennington et al. (2014); Kiros et al.
(2015); Conneau et al. (2017). To predict the para- 3 Qualitative Evaluation Framework
phrasing relation, they can compare the sentence
3.1 The ETPC Corpus
representations of the candidate paraphrases di-
rectly (ex.: cosine of the angle), and use a PI cor- ETPC (Kovatchev et al., 2018a) is a re-annotated
pus to learn a threshold. Alternatively they can use version of the MRPC corpus. It contains 5,801 text
the representations as features in a classifier. pairs. Each text pair in ETPC has two separate lay-
The complexity of paraphrasing has been em- ers of annotation. The first layer contains the tradi-
phasized by many researchers (Bhagat and Hovy, tional binary label (paraphrase or non-paraphrase)
2013; Vila et al., 2014; Benikova and Zesch, of every text pair. The second layer contains the
2017). Similar observations have been made for annotation of 27 “atomic” linguistic phenomena
Textual Entailment (Sammons et al., 2010; Cabrio involved in paraphrasing, according to the authors
and Magnini, 2014). Gold et al. (2019) study the of the corpus. All phenomena are linguistically
interactions between paraphrasing and entailment. motivated and humanly interpretable.
Despite the complexity of the phenomena, the
3a A federal magistrate in Fort Lauderdale
popular PI corpora (Dolan et al., 2004; Ganitke-
ordered him held without bail.
vitch et al., 2013; Iyer et al., 2017; Lan et al., 2017)
are annotated in a binary manner. In part it is due 3b He was ordered held without bail Wednesday
to lack of annotation tools capable of fine-grained by a federal judge in Fort Lauderdale, Fla.
annotation of relations. WARP-Text (Kovatchev
et al., 2018b) fills this gap in the NLP toolbox. We illustrate the annotation with examples 3a
The simplified corpus format poses a problem and 3b. At the binary level, this pair is annotated as
with respect to the quality of the PI task and the “paraphrases”. At the “atomic” level, ETPC con-
ways it can be evaluated. The vast majority of the tains the annotation of multiple phenomena, such
state-of-the-art systems in PI provide no or very as the “same polarity substitution (habitual)” of
little error analysis. This makes it difficult to in- “magistrate” and “judge” (marked bold) or the
terpret the actual capabilities of a system and its “diathesis alternation” of “...ordered him held”
applicability to other corpora and tasks. and “he was ordered by...” (marked underline).
Some researchers have approached the problem For the full set of phenomena, the linguistic rea-
of non-interpretability by evaluating the same ar- soning behind them, their frequency in the cor-
chitecture on multiple datasets and multiple tasks. pus, real examples from the pairs, and the anno-
Lan and Xu (2018b) apply this approach to Su- tation guidelines, please refer to Kovatchev et al.
pervised PI systems, while Aldarmaki and Diab (2018a).
(2018) use it for evaluating Unsupervised PI sys-
tems and general sentence representation models. 3.2 Evaluation Methodology
Linzen et al. (2016) demonstrate how by modi- We use the corpus to evaluate the capabilities of
fying the task definition and the evaluation the ca- the different PI systems implicitly. That means,
pabilities of a Deep Learning system can be de- the training objective of the systems remains un-
termined implicitly. The main advantage of such changed: they are required to correctly predict
569
the value of the binary label at the first annota- RQ 2 Which are the strong and weak sides of each
tion layer. However, when we analyze and evalu- individual system?
ate the performance of the systems, we make use
of both the binary and the atomic annotation lay- RQ 3 Are there any significant differences between
ers. Our evaluation framework is created to ad- the “performance profiles” of the systems?
dress our main research question (RQ 1):
RQ 4 Are there phenomena on which all systems
perform well (or poorly)?
RQ 1 Does the performance of a PI system on each
candidate-paraphrase pair depend on the dif- 4 PI Systems
ferent phenomena involved in that pair?
To demonstrate the advantages of our evaluation
We evaluate the performance of the systems in framework, we have replicated several popular Su-
terms of their “overall performance” (Accuracy pervised and Unsupervised PI systems. We have
and F1) and “phenomena performance”. selected the systems based on three criteria: pop-
“Phenomena performance” is a novelty of our ularity, architecture, and performance. The sys-
approach and allows for qualitative analysis and tems that we chose are popular and widely used
comparison. To calculate “phenomena perfor- not only in PI, but also in other tasks. The systems
mance”, we create 27 subsets of the test set, one use a wide variety of different ML architectures
for each linguistic phenomenon. Each of the sub- and/or different features. Finally, the systems ob-
sets consists of all text pairs that contain the cor- tain comparable quantitative results on the PI task.
2
responding phenomenon . Then, we use each of They have also been reported to obtain good re-
the 27 subsets as a test set and we calculate the sults on the MRPC corpus which is the same size
binary classification Accuracy (paraphrase or non- as ETPC. The choice of system allows us to best
paraphrase) for each subset. This score indicates demonstrate the limitations of the classical quan-
how well the system performs in cases that include titative evaluation and the advantages of the pro-
one specific phenomenon. We compare the perfor- posed qualitative evaluation.
mance of the different phenomena and also com- To ensure comparability, all systems have been
pare them with the “overall performance”. trained and evaluated on the same computer and
the same corpus. We have used the configurations
Prior to running the experiments we verified
recommended in the original papers where avail-
that: 1) the relative distribution of the phenom-
able. During the replication we did not do a full
ena in paraphrases and in non-paraphrases is very
grid-search as we want to replicate and thereby
similar; and 2) there is no significant correlation
contribute to generalizable research and systems.
(Pearson r <0.1) between the distributions of the
As such, the quantitative results that we obtain
individual phenomena. These findings show that
may differ from the performance reported in the
the sub-tasks are non-trivial: 1) the binary labels
original papers, especially for the Supervised sys-
of the pairs cannot be directly inferred by the pres-
tems. However, the results are sufficient for the
ence or absence of phenomena; and 2) the different
objective of this paper: to demonstrate the advan-
subsets of the test set are relatively independent
tages of the proposed evaluation framework.
and the performance on them cannot be trivially
reduced to overlap and phenomena co-occurrence. We compare the performance of five Supervised
and five Unsupervised systems on the PI task,
The “overall performance” and “phenomena
including one Supervised and one Unsupervised
performance” of a system compose its “perfor-
baseline systems. We also include Google BERT
mance profile”. With it we aim to address the rest
(Devlin et al., 2018) for reference.
of our research questions (RQs):
The Supervised PI systems include:
2
i.e. The “diathesis alternation” subset contains all pairs [S1] Machine translation evaluation metrics as
that contain the “diathesis alternation” phenomenon (such as hand-crafted features in a Random Forest classi-
the example pair 3a–3b). Some of the pairs can also con-
tain multiple phenomena: the example pair 3a–3b contains
fier. Similar to Madnani et al. (2012) (baseline)
both “same polarity substitution (habitual)” and “diathesis [S2] A replication of the convolutional network
alternation”. Therefore pair 3a–3b will be added both to the similarity model of He et al. (2015)
“same polarity substitution (habitual)” and to the “diathesis
alternation” phenomena subsets. Consequentially, the sum [S3] A replication of the lexical composition
of all subsets exceeds the size of the test set. and decomposition system of Wang et al. (2016)
570
[S4] A replication of the pairwise word interac- ID System Description Acc F1
tion modeling with deep neural network system by SUPERVISED SYSTEMS
He and Lin (2016) 1 MTE features (baseline) .74 .819
[S5] A character level neural network model by 2 He et al. (2015) .75 .826
Lan and Xu (2018a) 3 Wang et al. (2016) .76 .833
The Unsupervised PI systems include: 4 He and Lin (2016) .76 .827
[S6] A binary Bag-of-Word sentence represen- 5 Lan and Xu (2018a) .70 .800
tation (baseline) UNSUPERVISED SYSTEMS
[S7] Average over sentence of pre-trained 6 Bag-of-Words (baseline) .68 .790
Word2Vec word embeddings (Mikolov et al., 7 Word2Vec (average) .70 .805
2013) 8 GLOVE (average) .72 .808
[S8] Average over sentence of pre-trained Glove 9 InferSent .75 .826
word embeddings (Pennington et al., 2014) 10 Skip-Thought .73 .816
[S9] InferSent sentence embeddings (Conneau 11 Google BERT .84 .889
et al., 2017)
[S10] Skip-Thought sentence embeddings Table 1: Overall Performance of the Evaluated Systems
(Kiros et al., 2015)
In the unsupervised setup we first represent
each of the two sentences under the corresponding comparison of the different systems. However, it
model. Then we obtain a feature vector by con- also has several limitations.
catenating the absolute distance and the element- It does not provide much insight into the work-
wise multiplication of the two representations. ings of the systems and does not facilitate error
The feature vector is then fed into a logistic regres- analysis. In order to study and improve the per-
sion classifier to predict the textual relation. This formance of a system, a developer has to look at
setup has been used in multiple PI papers, more re- every correct and incorrect predictions and search
cently by Aldarmaki and Diab (2018). While the for custom defined patterns. The “overall perfor-
vector representations of BERT are unsupervised, mance” is also not very informative for a compar-
they are fine-tuned on the dataset. Therefore we ison between the systems. For example S3 (Wang
put them in a separate category (System #11). et al., 2016) and S4 (He and Lin, 2016) obtain the
same Accuracy score and only differ by 0.06 F1
5 Results score. With only looking at the quantitative eval-
uation it is unclear which of these systems would
5.1 Overall Performance generalize better on a new dataset.
Table 1 shows the “overall performance” of the
5.2 Full Performance Profile
systems on the 1725 text pairs in the test set.
Looking at the table, we can observe several reg- Table 2 shows the full “performance profile” of
ularities. First, the deep systems outperform the S3 (Wang et al., 2016), the supervised system
baselines. Second, the baselines that we choose that performed best in terms of “overall perfor-
are competitive and obtain high results. Since both mance”. Table 2 shows a large variation of the per-
baselines make their predictions based on lexical formance of S3 on the different phenomena. The
similarity and overlap, we can conclude that the accuracy ranges from .33 to 1.0. We also report the
dataset is biased towards those phenomena. Third, statistical significance of the difference between
the supervised systems generally outperform the the correct and incorrect predictions for each phe-
unsupervised ones, but without running a full grid- nomena and the correct and incorrect predictions
search the difference is relatively small. And fi- for the full test set, using the Mann–Whitney U-
nally, we can identify the best performing systems: test3 (Mann and Whitney, 1947).
S3 (Wang et al., 2016) for the supervised and S9 Ten of the phenomena show significant differ-
(Conneau et al., 2017) for the unsupervised. BERT ence from the overall performance at p <0.1. Note
largely outperforms all other systems. 3
The Mann–Whitney U-test is a non-parametric equiva-
The “overall performance” provides a good lence of T-test. The U-Test does not assume normal distribu-
overview of the task and allows for a quantitative tion of the data and is better suited for small samples.
571
OVERALL PERFORMANCE The demonstrated variance in phenomena per-
Overall Accuracy .76 formance and its statistical significance address
Overall F1 .833 RQ 1: we show that the performance of a PI sys-
PHENOMENA PERFORMANCE tem on each candidate-paraphrase pair depends on
the different phenomena involved in that pair or at
Phenomenon Acc p
least there is a strong observable relation between
Morphology-based changes the performance and the phenomena.
Inflectional changes .79 .21 The individual “performance profile” also ad-
Modal verb changes .90 .01 dresses RQ 2. The profile is humanly inter-
Derivational changes .72 .22 pretable, and we can clearly see how the system
Lexicon-based changes performs on various sub-tasks at different linguis-
Spelling changes .88 .01 tic levels. The qualitative evaluation shows that S3
Same polarity sub. (habitual) .78 .18 performs better when it has to deal with: 1) surface
phenomena such as “spelling changes”, “punctu-
Same polarity sub. (contextual) .75 .37
ation changes”, and “change of order”; 2) dic-
Same polarity sub. (named ent.) .73 .14
tionary related phenomena such as “opposite po-
Change of format .75 .44
larity substitution (habitual)”, “converse substi-
Lexico-syntactic based changes tution”, and “modal verb changes”. S3 performs
Opp. polarity sub. (habitual) 1.0 na worse when facing phenomena such as “negation
Opp. polarity sub. (context.) .68 .14 switching”, “ellipsis”, “opposite polarity substi-
Synthetic/analytic substitution .77 .39 tution (contextual)”, and “addition/deletion”.
Converse substitution .92 .07
Syntax-based changes 5.3 Comparing Performance Profiles
Diathesis alternation .83 .12 Table 3 shows the full performance profiles of all
Negation switching .33 na systems. The systems are identified by their IDs,
Ellipsis .64 .07 as shown in Table 1. In addition to providing a
Coordination changes .77 .47 better error analysis for every individual system,
Subordination and nesting .86 .01 the “performance profiles” of the different sys-
Discourse-based changes tems can be used to compare them qualitatively.
This comparison is much more informative than
Punctuation changes .87 .01
the “overall performance” comparison shown in
Direct/indirect style .76 .5
Table 1. Using the “performance profile”, we can
Syntax/discourse structure .83 .05
quickly compare the strong and weak sides of the
Other changes different systems.
Addition/Deletion .70 .05 When looking at the “overall performance”, we
Change of order .81 .04 already pointed out that S3 (Wang et al., 2016)
Contains negation .78 .32 and S4 (He and Lin, 2016) have almost identical
Semantic (General Inferences) .80 .21 quantitative results: 0.76 accuracy, 0.833 F1 for
Extremes S3 against 0.76 accuracy, 0.827 F1 for S4. How-
Identity .77 .29 ever, when we compare their “phenomena per-
Non-Paraphrase .81 .04 formance” it is evident that, while these systems
Entailment .76 .5 make approximately the same number of correct
and incorrect predictions, the actual predictions
and errors can vary.
Table 2: Performance profile of Wang et al. (2016)
Looking at the accuracy, we can see that S3
performs better on phenomena such as “Con-
that eight of them are also significant at p <0.05. verse substitution”, “Diathesis alternation”, and
The statistical significance of “Opposite polarity “Non-Paraphrase”, while S4 performs better on
substitution (habitual)”, and “Negation Switch- “Change of format”, “Opposite polarity substitu-
ing” cannot be verified due to the relatively low tion (contextual)”, and “Ellipsis”.
frequency of the phenomena in the test set. We performed McNemar paired test comparing
572
PHENOMENON PARAPHRASE IDENTIFICATION SYSTEMS
SUPERVISED UNSUPERVISED
S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11
OVERALL ACC. .74 .75 .76 .76 .70 .68 .70 .72 .75 .73 .84
Inflectional .77 .76 .79 .79 .75 .79 .75 .76 .78 .80 .84
Modal verb .84 .89 .90 .89 .91 .92 .89 .84 .81 .89 .92
Derivational .80 .83 .72 .73 .84 .80 .88 .86 .80 .77 .87
Spelling .85 .83 .88 .90 .89 .85 .89 .88 .85 .89 .94
Same pol. sub. (hab.) .74 .77 .78 .76 .76 .76 .76 .75 .76 .76 .85
Same pol. sub. (con.) .74 .74 .75 .74 .70 .71 .71 .71 .73 .73 .81
Same pol. sub. (NE) .74 .72 .73 .75 .64 .67 .65 .70 .73 .66 .80
Change of format .80 .79 .75 .84 .85 .82 .81 .80 .80 .71 .91
Opp. pol. sub. (hab.) 1.0 1.0 1.0 .50 1.0 1.0 1.0 1.0 1.0 1.0 1.0
Opp. pol. sub. (con.) .77 .84 .68 .84 .52 .84 .61 .77 .65 .52 .71
Synthetic/analytic sub. .73 .73 .77 .77 .74 .70 .72 .71 .73 .74 .83
Converse substitution .93 .93 .92 .86 .93 .86 .79 .79 .93 .79 .86
Diathesis alternation .77 .85 .83 .77 .83 .89 .85 .83 .84 .81 .85
Negation switching 1.0 .67 .33 .33 .33 .67 .33 .67 .33 .67 .33
Ellipsis .77 .71 .64 .74 .80 .65 .81 .74 .61 .71 .81
Coordination .92 .92 .77 .92 .77 .92 .85 .85 .92 .92 .92
Subord. & nesting .83 .84 .86 .84 .81 .81 .85 .86 .80 .85 .93
Punctuation .88 .90 .87 .87 .86 .87 .89 .89 .89 .88 .93
Direct/indirect style .84 .84 .76 .80 .76 .80 .80 .84 .80 .80 .92
Syntax/disc. struct. .80 .83 .83 .81 .78 .81 .80 .80 .76 .78 .82
Addition/Deletion .69 .68 .70 .72 .67 .64 .65 .66 .70 .67 .82
Change of order .82 .83 .81 .81 .77 .82 .82 .82 .83 .84 .89
Contains negation .78 .74 .78 .79 .78 .72 .74 .78 .75 .76 .85
Semantic (Inferences) .80 .89 .80 .81 .88 .90 .90 .92 .76 .79 .90
Identity .74 .75 .77 .77 .73 .72 .73 .73 .76 .74 .85
Non-Paraphrase .76 .77 .81 .75 .71 .55 .67 .68 .77 .79 .88
Entailment .80 .80 .76 .76 .88 .80 .84 .88 .92 .88 .76
573
Figure 1: Critical Difference diagram of the average ranks by phenomena
a larger quantitative difference. Table 5 shows value. Each phenomenon is plotted with its av-
the comparison between S3 and S5 (Lan and Xu, erage rank across the 11 evaluated systems. The
2018a). Ten of the phenomena show significant horizontal lines connect phenomena which rank is
difference with p <0.1 and seven with p <0.05. within CD of each other. Phenomena which are
These results answer our RQ 3: we show that not connected by a horizontal line have signifi-
there are significant differences between the “per- cantly different ranking. We can observe that each
formance profiles” of the different systems. phenomenon is significantly different from at least
half of the other phenomena.
5.4 Comparing Performance by Phenomena We can observe that some phenomena, such
The “phenomena performance” of the individual as “opposite polarity substitution (habitual)”,
systems clearly differ among them, but they also “punctuation changes”, “spelling”, “modal verb
show noticeable tendencies. Looking at the per- changes”, and “coordination changes” are sta-
formance by phenomena, it is evident that certain tistically much easier according to our eval-
phenomena consistently obtain lower than aver- uation, as they are consistently among the
age accuracy across multiple systems while other best performing phenomena across all systems.
phenomena consistently obtain higher than aver- Other phenomena, such as “negation switching”,
age accuracy. “addition/deletion”, “same polarity substitution
In order to quantify these observations and (named entity)”, “opposite polarity substitution
to confirm that there is a statistical significance (contextual)”, and “ellipsis” are statistically much
we performed Friedman-Nemenyi test (Demšar, harder, as they are consistently among the worst
2006). For each system, we ranked the perfor- performing phenomena across all systems. With
mance by phenomena from 1 to 27, accounting the exception of “negation switching” and “op-
for ties. We calculated the significance of the dif- posite polarity substitution (habitual)”, these phe-
ference in ranking between the phenomena using nomena occur in the corpus with sufficient fre-
the Friedman test (Friedman, 1940) and obtained quency. These results answer our RQ 4: we
a Chi-Square value of 198, which rejects the null show that there are phenomena which are easier or
hypothesis with p <0.01. Once we had checked harder for the majority of the evaluated systems.
for the non-randomness of our results, we com-
6 Discussion
puted the Nemenyi test (Nemenyi, 1963) to find
out which phenomena were significantly different. In Section 3.2 we described our evaluation
In our case, we compute the two-tailed Nemenyi methodology and posed four research questions.
test for k = 27 phenomena and N = 11 systems. The experiments that we performed and the anal-
The Critical Difference (CD) for these values is ysis of the results answered all four of them. We
12.5 at p <0.05. briefly discuss the implications of the findings.
Figure 1 shows the Nemenyi test with the CD By addressing RQ 1, we showed that the perfor-
574
mance of a system can differ significantly based switching”) do not appear with a sufficient fre-
on the phenomena involved in each candidate- quency. We release the code for the creation and
paraphrase pair. By addressing RQ 4, we showed analysis of the “performance profile” 4 .
that some phenomena are consistently easier or
harder across the majority of the systems. These 7 Conclusions and Future Work
findings empirically prove the complexity of para- We present a new methodology for evaluation,
phrasing and the task of PI. The results justify the interpretation, and comparison of different Para-
distinction between the qualitatively different lin- phrase Identification systems. The methodology
guistic phenomena involved in paraphrasing and only requires at evaluation time a corpus annotated
demonstrate that framing PI as a binary classifica- with detailed semantic relations. The training cor-
tion problem is an oversimplification. pus does not need any additional annotation. The
By addressing RQ 2, we showed that each sys- evaluation also does not require any additional ef-
tem has strong and weak sides, which can be iden- fort from the systems’ developers. Our methodol-
tified and interpreted via its “performance pro- ogy has clear advantages over using simple quan-
file”. This information can be very valuable when titative measures (Accuracy and F1 Score): 1) It
analyzing the errors made by the system or when allows for a better interpretation and error analysis
reusing it on another task. Given the Deep archi- on the individual systems; 2) It allows for a better
tecture of the systems, such a detailed interpreta- qualitative comparison between the different sys-
tion is hard to obtain via other means and metrics. tems; and 3) It identifies phenomena which are
By addressing RQ 3, we showed that two sys- easy/hard to solve for multiple systems and may
tems can differ significantly in their performance require further research.
on candidate-paraphrase pairs involving particu- We demonstrate the methodology by evaluating
lar phenomenon. These differences can be seen and comparing several of the state-of-the-art sys-
even in systems that have almost identical quan- tems in PI. The results show that there is a statisti-
titative (Acc and F1) performance on the full test cally significant relationship between the phenom-
set. These findings justify the need for a qualita- ena involved in each candidate-paraphrase pair
tive evaluation framework for PI. The traditional and the performance of the different systems. We
binary evaluation metrics do not account for the show the strong and weak sides of each system
difference in phenomena performance. They do using human-interpretable categories and we also
not provide enough information for the analysis or identify phenomena which are statistically easier
for the comparison of different PI systems. Our or harder across all systems.
proposed framework shows promising results. As a future work, we intend to study phenomena
Our findings demonstrate the limitations of the that are hard for the majority of the systems and
traditional PI task definition and datasets and the proposing ways to improve the performance on
way PI systems are typically interpreted and eval- those phenomena. We also plan to apply the evalu-
uated. We show the advantages of a qualitative ation methodology to more tasks and systems that
evaluation framework and emphasize the need to require a detailed semantic evaluation, and further
further research and improve the PI task. The test it with transfer learning experiments.
“performance profile” also enables the direct em-
Acknowledgements
pirical comparison of related phenomena such
as “same polarity substitution (habitual)” and We would like to thank Darina Gold, Tobias Hors-
“(contextual)” or “contains negation” and “nega- mann, Michael Wojatzki, and Torsten Zesch for
tion switching”. These comparisons, however, fall their support and suggestions, and the anonymous
outside of the scope of this paper. reviewers for their feedback and comments.
Our evaluation framework is not specific to the This work has been funded by Spanish Min-
ETPC corpus or the typology behind it. The istery of Science, Innovation, and Universities
framework can be applied to other corpora and Project PGC2018-096212-B-C33, by the CLiC re-
tasks, provided they have a similar format. While search group (2017 SGR 341), and by the APIF
ETPC is the largest corpus annotated with para- grant of the first author.
phrase types to date, it has its limitations as
some interesting paraphrase types (ex.: “negation 4
https://github.com/JavierBJ/paraphrase eval
575
References Hua He and Jimmy Lin. 2016. Pairwise word interac-
tion modeling with deep neural networks for seman-
Hanan Aldarmaki and Mona Diab. 2018. Evaluation tic similarity measurement. In Proceedings of the
of unsupervised compositional representations. In 2016 Conference of the North American Chapter of
Proceedings of COLING 2018. the Association for Computational Linguistics: Hu-
man Language Technologies (NAACL-HLT).
Darina Benikova and Torsten Zesch. 2017. Same same,
but different: Compositionality of paraphrase gran- Shankar Iyer, Nikhil Dandekar, and Kornl Csernai.
ularity levels. In Proceedings of RANLP 2017. 2017. First quora dataset release: Question pairs.
Rahul Bhagat and Eduard H. Hovy. 2013. What is a Yangfeng Ji and Jacob Eisenstein. 2013. Discrimina-
paraphrase? Computational Linguistics, 39(3):463– tive improvements to distributional sentence simi-
472. larity. In Proceedings of the 2013 Conference on
Empirical Methods in Natural Language Process-
Elena Cabrio and Bernardo Magnini. 2014. Decom- ing, EMNLP 2013, 18-21 October 2013, Grand Hy-
posing semantic inferences. att Seattle, Seattle, Washington, USA, A meeting of
SIGDAT, a Special Interest Group of the ACL, pages
Alexis Conneau, Douwe Kiela, Holger Schwenk, 891–896.
Loı̈c Barrault, and Antoine Bordes. 2017. Su-
pervised learning of universal sentence representa- Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov,
tions from natural language inference data. CoRR, Richard Zemel, Raquel Urtasun, Antonio Torralba,
abs/1705.02364. and Sanja Fidler. 2015. Skip-thought vectors. In
C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama,
Janez Demšar. 2006. Statistical comparisons of classi- and R. Garnett, editors, Advances in Neural Infor-
fiers over multiple data sets. J. Mach. Learn. Res., mation Processing Systems 28, pages 3294–3302.
7:1–30. Curran Associates, Inc.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Venelin Kovatchev, M. Antònia Martı́, and Maria
Kristina Toutanova. 2018. Bert: Pre-training of deep Salamó. 2018a. Etpc - a paraphrase identification
bidirectional transformers for language understand- corpus annotated with extended paraphrase typology
ing. arXiv preprint arXiv:1810.04805. and negation. In Proceedings of the Eleventh In-
ternational Conference on Language Resources and
Bill Dolan, Chris Quirk, and Chris Brockett. 2004. Un- Evaluation (LREC 2018).
supervised construction of large paraphrase corpora:
Exploiting massively parallel news sources. In Pro- Venelin Kovatchev, M. Antònia Martı́, and Maria
ceedings of Coling 2004, pages 350–356, Geneva, Salamó. 2018b. WARP-text: a web-based tool for
Switzerland. COLING. annotating relationships between pairs of texts. In
Proceedings of the 27th International Conference on
Samuel Fernando and Mark Stevenson. 2008. A se- Computational Linguistics: System Demonstrations,
mantic similarity approach to paraphrase detection. pages 132–136, Santa Fe, New Mexico. Association
Computational Linguistics UK (CLUK 2008) 11th for Computational Linguistics.
Annual Research Colloqium.
Wuwei Lan, Siyu Qiu, Hua He, and Wei Xu. 2017.
M. Friedman. 1940. A comparison of alternative tests A continuously growing dataset of sentential para-
of significance for the problem of m rankings. The phrases. In Proceedings of the 2017 Conference on
Annals of Mathematical Statistics, 11(1):86–92. Empirical Methods in Natural Language Process-
ing, EMNLP 2017, Copenhagen, Denmark, Septem-
Juri Ganitkevitch, Benjamin Van Durme, and Chris ber 9-11, 2017, pages 1224–1234.
Callison-Burch. 2013. Ppdb: The paraphrase
database. In HLT-NAACL, pages 758–764. The As- Wuwei Lan and Wei Xu. 2018a. Character-based
sociation for Computational Linguistics. neural networks for sentence pair modeling. In
Proceedings of the 2018 Conference of the North
Darina Gold, Venelin Kovatchev, and Torsten Zesch. American Chapter of the Association for Computa-
2019. Annotating and analyzing the interactions be- tional Linguistics: Human Language Technologies
tween meaning relations. In Proceedings of the 13th (NAACL-HLT).
Linguistic Annotation Workshop, pages 26–36, Flo-
rence, Italy. Association for Computational Linguis- Wuwei Lan and Wei Xu. 2018b. Neural network mod-
tics. els for paraphrase identification, semantic textual
similarity, natural language inference, and question
Hua He, Kevin Gimpel, and Jimmy Lin. 2015. Multi- answering. In Proceedings of COLING 2018.
perspective sentence similarity modeling with con-
volutional neural networks. In Proceedings of the Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg.
2015 Conference on Empirical Methods in Natural 2016. Assessing the ability of lstms to learn syntax-
Language Processing, pages 1576–1586. Associa- sensitive dependencies. Transactions of the Associ-
tion for Computational Linguistics. ation for Computational Linguistics, 4:521–535.
576
Nitin Madnani, Joel Tetreault, and Martin Chodorow.
2012. Re-examining machine translation metrics
for paraphrase identification. In Proceedings of the
2012 Conference of the North American Chapter
of the Association for Computational Linguistics:
Human Language Technologies, NAACL HLT ’12,
pages 182–190, Stroudsburg, PA, USA. Association
for Computational Linguistics.
H. B. Mann and D. R. Whitney. 1947. On a test of
whether one of two random variables is stochas-
tically larger than the other. Ann. Math. Statist.,
18(1):50–60.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Cor-
rado, and Jeffrey Dean. 2013. Distributed represen-
tations of words and phrases and their composition-
ality. In Proceedings of the 26th International Con-
ference on Neural Information Processing Systems -
Volume 2, NIPS’13, pages 3111–3119, USA. Curran
Associates Inc.
577
Study on Unsupervised Statistical Machine Translation for
Backtranslation
578
Proceedings of Recent Advances in Natural Language Processing, pages 578–582,
Varna, Bulgaria, Sep 2–4 2019.
Dataset Type Domain No. of Sentences
News-Commentary-v14 (Ru-En) Parallel News 290866
Newscrawl 2018 (Ru) Monolingual News 8669559
Newscrawl 2017 (Ru) Monolingual News 8233907
Newscrawl 2018 (En) Monolingual News 18113311
describe the datasets that we used for our experi- tem using only monolingual corpus (Lample et al.,
ments. Next, we describe our experimental setup 2018; Artetxe et al., 2018).
for carrying out the experiments. We then analyze In our study, we use an unsupervised statisti-
the results for NMT, USMT and other models on cal machine translation system based on Artetxe
Russian - English dataset. We conclude the paper et al. (2018) 1 . In their paper, they describe a novel
with discussion on future work. method to build a statistical machine translation
model using monolingual data without any paral-
2 Related Works lel data. The main idea is to learn word embed-
In this section, we will describe the main ideas dings for each language independently and use lin-
and experiments using backtranslation and mono- ear transformations to bring them to shared space.
lingual data. These embeddings are used to generate the phrase
Backtranslation was popularized by Sennrich table and then, the SMT is fine-tuned on a syn-
et al. (2016a) where they improved the state-of- thetic training set. In our work, we make use of
the-art in several language pairs. They trained a this model to generate more synthetic data to be
target to source machine translation model on the added to the parallel corpora.
aligned parallel corpus. The model was later used
to translate target-side monolingual sentences to 3 Data
the source language. These new synthetic sen-
We perform our experiments on the Russian - En-
tences were added to the training set and a new
glish language pair. For training the supervised
model is trained.
model (both the back-translated model and re-
On similar lines, Zhang and Zong (2016) use trained model), we use the Russian - English par-
the source-side monolingual data to train the NMT allel corpus from WMT 19 2 . To train our unsu-
model. They build a baseline NMT model and pervised model, we used monolingual corpora for
then use that model to generate the synthetic paral- Russian 3 and English 4 from Newscrawl 2017 and
lel data. They experiment on self-training as well 2018 datasets. For English, we consider only the
as multitask learning. Newscrawl 2018 as the monolingual data where as
He et al. (2016) introduced a dual learning for Russian we combine both Newscrawl 2017 and
mechanism for neural machine translation. They 2018 for the monolingual data.
train translation models in both directions and use
We provide more details regarding the data in
monolingual data to provide feedback on the qual-
Table 1.
ity of the translation. Their main contribution
was to treat machine translation as a reinforcement
3.1 Preprocessing
learning problem.
Hoang et al. (2018) proposed a simple tech- For normalizing punctuation in the parallel cor-
nique. They do show that iterative backtranslation pora, we use the default scripts provided by Moses
5 . We then perform true-casing followed byte-
improves the performance of the system on large
datasets. However, they observe that the improve- 1
We use this method as it did not require huge amounts of
ments in low resource datasets were not signifi- resource to train compared to other methods such as Artetxe
cant. et al. (2019)
2
As mentioned earlier, obtaining aligned paral- http://www.statmt.org/wmt19/index.html
3
http://data.statmt.org/news-crawl/ru/
lel corpora is expensive and cumbersome. Recent 4
http://data.statmt.org/news-crawl/en/
efforts in Unsupervised Machine Translation indi- 5
https://github.com/moses-
cate that it is possible to develop a competitive sys- smt/mosesdecoder/tree/master/scripts
579
pair encoding while training the NMT 6 (Sennrich Neural Machine Translation model. For example,
et al., 2016b). we use back-translated Russian monolingual cor-
pus to generate synthetic English sentences. These
3.2 Postprocessing English sentences are added to the training corpus
We remove byte pair encodings, detrucase and (with varying training corpora sizes) with English
detokenize all the translated sentences before val- as the source and Russian as the target. We then
idating the predictions against the test set. train an NMT model from scratch and report our
performance on an unseen test set.
4 Experimental Setup In the case USMT, we backtranslate the mono-
We describe our experimental setup in this section. lingual corpus using the USMT model and aug-
For our unsupervised statistical machine trans- ment the sampled training data. The training data
lation model, we use Monoses (Artetxe et al., is used to build the NMT model.
2018). We use the defaults for training the unsu- EN-RU
pervised model from the Monoses code. We train 20
the model on both Russian to English (ru-en) and BLEU Score NMT
English to Russian (en-ru) monolingual corpus. USMT
Monoses frameworks trains the model in 8
steps. Steps 1-7 involves training the word embed- 15
dings and bringing them to a shared space to build
an initial phrase table. In step 8, 2M sentences
are generated through backtranslation in both the
directions. The phrase table is fine-tuned for 3 it- 10
erations using the synthetic sentences to obtain the
final model.
We use OpenNMT for training the supervised
machine translation models (Klein et al., 2017).
We use the default architecture in OpenNMT, 25% 50% 75%
which includes an LSTM layer for encoding and
Sample Size
another LSTM layer for decoding. Furthermore,
we do not use the pretrained word embeddings. Figure 1: BLEU scores for NMT backtranslation and
For both NMT and retraining the NMT model, we USMT backtranslation for EN-RU
use the same architecture with default values.
To train the USMT model and NMT models, we
used a system with 4x vCPUS, NVIDIA Tesla K80 6 Results
GPU and 61 GB of RAM. The USMT model took
We perform all the experiments mentioned in the
about 2 weeks to complete training where as the
Table 2. Each row has a key which corresponds to
NMT model usually took about 4-5 hours.
the type of training corpora. The training corpora
5 Experiments includes Only Parallel corpora, Only Monolingual
Corpora and a sampled combination of the Parallel
Our primary aim is to show that we can use an un- Corpora and the backtranslated Monolingual Cor-
supervised machine translation model to improve pora.
the performance of NMT systems. At the same We obtain the baseline NMT model using the
time, we want to have a fair validation and investi- parallel corpus without any additional synthetic
gate numerous scenarios where our approach per- sentences. As mentioned earlier, we train the par-
forms well. allel corpora using the default encoder-decoder ar-
Therefore, in this experiment, we use mono- chitecture from OpenNMT. Furthermore, we train
lingual data to generate synthetic sentences using a baseline USMT model using the monolingual
both NMT and USMT separately and then, each of corpora of English and Russian. We report the re-
them is used to augment the training data for the sults in Table 2.
6
https://github.com/rsennrich/subword- From the baselines of NMT and USMT, it is
nmt/tree/master/subword nmt very clear that NMT outperforms USMT when
580
ru-en en-ru
Training Corpus NMT USMT NMT USMT
Only Parallel Corpora 17.65 - 15.54 -
Only Monolingual Corpora - 15.58 - 8.07
10% (∼29K) + Backtranslated Corpora 13.17 16.34 9.31 11.57
25% (∼72.5K) + Backtranslated Corpora 14.81 17.17 12.02 12.63
50% (∼145K) + Backtranslated Corpora 19.63 18.46 14.67 13.18
75% (∼217K)+ Backtranslated Corpora 20.17 19.34 15.74 13.97
Table 2: The BLEU scores for the different experiments. The training corpus with x% indicates the percentage of
aligned training pairs randomly sampled from the parallel corpora.
581
8 Code Guillaume Klein, Yoon Kim, Yuntian Deng, Jean
Senellart, and Alexander Rush. 2017. OpenNMT:
To facilitate reconstruction of our pa- Open-source toolkit for neural machine translation.
per, we are releasing the code - In Proceedings of ACL 2017, System Demonstra-
https://github.com/anush6/USMT For Backtranslation tions, pages 67–72, Vancouver, Canada. Association
for Computational Linguistics.
9 Acknowledgement Philipp Koehn and Rebecca Knowles. 2017. Six chal-
lenges for neural machine translation. In Pro-
We would like to thank our anonymous reviewers ceedings of the First Workshop on Neural Machine
for their helpful feedback and comments. Translation, pages 28–39, Vancouver. Association
for Computational Linguistics.
582
Towards Functionally Similar Corpus Resources for Translation
583
Proceedings of Recent Advances in Natural Language Processing, pages 583–592,
Varna, Bulgaria, Sep 2–4 2019.
made by the author. This notion reflects the dif- ing genre annotations that exist in these corpora
ferences in the linguistic make-up of texts and it (see Table 4). We want to reduce the probability
relies on frequencies of lexicogrammatic features that the differences observed in the translations are
such as passive voice, relative clauses or personal down to the differences between the sources and
pronouns. It is assumed that the observed linguis- are not genuine translational or cultural effects.
tic variation captures the possible combinations of The rest of the paper is structured as follows.
field, tenor and mode, the most prominent factors Section 2 has a brief overview of the topologi-
of communication (suggested by Halliday (1985)). cal approach to the text characterization and the
On the other hand, genres are understood as research related to corpus similarity and genre
conventionally recognizable text categories that classification. In Section 3 we describe our ap-
can be established on a number of external crite- proaches to FTDs modelling and report the results
ria, referring to the function of the text and its situ- of the intrinsic evaluation of the models. A selec-
ational constrains. According to Lee (2001), most tion of BNC genres is used to evaluate the models
existing corpora rely on the text-external approach against an independent judgment and to test the
to text categorization and the choice of parameters clustering approaches to be used in the real-life
behind it is guided by practical considerations in task (Section 4). Section 5 presents a study that
each case. It has been shown how little compa- showcases the application of the functional vec-
rability there is between the genre classifications tors to computing the most similar parts of the two
used to annotate different corpora (Sharoff, 2018). corpora. In Section 6 we aggregate the analytic
TS researchers interested in register variation find results and highlight important findings.
that existing corpora provide “limited background
information on the genres ... and how they were 2 Related Research
defined” and choose to set up annotation tasks to
The practical needs to describe and compare cor-
reorganize the corpora (Delaere, 2015).
pora have made ‘corpusometry’ a prominent area
One translationally relevant common footing to of research in corpus linguistics. Below we outline
compare texts from corpora with divergent or ab- the two major approaches to measuring similar-
sent genre annotation is to rely on their function. ity and describing the corpora contents. The first
On the one hand, text function is an important fac- one is based on lexical features and yields a the-
tor in translation, as texts aimed at informing the matic description of corpus texts. It is one of the
reader are translated differently from texts aimed most prominent methods of measuring similarity
at promoting goods and services (Lapshinova- between texts and/or building comparable corpora.
Koltunski and Vela, 2015). On the other hand, text For example, Kilgarriff and Salkie (1996) put
functions can be used to produce continuous rather forward a corpus homogeneity/similarity measure
than discrete text descriptions and account for hy- based on calculating χ2 statistics from frequency
brid texts. In this research we explore the poten- lists or N keywords. A lexical approach to esti-
tial of Functional Text Dimensions (FTD), hand- mate the corpus composition is taken by Sharoff
annotated for English and Russian (Sharoff, 2018) (2013). This research compared the results of
to produce text representations and to build func- clustering and topic modelling as ways to rep-
tionally comparable corpora for TS research. resent a corpus content using keywords statis-
The aim of the present study is solve a practi- tics. In Li et al. (2018), the authors com-
cal task of creating research corpora for the study pared the performance of several bilingual vo-
of translational tendencies in English-German and cabulary overlap measures on a specifically de-
English-Russian translation. To this end, we de- signed corpus with known comparability levels
velop a method to build a reasonably big and func- and found that frequencies of words with a sim-
tionally homogeneous intersection of the three text ple Presence/Absence weighting scheme outper-
collections: CroCo, an English-German parallel formed other approaches.
corpus (Hansen-Schirra et al., 2006), and the stu- Another approach to measuring corpora has to
dents and professional collections from RusLTC, a do with calculating frequencies of a range of lex-
English-Russian parallel corpus (Kutuzov and Ku- icogrammatic features (tense forms, modals) that
nilovskaya, 2014). Our major motivation for this allegedly reflect linguistically relevant parameters
research is to find a way to reconcile the diverg- of the communicative situations. This text-internal
584
approach to the text categorization can be best ex- is particularly relevant to our task for three rea-
emplified by Biber’s work (Biber, 1988). He used sons: (1) it provides a solid theoretically grounded
several dozens of hand-picked text-internal fea- approach for comparing texts coming from differ-
tures to place a text along each of the six dimen- ent or unknown sources and for producing com-
sions (e.g. involved vs informational production parative descriptions for the corpora at hand, (2)
or abstract vs non-abstract information). Biber’s it is focused on functional and communicative pa-
multidimensional approach to describing text vari- rameters of texts that are particularly important in
ation has been criticized for lack of interpretabil- TS, (3) this framework, like Biber’s, provides a
ity and, more importantly, for being loosely re- flexible way to represent texts functionality along
lated to any external and situational factors, which a few dimensions instead of squeezing texts into
cab be a socially more important reality for text the atomic genre labels. In effect, FTD framework
categorization than linguistic features. The lat- is a way to produce functional text vectors that po-
ter can throw together texts that are perceived as sition each individual text in a multidimensional
belonging to different genres (Lee, 2001). The functional space and help to account for variation
attempts to classify genres, particularly, as anno- within and across text categories.
tated in the BNC and limited to a selection of ma-
jor ‘tried and tested’ three or four top-level cat- 3 Modelling: Setup and Results
egories, have shown that the 67 Biber’s features
can be an overkill for a task like that. Lijffijt and The annotated data from the FTD project was used
Nevalainen (2017) report over 90% classification to learn models that predicted 10-dimensional vec-
accuracy for the BNC four major genres on just tors for the English texts in our research corpora.
pairs of surface features (such as frequencies of Further on, we used these vectors to compare texts
nouns and pronouns, values of type-to-token ra- and to get functionally similar subcorpora for a
tio and sentence length). The results from Kilgar- subsequent TS research (Section 5).
riff and Salkie (1996); Xiao and McEnery (2005) The annotated data for English consists of 1624
indicate that the most frequent words can cope chunks of texts that count about 2 mln tokens from
with the four major BNC categories as well. More two different sources: 5gthe Pentaglossal cor-
specifically, Xiao and McEnery (2005) show that pus (Forsyth and Sharoff, 2014) and ukWac (Ba-
keyword analysis can be used to replicate Biber’s roni et al., 2009). We used the annotations for
results. In effect they analyze differences in the the 10 most prominent FTD described in Sharoff
frequencies of mostly functional words that are (2018). Each dimension received a score on a 4-
key to genre identification. In a setting simi- point Likert scale that reflects the proximity of a
lar to Biber’s, Diwersy et al. (2014) use 29 lexi- text to the suggested functional prototype. The
cogrammatic features and mildly-supervised ma- inter-annotator agreement is reported at Krippen-
chine learning methods to tease apart genres an- dorff’s α >0.76. We refer the reader to Sharoff
notated in CroCo. The visualizations they provide (2018) for more details on the FTD framework.
indicate that they have managed to clearly separate We used two modelling approaches to learn
only fiction and instruction of the eight genres in functional vectors from the annotated dataset: a
their experiment. multi-label task in a deep neural network architec-
ture and a set of binary classifiers in a traditional
This demonstrates that describing genres needs machine learning setting. The respective models
a sophisticated approach that takes into account produced two types of functional vectors, which
a multitude of criteria such as topic and situ- demonstrated comparable performance in several
ated linguistic properties. This research contin- evaluation settings. This paper investigates the
ues the investigation of the functional aspect of differences between, and adequacy of, these two
genre shaped in Sharoff’s Functional Text Dimen- types of functional vectors.
sions (Sharoff, 2018). Sharoff’s work establishes In the neural networks scenario, we used a bidi-
a text-external framework to capture human per- rectional LSTM with an attention layer and two
ception of the texts functionality (as distance to types of text input. Firstly, texts received mixed
a functional prototype) and to link it to any text- token-PoS representations suggested by Baroni
internal representations, with the aim of predicting and Bernardini (2006), biLSTMmix throughout
the functional setup of unknown texts. This work this paper and in Table 1). The 1500 most fre-
585
quent words were kept in their surface form, while Figure 1. Distribution of predictions for the mod-
the rest of the tokens were converted into their eling approaches
PoS. For example, a sentence “It was published
in 1931 by one of New York’s major publishers.”
was transformed into “It was VERB in [#] by one
of PROPN PROPN major NOUN.” The embed-
dings for PoS were initialized as random vectors
and trained in the Embedding layer. Secondly,
we used lemmatised texts, with stop words fil-
tered out (biLSTMlex in Table 1). For both sce-
narios we used pre-trained word embeddings of
size 300, trained on the English Wikipedia and
CommonCrawl data, using the skip-gram model,
from the WebVectores database (Kutuzov et al.,
2017). The preliminary experiments showed that
cross entropy as the loss function with the Adam
optimizer performed best (Kingma and Ba, 2014).
We trained the models for 10 epochs. In the ML
case, we reformulated the task as the binary clas-
sification task and learnt a classifier for each FTD. which penalizes model errors equally regardless of
To this end, we binarized the existing human anno- class distributions.
tations by converting ‘0.5’ score to 0 and ‘2’ to 1. From the statistics in Table 1, it follows that
To get the real-valued functional vectors we used the deep learning approach is more accurate in de-
the probabilities returned for the positive class for termining the text functionality than the classical
each FTD on the assumption that the model would algorithms, and the mixed representations work
return higher probabilities for texts with a clearer best.
functional makeup. We experimented with fea- The difference between the models perfor-
tures (TF-IDF and Biber’s 67 text-internal regis- mance, however, is quite slim: it is in the second
ter features) and with different algorithms (Sup- decimal digit only. A brief glance at the values
port vector machines (SVM), RandomForest (RF), of the functional vectors components (i.e. values
Logistic Regression (LogReg)). SVM and RF re- predicted for each FTD) returned by the models
sults below pertain to the experiments with the reveals the differences in how the models arrive at
grid search optimized for macro F1 score. TF-IDF the same overall result. Figure 1 shows the prob-
representation proved to be inferior to the Biber’s ability density for the values produced by the best
features and was excluded from the results below. performing models in each learning setup.
We added a dummy classifier which randomly pre- Figure 1 demonstrates that biLSTM, unlike the
dicts a class with respect to the class distribution traditional ML algorithms, tends to predict near-
as a baseline. For register feature extraction (the zero values, with up to 7-11% of the training texts
Biber’s features) we used MAT for English (Nini, receiving values smaller than 0.2 on the strongest
2015). ‘dominant’ dimension.
In the next section we will show how the pre-
To use a comparable performance metrics for
dictions of these two models correlate with the
the two learning approaches, the annotations and
experiment-independent judgment we can sug-
the models predictions were transformed into
gest.
multi-hot vectors.
In Table 1 we report the standard measures av- 4 Evaluation on BNC Categories
eraged over 10 FTDs on the 10-fold cross valida-
tion for the six experiments. We accounted for 4.1 Genre Classification
the severe class imbalances in all training settings To test the functional vectors on the data out-
by using ‘class weight=balanced’ option, strati- side the annotated dataset, we constructed a cor-
fied (multi-label) split with cross-validation and, at pus with the ‘known’ genre composition. To this
the evaluation stage, by choosing macro-averaging end, we followed Lee’s scheme for the BNC text
586
FTD perspective FTD minority class Samples perspective
P R F1 F1 F1 neg humming loss
biLSTMmix .804 .767 .776 .609 .640 .902
biLSTMlex .787 .747 .757 .576 .596 .895
RF .732 .756 .723 .532 .517 .846
SVM .667 .531 .510 .095 .059 .844
LogReg .664 .734 .659 .480 .527 .753
dummy .504 .504 .504 .169 .134 .737
587
Figure 2. Average values on the 10 FTDs by BNC 4 clusters on our BNC selection, was returned
genres predicted by biLSTMmix model for both biLSTMmix and RandomForest vectors
at damping=0.9 and preference=-12. The clus-
ters, quite predictably, were built around (1) fic-
tion, (2) reportage (3) non-academic + editorial (4)
academic + instructions. Vectors learnt on lemma-
tized embeddings (biLSTMlex) were not able to
converge to this solution.
An alternative clustering technique used in this
research was KMeans algorithm with the Elbow
Criterion method to determine k. The latter is
based on measuring the sum of the squared dis-
tances between each member of the cluster and
its centroid. k is defined as the number of clus-
ters after which this value drops abruptly. How-
ever, this method needs to be applied with regard
to the task at hand and some understanding of the
data. For BNC k was automatically put at 2, be-
cause of the imbalance in our collection towards
fiction: more than half of the texts were fiction.
The best KMeans result (ARI=0.92) was regis-
to the true genre category following the strongest tered on the RandomForest model vectors for k=5.
prediction, if we map genres to FTDs as follows: It was superior to the best biLSTM result in that
fiction : fictive, academic : scitech, reportage : it separated the instruction cluster. Clustering on
news, instruction : instruction, non-academic : ar- Biber’s features and keywords did not achieve ARI
gumentative. However, reducing a functional vec- of more than 0.2 for either Affinity Propagation or
tor to just the strongest component would be unfair KMeans.
to the functionally hybrid texts that fall under the
genre labels of non-academic and editorial in our 5 Case Study: CroCo and RusLTC
BNC slice.
In this section we report the results of a case study
where we used the functional vectors to get com-
4.2 Testing the Clustering Method on BNC
parable functional clusters from several text col-
The ultimate goal of this work is to produce a func- lection. Our data comes from the English parts
tional intersection of two corpora, i.e. to find func- of the three parallel corpora: RusLTC, including
tionally comparable texts in several sets. In this student and professional translations subcorpora
section we apply two clustering techniques to the that have different English sources, and the CroCo
BNC selection to determine which text representa- corpus. As can be seen from Table 4, the three
tion and clustering approach is better at matching text collections vary in size, have diverging genre
the annotated genres as class labels. setup, and there is no way to tell whether the same
In the first clustering scenario we ran Affinity categories include the same texts. In this work
Propagation on a square matrix of pair-wise corre- CroCo was chosen as the normative corpus, i.e.
lations pre-computed as euclidean similarities for the starting point for the comparison and cluster-
the 650 BNC texts. This approach is attractive be- ing operations.
cause it does not require the value of k, which is The first step in solving our practical task with
difficult to deduce in the real application context. KMeans was to determine k. K-value was iden-
We searched through the combinations of parame- tified as n+1, where n is the number of the most
ters to get the highest score for the Adjusted Rand populous groups formed by the texts with a spe-
Index (ARI), a clustering metric, which returns the cific FTD as the dominant function (see Figure 3,
proportion of text pairs that were assigned to the which show the ratio of texts with a specific dom-
correct cluster, given the gold standard classes. inant function). For the tree corpora in our exper-
The best clustering solution, with ARI=0.92 and iment it seems possible to set k to 5 or 6. Our ex-
588
CroCo RusLTC(stu) RusLTC(pro) Figure 3. Ratio of texts by the dominant FTD in
EN>GE EN>RU EN>RU the research corpora as predicted by biLSTM
Tokens 240 K 213 K 1.2 M
Texts 110 360 517
Acad(12) Media(417)
Essay(29) Adverts(12) Popsci (100)
Fiction(10) Educat(58)
Instr(10) Essay(131)
Business(13) Fiction(12)
Genres
Popsci(11) Info(143)
Speech(14) Interview(3)
Tourist(11) Letters(3)
Web(12) Speech(12)
Tech(15)
589
can be explained by two factors: first, these func- the BNC genres (5 clusters, ARI=0.92) only for
tions can rely on lexical features for their expres- the RandomForest vectors. The functional vec-
sion and, second, they are often annotated as a sec- tors learnt on embedded mixed representations
ond strong dimension, unlike the mutually exclu- achieved ARI=0.58 for any k in the range from
sive (genre-pivotal and relatively easy to predict) 4 to 8. Note, however, that any functional vectors
FTD such as fiction, instruction, news and scitech. were by far better in this task than the baselines:
To support this argument: in the classification task we failed to produce any good clustering results
on the six hand-picked BNC genre categories, the for our BNC selection on the lexical and on the
raw Biber’s features performed a bit better than raw register features.
the functional vectors learnt on them, while the
biLSTMmix vectors demonstrated even less skill 7 Conclusions
in recognizing our select BNC genres, where the
majority of texts are of the easy-to-recognize type. This paper presents an approach to deal with a
On RusLTC(pro) corpus which consists of mass practical issue of constructing functionally compa-
media texts and popular scientific texts, biLSTM- rable corpus resources. We proposed a method to
mix returns no reliable predictions for the stagger- measure functional comparability of the resources
ing 19% of texts. This analysis shows that FTD at hand and to produce their functionally homoge-
detection can benefit from combining vectorized neous intersection. The method offers a way to
and statistical register features, which we leave for verify the researcher’s judgments about the cor-
future work. pora comparability which are usually based on
pre-existing corpus annotation schemes and re-
Second, though our modelling approaches per searcher’s intuition. We show that texts can be
se are not directly comparable, because they had described externally via a reference to a number
different objectives and operated on different text of communicative functions and that the functions
representations, we can evaluate their usefulness are reflected via text-internal linguistic features.
for the practical task of predicting functional prop- We found that functional text representations offer
erties of texts. The inspection of the real-valued a better clustering result for a corpus with ‘known’
vectors indicates that the vectors learnt within the functions in comparison to keywords and linguis-
classification task setting overestimated the texts tic register features. They can be effectively used
functionality (i.e. produced noisy predictions) and to identify functionally homogeneous subsets of
were less adequate in determining the functions hi- texts in a given text collection and to match them
erarchy as manifested in the human scores. The to functionally comparable sets from another cor-
two approaches had very similar overall perfor- pus. The cross-linguistic extension of this research
mance in the intrinsic evaluation and in the BNC (left for future work) is supposed to equip a re-
genre classification task, though in the real ap- searcher with a corpus of non-translations in the
plication they produce only partially overlapping TL functionally comparable to the ST. Such a ref-
clusters. This is probably because the models erence corpus would effectively represent the ex-
are focusing different properties of texts that are pected TL textual fit (Chesterman, 2004) that is
equally relevant for fulfilling text functions, but necessary to estimate specificity of translations.
are more or less pronounced in individual real
texts. It seems reasonable to use the union of the Acknowledgments
two sets for practical purposes. Besides, the mod-
els are different in terms of processing effort re- The first author has been supported by the Russian
quired, with the model on Biber’s features less eas- Foundation for Basic Research within Project No.
ily applicable to big corpora. 17-06-00107.
Third, the effectiveness of the functional repre-
sentation was ultimately tested in the BNC clus- References
tering task. While for Affinity Propagation on
pair-wise similarities the type of functional vec- Marco Baroni and Silvia Bernardini. 2006. A new
approach to the study of translationese: Machine-
tors did not matter, the better-performing KMeans learning the difference between original and trans-
proved to be sensitive to the difference in the func- lated text. Literary and Linguistic Computing
tional vectors and managed to find a good fit to 21(3):259–274. https://doi.org/10.1093/llc/fqi039.
590
Marco Baroni, Silvia Bernardini, Adriano Ferraresi, Ekaterina Lapshinova-Koltunski and Mihaela Vela.
and Eros Zanchetta. 2009. The WaCky wide web: 2015. Measuring ’Registerness’ in Human and Ma-
a collection of very large linguistically processed chine Translation: A Text Classification Approach.
web-crawled corpora. Language Resources and Proceedings of the Second Workshop on Discourse
Evaluation 43(3):209–226. in Machine Translation (September):122–131.
Douglas Biber. 1988. Variations Across Speech and David Lee. 2001. Genres, registers, text types, do-
Writing. Cambridge University Press. mains, and styles: clarifying the concepts and nav-
igating a path through the BNC jungle. Language
Andrew Chesterman. 2004. Hypotheses about Learning and Technology 5(3):37–72.
translation universals. Claims, Changes and
Challenges in Translation Studies pages 1–14. Bo Li, Eric Gaussier, and Dan Yang. 2018. Mea-
https://doi.org/10.1075/btl.50.02che. suring bilingual corpus comparability. Nat-
ural Language Engineering 24(4):523–549.
Isabelle Delaere. 2015. Do translations walk the line? https://doi.org/10.1017/S1351324917000481.
Visually exploring translated and non-translated
texts in search of norm conformity. Phd, Ghent Uni- Jefrey Lijffijt and Terttu Nevalainen. 2017. A simple
versity. model for recognizing core genres in the bnc. In
Big and Rich Data in English Corpus Linguistics:
Sascha Diwersy, Stefan Evert, and Stella Neumann.
Methods and Explorations, University of Helsinki,
2014. A weakly supervised multivariate approach to
VARIENG eSeries, volume 19.
the study of language variation. In Aggregating di-
alectology, typology, and register analysis. linguis- Albrecht Neubert. 1985. Text and Translation. Enzyk-
tic variation in text and speech, Walter de Gruyter, lopdie.
Berlin, pages 174–204.
Stella Neumann. 2013. Contrastive register variation.
Richard Forsyth and Serge Sharoff. 2014. Document A quantitative approach to the comparison of En-
dissimilarity within and across languages: a bench- glish and German. Mouton de Gruyter, Berlin,
marking study. Literary and Linguistic Computing Boston.
29:6–22.
Andrea Nini. 2015. Multidimensional Analysis Tagger
Michael A. K. Halliday. 1985. An Introduction to
(v. 1.3).
Functional Grammar. Edward Arnold, London.
Christiane Nord. 2006. Translating as a purposeful
Silvia Hansen-Schirra, Stella Neumann, and Mihaela
activity: a prospective approach. TEFLIN Journal
Vela. 2006. Multi-dimensional annotation and
17(2):131–143.
alignment in an English-German translation corpus.
In Proc 5th Workshop on NLP and XML: Multi- Randolph Quirk, Sidney Greenbaum, Geoffrey Leech,
Dimensional Markup in Natural Language Process- and Jan Svartvik. 1985. A Comprehensive Grammar
ing at EACL. Trento, pages 35–42. of the English Language. Longman.
Adam Kilgarriff and Raphael Salkie. 1996. Corpus Katharina Reiss and Hans J Vermeer. 1984. Ground-
similarity and homogeneity via word frequency. In work for a general theory of translation. Tubingen:
Proceedings of Euralex. volume 96. Niemeyer 101.
Diederik P Kingma and Jimmy Ba. 2014. Adam: A
Serge Sharoff. 2013. Measuring the distance between
method for stochastic optimization. arXiv preprint
comparable corpora between languages. In Building
arXiv:1412.6980 .
and Using Comparable Corpora, Springer, pages
Haidee Kruger and Bertus Van Rooy. 2010. The fea- 113–130. http://www.tausdata.org/.
tures of non-literary translated language: a pilot
study. Proceedings of Using Corpora in Contrastive Serge Sharoff. 2018. Functional Text Dimensions for
and Translation Studies, England, July 2010 . annotation of Web corpora. Corpora 13(1):65–95.
Andrei Kutuzov, Murhaf Fares, Stephan Oepen, and Alexandr Shveitzer. 1973. Translation and Lin-
Erik Velldal. 2017. Word vectors, reuse, and replica- guistics: Informational newspaper and mili-
bility: Towards a community repository of large-text tary publicist texts in translation [Perevod i
resources. In Proceedings of the 58th Conference lingvistika: O gazetno-informacionom i voienno-
on Simulation and Modelling. Linköping University publicisticheskom perevode]. Voenizdat.
Electronic Press, pages 271–276.
Zhonghua Xiao and Anthony McEnery. 2005.
Andrey Kutuzov and Maria Kunilovskaya. 2014. Rus- Two Approaches to Genre Analysis: Three
sian Learner Translator Corpus: Design, Research Genres in Modern American English.
Potential and Applications. In Text, Speech and Di- Journal of English Linguistics 33:62–82.
alogue: 17th International Conference, TSD 2014, https://doi.org/10.1177/0075424204273957.
Brno, Czech Republic, September 8-12, 2014, Pro-
ceedings. Springer, volume 8655, page 315.
591
Appendix: Supplementary material
CroCo clusters as determined for the two alternative functional representations against the existing an-
notation
biLSTMmix RandomForest
texts description homo genres texts description homo genres
Cluster 0 12 instr:0.91, .685 INSTR 9, 15 instr:0.71, .687 INSTR 10,
promo:0.19, WEB 3 promo:0.57, WEB 5
info:0.09 info:0.43
Cluster 1 43 argum:0.83, .706 ESSAY 27, 47 argum:0.7, .626 ESSAY 22,
new:0.11, SPEECH 12, per- SPEECH 13,
per- SHARE 2, sonal:0.62, SHARE 7,
sonal:0.08 POPSCI 1, new:0.54 POPSCI 3,
WEB 1 FICTION 1,
TOU 1
Cluster 2 12 info:0.59, .588 TOU 7, 39 scitech:0.5, .487 TOU 10,
promo:0.19, WEB 2, new:0.5, POPSCI
eval:0.14 POPSCI 1, info:0.48 8, WEB 7,
SPEECH 1, ESSAY 7,
SHARE 1 SHARE 6,
SPEECH 1
Cluster 3 24 fiction:0.27, .406 FICTION 9 fictio:0.85, .672 FICTION 9
scitech:0.13, 10, POPSCI eval:0.48,
argum:0.1 8, WEB 3, per-
ESSAY 2, sonal:0.35
SPEECH 1
Cluster 4 19 promo:0.7, .527 SHARE 10,
info:0.15, TOU 4, WEB
argum:0.1 3, POPSCI 1,
INSTR 1
592
Question Similarity in Community Question Answering:
A Systematic Exploration of Preprocessing Methods and Models
Florian KunnemanA , Thiago Castro FerreiraB , Emiel KrahmerB , and Antal van den BoschA,C
A
Centre for Language Studies, Radboud University, The Netherlands
B
Tilburg center for Cognition and Communication, Tilburg University, The Netherlands
C
KNAW Meertens Institute, Amsterdam, The Netherlands
593
Proceedings of Recent Advances in Natural Language Processing, pages 593–601,
Varna, Bulgaria, Sep 2–4 2019.
how these models perform in comparison to (or
combined with) more traditional question similar- Y
T RLM (Q1 , Q2 ) = (1 − σ)Ptr (w|Q2 ) + σPlm (w|C)
ity techniques.
w∈Q1
In this paper we therefore systematically com- X
Ptr (w|Q2 ) = α Sim(w, t)Plm (t|Q2 )+
bine and compare the SoftCosine metric and the t∈Q2
Syntactic Tree Kernels with the more traditional (1 − α)Plm (w|Q2 )
BM25 and translation-based models. Moreover, (1)
we analyze the impact of preprocessing steps
(lowercasing, suppression of punctuation and stop Sim(w, t) denotes a similarity score among
words removal) and word-similarity metrics based words w and t. In the original study, this sim-
on different distributions (word translation proba- ilarity metric is the word-translation probability
bility, Word2Vec, fastText and ELMo). P (w|t) obtained by the IBM Translation Model
The experiments were mainly conducted on the 1 (Brown et al., 1993). Furthermore, C denotes a
data of SemEval 2016-2017 - Task 3, based on the background corpus to compute unigram probabil-
Qatar Living corpus. As a secondary goal, we also ities in order to avoid 0 scores.
evaluated our main models in classifying question SoftCosine is the ranking function used by Sim-
duplicates on the Quora dataset, so as to assess BOW (Charlet and Damnati, 2017), the winning
whether the results that we find apply to different system of the question similarity re-ranking task of
datasets. SemEval 2017 (Nakov et al., 2017). The method
Results show that the choice of a preprocess- is similar to a cosine similarity between the tf-idf
ing method and a word-similarity metric have a bag-of-words of the pair of questions, except that
considerable impact on the final results. We also it also takes into account word-level similarities as
show that the combination of all the analyzed ap- a matrix M . Given X and Y as the respective tf-
proaches leads to results competitive with related idf bag-of-words for questions Q1 and Q2 , Equa-
work in question-similarity. tion 2 summarizes the SoftCosine metric.
2 Models
X tM Y
Sof tCos(X, Y ) = √ √
We compare two traditional and two recent ap- X tM X Y tM Y
Xn Xn
proaches in this study: BM25, Translation- X tM X = Xi Mij Yj (2)
Based Language Model (TRLM), SoftCosine and i=1 j=1
Smoothed Partial Tree Kernels (SPTK - Syntactic Mij = max(0, cosine(Vi , Vj ))2
Tree Kernels).
As Sim(w, t) in Equation 1, Mij represents
BM25 is a fast information retrieval technique the similarity between the i-th word of question
(Robertson et al., 2009) used as a search engine Q1 and the j-th one in question Q2 . cosine is
in the first step of the shared task by many studies. the cosine similarity, and Vi and Vj are originally
We used the implementation of BM25 provided by 300-dimension embedding representations of the
gensim5 as a baseline. words, trained on the unannotated part of the Qatar
living corpus using Word2Vec (Mikolov et al.,
Translation-Based Language Model (TRLM) 2013) with a context window size of 10.
is a question similarity ranking function, first in-
Smoothed Partial Tree Kernels (SPTK) are
troduced by Xue et al. (2008). The method com-
the basis of KeLP (Filice et al., 2016), a system in-
bines a language model with a word translation
troduced by Croce et al. (2011). SPTK applies the
system technique, and is known to obtain better
kernel trick by computing the similarity of ques-
results on the question similarity task than BM25
tion pairs based on the number of common sub-
and only the language model (Jeon et al., 2005).
structures their parse trees share. The difference
Equation 1 summarizes the TRLM ranking score
with Partial Tree Kernels (PTK) (Moschitti, 2006)
between questions Q1 and Q2 :
is that SPTK also considers word relations.
5
https://radimrehurek.com/gensim/ Besides the different variations of the model,
summarization/bm25.html which are well explained in Moschitti (2006) and
594
Filice et al. (2016), we designed SPTK in the fol-
lowing form. Equation 3 portrays the notation of EnsSP T K(Q1 , Q2 ) = γφ(Q1 , Q2 )+
the similarity metric among two questions’ con- (7)
(1 − γ)Φ(SP T K(TQ1 , TQ2 ))
stituency trees, i.e. TQ1 and TQ2 .
We will compare the performance of the ensem-
T K(TQ1 , TQ2 ) =
X X
∆(n1 , n2 ) (3)
ble implementation with and without SPTK. For
n1 ∈NT
Q1
n2 ∈NT
Q2
distinction, in the following sections we will refer
to the former ensemble method as Ensemble and
NTQ1 and NTQ2 are the respective sets of nodes the latter as EnsSPTK.
of parse trees TQ1 and TQ2 . ∆(n1 , n2 ) is com-
puted in distinct forms according to three condi- 3 Experiments
tions. (1) If the production rules of TQ1 on n1 3.1 Data
and TQ2 on n2 are different, then ∆(n1 , n2 ) = 0.
(2) If n1 and n2 are similar preterminals, then Qatar Living We ran our experiments on the
∆(n1 , n2 ) = Sim(wn1 , wn2 ), where Sim is sim- data of SemEval 2016-2017 - Task 3 based on
ilar to Mij in Equation 2, as well as wn1 and wn2 the Qatar Living corpus6 . Its training split con-
are the terminal words for n1 and n2 , respectively. sists of 267 target questions, 2,669 related ques-
(3) If the production rules of TQ1 on n1 and TQ2 tions (around 10 for each target question), and
on n2 are the same and both are not preterminals, 26,690 comments (around 10 per related ques-
then tion). The development split and test sets of
2016 and 2017 have 50, 70, and 88 target ques-
tions, respectively (with the same proportion of
child(n1 )
Y related questions and comments as the training
∆(n1 , n2 ) = ∆(child(n1 )j , child(n2 )j ) (4)
j=1
set). Given a target question, each of its related
questions, retrieved by Google, was manually an-
So given a pair of constituency tree questions notated as “Perfect Match”, “Relevant” or “Irrel-
p = hTQ1 , TQ2 i to have their relevance scored and evant”. The shared-task also provided a large
a training set of pair trees C, features are extracted unannotated dataset of Qatar Living, with 189,941
in the following way: questions and 1,894,456 comments. In the Qatar
Living corpus, each question is formed by a sub-
ject and a body. For the models BM25, TRLM
SP T K(TQ1 , TQ2 ) = {T K(TQ1 , Tc1 ) + T K(TQ2 , Tc2 )}
and SoftCosine, we treat a question combining the
(5)
subject and body into a single text, whereas we
where hTc1 , Tc2 ii ∈ C only use the subject for SPTK.
The extracted kernel is used in Support Vector Quora To mitigate duplicate question pages at
Machines Φ, whose output decision function is the scale, Quora motivated the development of auto-
relevance score among TQ1 and TQ2 . mated ways of detecting these questions by re-
Ensemble is the method we propose to com- leasing a dataset with 400,000 pairs of questions
bine the relevance scores produced by the previ- together with a label for each entry indicating
ous approaches into a single model. Given ques- whether they are semantically identical (i.e., du-
tions Q1 and Q2 , we trained a Logistic Regression plicates) or not7 . We used this dataset to evaluate
φ(Q1 , Q2 ) with the relevance scores of BM25, our most relevant models in the task of detecting
TRLM and SoftCosine as features: question duplicates.
3.2 Settings
Ensemble(Q1 , Q2 ) = φ(Q1 , Q2 ) (6) For the Translation model (TRLM), C was com-
puted based on the training questions of the dataset
After empirically testing different settings, we used in the evaluation (e.g., Qatar Living or
found that the integration of SPTK in the ensem-
6
ble method was most effective when interpolating http://alt.qcri.org/semeval2017/
task3/index.php?id=data-and-tools
its relevance score separately with the outcome of 7
http://qim.fs.quoracdn.net/quora_
formula 6. Equation 7 denotes the model: duplicate_questions.tsv
595
Quora). In the Qatar Living corpus, the unanno- task, we evaluated all the models using the word-
tated part of the data is also used to compute C. translation probabilities, plus the cosine similar-
Across the experiments, hyperparameters of the ity measure depicted in Equation 2. In the latter,
models such as σ and α of TRLM and γ of Ensem- besides Word2Vec representations, we also tested
ble with SPTK were optimized in the development fastText (Bojanowski et al., 2017), a distribution
split of the data through Grid Search. Moreover, which takes character-level information and tends
Support Vector Machines in SPTK and Logistic to overcome spelling variations, and the top layer
Regression in Ensemble were implemented based of ELMo (Peters et al., 2018). To equalize the tri-
on the Scikit-Learn toolkit (Pedregosa et al., 2011) als, the data used by the models were lowercased
and had their hyperparameters tuned by cross- and stripped of stop words and punctuation.
validation on the training set.
4 Results
3.3 Evaluation
The first section of Table 1 lists the MAP of the
In the SemEval shared-task, the question- preprocessing methods in the development part of
similarity task was treated as a binary classifica- the corpus for each model. Although the best com-
tion task, where the models aim to predict whether bination of preprocessing methods differs between
a related question is “Perfect Match/Relevant” or models, we see that preprocessing the data is ben-
“Irrelevant”. We evaluate the models using the eficial for the performance of all models, except
Mean Average Precision (MAP) as the main met- for SPTK. Between the best results, we see that
ric, and also report F-Score for the classification suppression of punctuation is beneficial for all the
models. In the Quora dataset, we evaluated the models, while the removal of stopwords and low-
performance of our models in predicting question ercasing are detrimental to BM25 and TRLM, re-
duplicates also using the F-Score measure. spectively.
The lower part of Table 1 lists the performance
3.4 Experiment 1: Preprocessing of each model according to the different word-
From the top models for question similarity, lit- level similarity metrics. The use of Word trans-
tle is known about the design process of their lation probabilities appears the under-performing
preprocessing methods. Filice et al. (2016) do method out of the five, showing the power of the
not report on the preprocessing that they applied, continuous word representations. Surprisingly, we
and Charlet and Damnati (2017) lowercased the do not observe an improvement of fastText over
text as well as removed stopwords and punctu- Word2Vec representations. Even though CQA fo-
ation. So in our first experiment, we evaluated rums may have very noisy texts, the character-
BM25, TRLM, SoftCosine, Ensemble and En- level information that fastText takes into account
sSPTK with 3 preprocessing methods (and all apparently does not help. Using the top layer of
combinations of them): lowercasing, removal of ELMo concatenated with Word2Vec representa-
stopwords8 and suppression of punctuation. For tions leads to the best results in encoding the re-
SPTK we only apply lowercasing, since its con- lation between words, except with TRLM.
stituency trees contain punctuation and stopwords
Final Results Table 2 lists the results of the
as terminals. The preprocessing methods were ap-
models with their best settings in the test sets
plied in the training, development, test and unan-
of SemEval 2016-2017: BM25 with lowercased
notated parts of the data, such that probabili-
data without punctuation; TRLM with Word2Vec
ties and word distributions (e.g., word translation
without stop words and punctuation; SoftCosine
probability, Word2Vec, etc.) were affected.
with Word2Vec+ELMo, lowercased data with-
3.5 Experiment 2: Word-Similarity out stop words and punctuation; and SPTK with
Word2Vec+ELMo and lowercased data. The table
A central component of all of the evaluated mod- also shows the results of the best baseline (e.g.,
els except BM25 is the use of a word-similarity Google) and the winners of the SemEval 2016-
metric. To evaluate which distribution better cap- 2017 challenges. As expected, our best models
tures the similarity between two words for the were the ensemble approaches (e.g., Ensemble and
8
We used the list of English stopwords provided by the EnsSPTK), which combine the ranking scores of
NLTK framework (Bird and Loper, 2004) all the other evaluated approaches and outperform
596
Preproc. BM25 TRLM SoftCosine SPTK Ensemble EnsSPTK
L.S.P. 68.80 68.43 72.75 - 71.62 72.40
L.S. 67.31 63.25 69.15 - 69.50 71.29
L.P. 69.95 68.42 65.33 - 68.70 69.16
S.P. 66.03 68.65 68.56 - 68.67 70.37
L. 67.07 66.42 63.68 54.34 67.04 67.41
S. 63.77 64.53 67.01 - 67.85 68.36
P. 65.05 64.38 60.04 - 65.31 66.66
- 63.52 64.95 60.66 54.44 63.08 64.31
Metric BM25 TRLM SoftCosine SPTK Ensemble EnsSPTK
Translation - 68.43 70.75 48.10 70.80 70.80
Word2Vec - 72.90 72.75 54.44 71.40 72.64
fastText - 70.93 71.07 53.49 71.92 71.92
Word2Vec+ELMo - 71.41 73.89 54.78 73.90 74.63
fastText+ELMo - 70.56 73.43 54.77 73.73 73.73
Table 1: MAP results on the different preprocessing and word-relation metric conditions in the develop-
ment set. In the first part, L., S. and P. denote lowercase, stop words removal and punctuation suppression
methods respectively.
597
Quora results Based on the previous results,
we also evaluated the performance of our best
question-similarity model, Ensemble9 , in classi-
fying question duplicates on the Quora dataset.
Table 3 depicts the results of our ensemble
method with and without preprocessing and
using two similarity metrics (Word2Vec and
Word2Vec+ELMo). The best F-Score was
obtained by the version which preprocesses
the questions and represents the words with
Word2Vec+ELMo. Results were statistically sig-
nificant with p < 0.05 according to the McNe-
mar’s test.
5 Error Analysis
To obtain insight into the improvement by prepro-
cessing setting, in Figure 1 we present the per-
centage of similar questions that were ranked bet-
ter (placed on a higher position formerly occupied
by a non-similar), equally or worse (switched a
lower position with a non-similar) for each model-
preprocessing combination in the Qatar Living
corpus. The graph shows that each preprocessing
manipulation results in both improved rankings
Figure 1: Percentage of similar questions that
and worsened rankings. The model that is least af-
were ranked better, equally or worse after any of
fected by the preprocessing steps is BM25, which
the preprocessing manipulations or similarity met-
shows to be a stable baseline. Most gain is seen for
rics combined with each system, in comparison
the SoftCosine model with all preprocessing steps,
to the standard setting without preprocessing (first
where 38% of the duplicates are ranked better and
graph) or the standard similarity metric for each
only 9% is ranked lower than a non-duplicate. Re-
system (second graph).
garding the preprocessing steps applied in isola-
tion, lowercasing leads to most changes for TRLM
and BM25, while SoftCosine is most affected after The performance patterns presented in Figure 1
removing stopwords. show that the SoftCosine metric is affected most
The changes in performance by similarity met- by the presence of stopwords. Explicit evidence is
ric are also presented in Figure 1. The highest presented in Figure 2, which depicts the scores of
gains are seen for the TRLM model, which yields the SoftCosine settings with and without prepro-
an improved ranking for over 20% of the dupli- cessing in relation to the number of stopwords in
cates and a poor re-ranking for 12% to 14% when a a question-pair. The setting without preprocessing
similarity metric other than alignment is used. The shows a correlation with the number of stopwords:
SPTK model is not helped by a different similar- the similarity score goes up as the number of stop-
ity metric, with the most detrimental effect when words increases. The setting with preprocessing
combining the model with the translation align- is, as expected, robust to the number of stopwords.
ment or fastText. The SoftCosine model with the This shows that the SoftCosine metric is consider-
default Word2Vec is also rather robust, with only ably affected by the inclusion of stopwords, which
a slight improvement when applying one of the hampers performance for the task of question sim-
ELMo metrics. These metrics do affect the rank- ilarity.
ings considerably, but lead to fairly equal improve- In Table 4 we present examples of question
ments and declines of the ranking quality. pairs in the Qatar Living development set along
9
Given the size of the Quora dataset, computing the kernel with their Gold standard label and the preprocess-
trick of SPTK would be intractable. ing steps or model that yielded a proper ranking
598
Figure 2: Similarity score in relation to the number of stopwords in a question pair, for the SoftCosine
settings with and without preprocessing.
Gold
Question Best per-
Target question (subject - body) Related question (subject - body) stan-
ID formance
dard
QLing after working hours - how Surfing QL during office hours - How Stopword
Q314 many of you logon to QL after your many hour(s) everyone spend surfing Similar removal;
working hours? QL during office hours? SoftCosine
Beach cleaning - I am planning to
NOT ONE PUBLIC BEACH IN
organize a community service
DOHA? - Surrounded by pleasant Not No lower-
Q278 specifically beach cleaning to be
waters and not ONE public beach! similar casing
carried out by our company staff. Any
Ridiculous.
good suggestion of a beach?
Water theme park in qatar - Hai
Friends........ Any one knows the any water theme park in qatar? - Do
Q293 location of watar theme park in qatar; you know about any water theme park Similar SPTK
Is it beatiful? childrens have enough in qatar?
ride?? and howmuch fee Thanks
for this pair. The first example, with question good assessment by SPTK is likely due to the cen-
ID Q314, is of a similar question pair that was tral phrase ‘watar theme park in qatar’, which is
most often ranked in a high position by settings recurring, albeit with a different spelling, in the
that included stopword removal and the SoftCo- related question.
sine model. Stopword removal shows particularly
effective for the target question by removing 7 of 6 Discussion and Conclusion
the 15 words, and SoftCosine best matches ‘of- Until now, careful preprocessing and smart com-
fice hours’ to ‘working hours’. The second ques- bining of methods have remained understudied in
tion pair is exemplary of cases where preprocess- the field of community question answering. Our
ing is actually detrimental. The related question, results highlight that both pay off, yielding state-
not similar to the target question, is partly written of-the-art results. Our findings show that lower-
in capitals. After lowercasing, the word ‘beach’ casing the input and removing both punctuation
is matched with the target question, which might and stopwords yields the most robust outcomes,
result in a higher similarity score than questions especially for the SoftCosine metric. In addition,
that are actually similar. The final example, with representing the meaning of words by means of
ID Q293, is particularly well ranked by the SPTK Word2Vec combined with the top layer of ELMo
model. On its own, SPTK did not compete with is the most beneficial word similarity implemen-
the other models in our study, but the focus on tation. Combining several metrics implemented
syntactic tree kernels could add a valuable angle with these optimal settings into an ensemble sys-
to the similarity assessment. In this example, the tem based on logistic regression yields the best
599
performance in terms of F1-score, being compet- Delphine Charlet and Geraldine Damnati. 2017. Sim-
itive with the winners of the SemEval tasks, and bow at semeval-2017 task 3: Soft-cosine semantic
similarity between questions for community ques-
using fewer components.
tion answering. In Proceedings of the 11th Interna-
The error analysis showed that the BM25 model tional Workshop on Semantic Evaluation (SemEval-
is most stable across different preprocessing met- 2017). pages 315–319.
rics, while the SoftCosine model mostly profits
Danilo Croce, Alessandro Moschitti, and Roberto
from preprocessing. Given the semantic matching Basili. 2011. Structured lexical similarity via convo-
that is done as part of SoftCosine and is absent in lution kernels on dependency trees. In Proceedings
BM25, we can infer that preprocessing is an im- of the Conference on Empirical Methods in Natu-
portant prerequisite for effectively ranking ques- ral Language Processing. Association for Computa-
tional Linguistics, pages 1034–1046.
tion pairs based on semantic links.
Most of our experimentation was conducted on Simone Filice, Danilo Croce, Alessandro Moschitti,
the Semeval dataset, in which similarity between and Roberto Basili. 2016. Kelp at semeval-2016
questions is labeled. We also showed that ad- task 3: Learning semantic relations between ques-
tions and answers. In Proceedings of the 10th
justing preprocessing and word similarity settings International Workshop on Semantic Evaluation
leaded to better results in the task of identifying (SemEval-2016). pages 1116–1123.
question duplicates, in the Quora dataset. More
Marc Franco-Salvador, Sudipta Kar, Thamar Solorio,
research is needed to see whether the patterns that
and Paolo Rosso. 2016. Uh-prhlt at semeval-
we find are dataset-independent. 2016 task 3: Combining lexical and semantic-based
In future work we aim to compare the optimal features for community question answering. In
models from our current study in a real-world set- Proceedings of the 10th International Workshop
ting, by running A/B testing on a open-domain on Semantic Evaluation (SemEval-2016). Associa-
tion for Computational Linguistics, pages 814–821.
CQA platform. Through clicks and likes by the https://doi.org/10.18653/v1/S16-1126.
users of such a platform, we can obtain insights
into the value of these models when applied in the Jiwoon Jeon, W Bruce Croft, and Joon Ho Lee. 2005.
Finding similar questions in large question and an-
wild with many different question topics. swer archives. In Proceedings of the 14th ACM in-
ternational conference on Information and knowl-
Acknowledgements edge management. ACM, pages 84–90.
This work is part of the research programme Dis- Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig.
cussion Thread Summarization for Mobile De- 2013. Linguistic regularities in continuous space
vices, which is financed by the Netherlands Or- word representations. In Proceedings of the 2013
Conference of the North American Chapter of the
ganisation for Scientific Research (NWO). We Association for Computational Linguistics: Human
thank the reviewers for their valuable comments. Language Technologies. pages 746–751.
600
Linguistics, San Diego, California, pages 525–545.
http://www.aclweb.org/anthology/S16-1083.
Lawrence Page, Sergey Brin, Rajeev Motwani, and
Terry Winograd. 1999. The pagerank citation rank-
ing: Bringing order to the web. Technical report,
Stanford InfoLab.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Pretten-
hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-
sos, D. Cournapeau, M. Brucher, M. Perrot, and
E. Duchesnay. 2011. Scikit-learn: Machine learning
in Python. Journal of Machine Learning Research
12:2825–2830.
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt
Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. 2018. Deep contextualized word rep-
resentations. In Proceedings of the 2018 Confer-
ence of the North American Chapter of the Associ-
ation for Computational Linguistics: Human Lan-
guage Technologies, Volume 1 (Long Papers). Asso-
ciation for Computational Linguistics, pages 2227–
2237. https://doi.org/10.18653/v1/N18-1202.
601
A Classification-Based Approach to Cognate Detection Combining
Orthographic and Semantic Similarity Information
602
Proceedings of Recent Advances in Natural Language Processing, pages 602–610,
Varna, Bulgaria, Sep 2–4 2019.
the existing research and methodologies to cog- netically transcribed words by taking the edit dis-
nate detection. Section 3 describes the data and tance between them. For this research, we opted
annotation process used to create the context- to only focus on the orthographic proximity, as
independent gold standard for English-Dutch cog- sound metrics require an additional phonetic tran-
nate pairs, while Section 4 gives an overview of scription, thus making them less-likely to be ap-
the experimental setup and the two types of in- plied on large data sets. Moreover, Schepens et
formation sources, viz. orthographic and semantic al. (2013) find that there is a high consistency be-
similarity features, that were used. In section 5, we tween orthographic and phonetic similarity mea-
report on the results of our classifier (1) incorpo- sures for Dutch-English cognate pairs.
rating only orthographic features and (2) combin-
ing orthographic and semantic similarity features.
Section 6 concludes this paper and gives directions Whereas orthographic and phonetic features
for future research. have often been employed to model the similar-
ity between candidate cognate pairs, semantic in-
2 Related Research formation has often been ignored. Mitkov (2007)
beliefs that this is another result of the main fo-
Extensive lists of known cognates and false friends cus of investigation, which is the identification of
are hard to find and expensive to compose, since cognates rather than distinguishing cognates from
they require a considerable amount of time and ef- false friends. Semantic evidence is, however, an
fort from trained lexicographers (Schepens et al., important information source, as it can not only
2012). Especially for low resource languages, be used to represent the semantic (dis)similarity
this constitutes a serious issue. Therefore, most between word pairs, but it can also further in-
NLP research on cognates has mainly focused crease the accuracy of cognate detection systems.
on the automatic detection of such cognate pairs. In his own research, Mitkov (2007) distinguishes
In the literature, there are three main methods between two types of semantic approaches: taxo-
to identify cognates: orthographic, phonetic and nomic and distributional semantic similarity mea-
semantic approaches. The oldest approaches to sures. Whereas the first group relies on the tax-
tackle this task involve simple string similarity onomic structure of a resource such as WordNet
metrics as the longest common subsequence ra- (Miller, 1995), the second approach relies on large
tio (Melamed, 1999) or the normalized Leven- corpora. The latter methods are based on the
shtein distance (Levenshtein, 1965). More re- Distributional Hypothesis (Harris, 1954), which
cently, however, the attention has been drawn to states that words that appear in similar contexts
machine learning techniques. For instance, Frunza tend to share similar meanings. The different ap-
et al. (2007) combine several orthographic simi- proaches that leverage this principle are typically
larity measures to train a machine classifier, while divided into two categories: count-based meth-
Gomes et al. (2011) design a new similarity met- ods, such as Latent Semantic Analysis, and pre-
ric that is able to learn spelling differences across dictive methods, such as neural probabilistic lan-
languages. guage models, which have gained a lot of popular-
Different types of approaches can also be com- ity in today’s NLP community. On the one hand,
bined to distinguish cognates, e.g. Kondrak et count-based models count how often a given target
al. (2004) join orthographic and phonetic informa- word co-occurs with its neighbor words in a large
tion to distinguish between similar drug names. text corpus, after which the resulting counts are
In order to capture the phonetic similarity be- mapped to a dense vector for each word. On the
tween words, Konrak (2000) further developed a other hand, predictive models directly try to pre-
software package, called ALINE, which portrays dict a word from its neighbors in terms of learned
phonemes as vectors of phonetic features, thus dense embedding vectors (Baroni et al., 2014).
creating a phonetic similarity measure. Neverthe- Word2vec (Mikolov et al., 2013) is a particularly
less, Heeringa et al. (2010) find that simple pho- computationally-efficient and popular example of
netic transcriptions still seem to outperform pho- predictive models for learning word embeddings
netic similarity metrics that are based on phonetic from raw text. In this research, we will incorpo-
features. Hence, Schepens et al. (2013) propose to rate the more recent fastText word embeddings as
calculate a substitution cost for each pair of pho- implemented by Bojanowski et al. (2017).
603
3 Data Creation 4. Proper name: proper nouns (e.g. persons,
companies, cities, countries, etc.) and their
To train and evaluate the cognate detection system,
derivations (e.g. American).
we created a novel context-independent gold stan-
dard by manually labelling English-Dutch pairs of 5. Error: word alignment errors and compound
cognates and false friends in bilingual term lists. nouns of which one part is a cognate but the
In this section, we describe how the lists of candi- other part is missing in one of the languages
date cognate pairs were compiled on the basis of (e.g. peripherals - aansturingsperipherals).
the Dutch Parallel Corpus (Macken et al., 2011)
and how a manual annotation was performed to 6. No standard: words that do not occur in the
create a gold standard for English-Dutch cognate dictionary (e.g. num connectors) and num-
pairs. bers (e.g. adm12006e, VI).
3.1 List of Candidate Cognate Pairs To decide on the correct label, we adopted a
To select a list of candidate cognate pairs, unsu- context-independent approach applying the fol-
pervised statistical word alignment using GIZA++ lowing procedure: (1) for every candidate cognate
(Och and Ney, 2003) was applied on the Dutch pair, the dictionary Grote Van Dale1 (henceforth:
Parallel Corpus (DPC). This high-quality paral- VD) was consulted; (2) the English word is looked
lel corpus for Dutch, French and English consists up in the VD, e.g. salon, (3) the Dutch translation
of more than ten million words and is sentence- is inspected in the VD, e.g. salon: “nice room”
aligned. It contains five different text types and is and salon: “(room for) gathering of people (e.g.
balanced with respect to text type and translation from the literary world)”.
direction. The automatic word alignment on the Based on the previously obtained information,
English-Dutch part of the DPC resulted in a list a decision is made: in case all meanings of the
containing more than 500,000 translation equiva- Dutch word correspond with the English word,
lents. A first selection was performed by applying we consider them “cognates”, in case only part
the Normalized Levenshtein Distance (NLD) (as of the Dutch meanings correspond with the En-
implemented by Gries (2004)) on this list of trans- glish word, we consider them “partial cognates”,
lation equivalents and only considering equiva- in case the words have different meanings, we con-
lents with a distance smaller than or equal to 0.5. sider them “false friends”. An example of par-
This resulted in a list with 28,503 Dutch-English tial cognates is the pair agent-agent: the Dutch
candidate cognate pairs, which was manually la- agent refers both to (1) a police man and to (2)
beled. a representative (e.g. business representative). As
only the second meaning of the Dutch word is ex-
3.2 Creation of Gold Standard pressed by the English agent, these words are con-
To create the gold standard for cognate detection, sidered partial cognates.
an extensive set of guidelines was established (La- Two important observations should be made.
bat et al., 2019). The guidelines propose a clearly Firstly, we accorded more fine-grained labels in
defined method for the manual labeling of the fol- the gold standard that are described in great detail
lowing six categories: in the annotation guidelines (Labat et al., 2019).
For cognates, a distinction was, for instance,
1. Cognate: words which have a similar form made between cognates of which Part-of-Speech
and meaning in all contexts. Conform with (PoS) and meaning are identical in both languages,
our working definition for cognates, the cognates that differ in PoS (e.g. organisatie-
source and target words do not need to be et- organizing) and cognates that differ in agreement
ymologically related. (e.g. organisatie-organisations). Secondly, it is
important to note that a successful dictionary look-
2. Partial cognate: words which have a simi- up never overruled the “proper name” annotation.
lar form, but only share the same meaning in The resulting gold standard is context-
some contexts. independent. Hence, it can be used for both
the development and the evaluation of machine
3. False friend: words which have a similar
form but a different meaning. 1
https://www.vandale.be/
604
learning models that deal with cognate detection. • Prefix divides the length of the shared pre-
Besides its applications in natural language fix by the length of the longest cognate in the
processing, the gold standard can also form an pair.
important new resource for further research on
cognates in linguistics, translation studies or • Dice (Brew and McKelvie, 1996) divides the
psycho-linguistics. number of common bigrams times two by the
total number of bigrams in the cognate pair,
4 Classification 2 × |bigrams(x)| ∩ |bigrams(y)|
as in .
|bigrams(x)| + |bigrams(y)|
This section describes the experimental setup and
the two types of information sources, viz. ortho- • Dice (trigrams) differs from Dice in that it
graphic similarity and semantic similarity, that uses trigrams instead of bigrams.
were incorporated for the experiments. • XDice is a variant of Dice as it uses bigrams
4.1 Experimental Setup that are created out of trigrams by deleting
the middle letter in them.
In this paper, cognate detection was approached
as a supervised classification task. To this end, we • XXDice incorporates the string positions of
applied Support Vector Machines as implemented the bigrams into its metric. Therefore, the de-
in sklearn (Pedregosa et al., 2011). nominator is no longer multiplied by two, but
The data set used for the binary classifica- 2
by .
tion experiments consisted of the COGNATE 1 + (pos(x) − pos(y))2
pairs (labels “cognate” and “partial cognate”) and • LCSR stands for the longest common subse-
NON-COGNATE pairs (labels “error” and “false quence ratio, which is two times the length
friend”). The categories of “proper name” and “no of the longest subsequence over the summed
standard” were removed from the data set as they length of both sequences.
are always identical translations and would boost
the performance of the system in an artificial way. • NLS or the Normalized Levenshtein Similar-
Table 1 gives an overview of the distribution of the ity equals one minus the minimum number of
two classes in the gold standard data set. edits required to change one string sequence
to another.
Cognate Non- Total
cognate pairs • LCSR (bigrams), NLS (bigrams), LCSR
GS 9,855 4,763 14,618 (trigrams), and NLS (trigrams) differ from
their standard metrics in that they use, respec-
Table 1: Distribution of the “cognate” and “non- tively, bigrams and trigrams to calculate their
cognate” class labels in the gold standard (GS). results.
In order to train and test the system, we per- • Jaccard index models the length of the in-
formed 5-fold cross-validation for which we fixed tersection of both cognate strings over the
our 5 subsamples. Hyperparameter optimisa- length of the union of these strings.
tion was performed by means of a 5-fold cross- • Jaro-Winkler similarity is the complement
validation grid search on the training folds, result- of the Jaro-Winkler distance. Word pairs that
ing in the following values: kernel = RBF, C = 5, from their beginning correspond to a set pre-
class weight = None and gamma = 5. fix length will receive higher scores.
4.2 Orthographic Similarity Features • Spsim option 1 and Spsim option 2 are
Fifteen different string similarity metrics were the only metrics which require supervised
applied on the candidate cognates to measure training, in order to learn grapheme map-
the formal relatedness between source and target pings between language pairs (Gomes and
words. Eleven of these fifteen metrics were also Pereira Lopes, 2011). They are trained
used by Frunza et al. (2007). The following list by performing 5-fold cross-validation on the
briefly summarizes the orthographic features im- positive instances (i.e. cognates) in the data
plemented: set. Therefore, we created two different train
605
sets: option 1 includes cognate pairs which Once the results for the Dutch and English
differ in agreement or PoS-tags, while option monolingual embeddings were loaded, the Dutch
2 only includes cognates and partial cognates. embeddings were mapped to the English vector
space by means of a pre-trained alignment ma-
4.3 Semantic Information trix (Smith et al., 2017). Since the embeddings are
Besides features that model formal similarity be- then situated in the same vector space, one can eas-
tween word pairs, we also included semantic in- ily compute the cosine similarity between the two
formation in our classifier. We opted for word words of a candidate cognate pair. Subsequently,
embeddings, as these have shown to be very ef- this cosine similarity was used as a semantic fea-
fective for various NLP tasks. In addition, word ture for our machine learning system.
embeddings have not yet been used for the task of
cognate identification. For the purpose of this re- 5 Experimental Results
search, we worked with fastText word embeddings This section describes the classification results for
that were pre-trained on the Wikipedia corpus with two sets of experiments, namely (1) a classifier
the skip-gram model proposed by Bojanowski et incorporating fifteen orthographic similarity fea-
al. (2017). The model was trained with the default tures and (2) a classifier combining the same set
parameters and the length of the vector was set to of orthographic similarity features with a semantic
300. A disadvantage of using text-formatted pre- feature resulting from computing the cosine simi-
trained embeddings is that we could not generate larity between the word embeddings of the cog-
embeddings for all words in the gold standard list. nate pair.
As a result, we only obtained word embeddings
for 12,433 instances, while we have orthographic 5.1 Experiment 1: Orthographic Features
information for 14,618 instances. Table 2 gives an
A first set of experiments was conducted to evalu-
overview of the distribution of the two classes in
ate the performance of the orthographic similarity
the full and reduced gold standard data sets. The
features for the task of cognate detection. Table 3
experimental results that we obtained for this sub-
lists the averaged precision, recall and F1-score for
set are presented in Section 5.2.
all individual orthographic similarity features and
their combination.
Cognate Non- Total
cognate pairs The results show a very good performance of
the classifier combining all orthographic similar-
Ortho 9,855 4,763 14,618
ity information (average F-score of 84%). Espe-
Semantic 8,935 3,498 12,433
cially precision improves considerably when com-
Table 2: Distribution of the “cognate” and “non- bining the different orthographic similarity met-
cognate” class labels in the full (Ortho) and re- rics. When looking into the results for the individ-
duced (Semantic) gold standard data sets. ual features, it is clear that some metrics perform
very well in isolation, such as LCSR and NLS,
We chose to work with fastText embeddings in- which obtain F-scores of around 85% for the pos-
stead of regular Word2Vec embeddings because itive class (“Cognates”) with good balance of pre-
the former model uses n-grams to train its em- cision and recall.
beddings. In contrast to the Word2Vec mod- To get further insight in the informativeness of
els, fastText can create word embeddings for out- the various orthographic features, we also trained
of-vocabulary words, which is especially impor- a conditional inference tree and random forest on
tant for low-frequent words. Although the cur- the cognate data set. Figure 1 visualizes the model
rent research only works with pre-trained word learned by the conditional inference tree at depth
entries, in future research we plan to add out-of- 3. The tree indicates which orthographic metric is
vocabulary words by training word embeddings on the most important for that node in the tree. As can
domain-specific corpora more similar to the DPC be observed in Figure 1, the longest common sub-
corpus that was used to extract the list of candi- sequence ratio is overall the most influential met-
date cognate pairs. This way, we hope to construct ric, followed by SpSim (option 1) and the Jaro-
embeddings for all word pairs in the gold standard Winkler similarity.
list. In addition to the conditional inference tree, a
606
Cognates Non-cognates Average score
Metric Prec Rec F-score Prec Rec F-score Prec Rec F-score
Prefix 77.43 87.84 82.31 65.17 47.03 54.62 71.30 67.44 68.46
Dice 73.38 91.99 81.63 65.04 30.91 41.84 69.21 61.45 61.73
Dice (3gr) 73.28 91.88 81.53 64.63 30.67 41.59 68.95 61.28 61.56
Jaccard 73.83 91.53 81.73 65.22 32.86 43.69 69.52 62.19 62.71
XDice 70.85 96.26 81.62 70.03 18.08 28.73 70.44 57.17 55.18
XXDice 76.10 92.54 83.52 72.15 39.88 51.35 74.12 66.21 67.43
LCSR 82.15 89.30 85.47 72.65 59.93 65.66 77.40 74.62 75.57
NLS 82.39 86.03 84.24 68.47 61.84 64.95 75.43 73.93 74.59
LCSR (2gr) 76.92 81.28 79.03 56.16 49.52 52.58 66.54 65.40 65.80
NLS (2gr) 76.80 81.02 78.84 55.74 49.31 52.26 66.27 65.17 65.55
LCSR (3gr) 73.28 91.88 81.53 64.63 30.67 41.59 68.95 61.28 61.56
NLS (3gr) 73.34 91.60 81.46 64.16 31.10 41.87 68.75 61.35 61.67
Jaro-Winkler 77.06 90.72 83.33 69.72 44.10 53.99 73.39 67.41 68.66
SpSim (opt.1) 86.01 79.01 82.35 62.83 73.38 67.68 74.42 76.19 75.02
SpSim (opt.2) 83.36 80.37 81.82 62.21 66.76 64.37 72.79 73.56 73.10
Combined 89.33 90.63 89.97 80.63 77.60 78.78 84.68 84.11 84.38
Table 3: Precision (Prec), Recall (Rec) and F1-score for the individual orthographic similarity features
and for the classifier combining all features (%).
random forest was trained in order to further inves- this set of experiments is only conducted on that
tigate the importance of each metric. Since a ran- part of the data set for which word embeddings
dom forest uses lots of seeds (in our case: 123) in were retrieved, we also added the updated perfor-
order to decide on the importance of each variable mance scores for all individual orthographic met-
individually, it provides a somewhat more repre- rics on this reduced data set.
sentative, validated picture of the influence of dif- The classification results listed in Table 4 show
ferent metrics. An additional Somers’ D value was some interesting findings. First of all, the em-
computed for the random forest in order to check beddings in isolation already obtain good classi-
the goodness of fit. With a correlation score of fication results for the “Cognates” class (F-score
0.9528739, our random forest forms a good model of 89.14%). Second, the classifier combining or-
for unseen data. Figure 2 shows that the model thographic and semantic similarity features clearly
agrees with the conditional inference tree in that outperforms the classifier only incorporating or-
it also classifies LCSR, SpSim (option 1) and the thographic information.
Jaro-Winkler similarity as important metrics for
An analysis of the output reveals that the se-
the identification of cognates. It does, however,
mantic information indeed helps to detect cognate
provide some additional information, as it shows
pairs showing less orthographic resemblance
that the normalized Levenshtein similarity is also
(e.g. east–oost, older–ouderen, widespread–
very influential for this binary classification task.
wijdverbreid, asleep–slaap, sweating–zweten,
shame–schaamte, belief–geloof, whole–hele,
5.2 Experiment 2: Orthographic and
swarm–zwerm, overheated–oververhitte). In
Semantic Features
addition, the word embedding information also
generates less false negatives. Examples of pairs
In a second set of experiments, we combined all that were wrongly labeled as cognates by the
orthographic similarity features with a semantic classifier relying on orthographic information
feature expressing the cosine similarity between and that are correctly labeled as non-cognates
the two word embeddings. Table 4 shows the re- by the combined classifier are: affects–effecten,
sult of the classifiers incorporating (1) only seman- unlocking–blokkering, investments–instrument,
tic information and (2) a combination of ortho- slit–gesplit, provide–profielen, brazier–branden,
graphic and semantic similarity information. As might–high, where–wateren. On the other hand,
607
Figure 1: Conditional Inference Tree with depth 3 trained on the orthographic similarity features.
Table 4: Precision (Prec), Recall (Rec) and F1-score for the classifiers incorporating the fifteen indi-
vidual orthographic features, the classifier incorporating only semantic information (Sem), the classifier
incorporating the combined orthographic information (Ortho) and the classifier incorporating both or-
thographic and semantic similarity features (Ortho + Sem).
the combined classifier rarely introduces addi- 6 Conclusion and Future Work
tional false negatives (seven instances in total,
e.g. lead–leiden, include–inhouden, docker– This paper presents preliminary experiments for
dokwerker) or additional false positives (three combining orthographic and semantic similarity
instances in total: told–toen, escapologist– information for cognate detection. The experi-
escapist, because–bepaalde). mental results already show promising scores for
608
Figure 2: Random Forest indicating the importance the different orthographic similarity features for the
current cognate identification task.
a classifier using merely orthographic similarity guage pair. In addition, this will enable us to per-
information. The results, however, revealed that form trilingual machine learning experiments and
adding semantic information capturing the cosine to gain useful insights into cross-lingual cognate
similarity between the word embeddings of the detection.
Dutch and English terms further improves the
classification results considerably. As a result, we
can conclude that combining orthographic and se- References
mantic similarity information is a viable approach M. Baroni, G. Dinu, and G. Kruszewski. 2014. Don’t
to automatic cognate detection. count, predict! A systematic comparison of context-
counting vs. context-predicting semantic vectors. In
As we presented proof-of-concept results in Proceedings of the 52nd Annual Meeting of the As-
this research, there is still a lot of room for fu- sociation for Computational Linguistics (Volume 1:
ture research. Firstly, the implementation of al- Long Papers). pages 238–247.
ternative word embeddings is an important di- S. Bergsma and G. Kondrak. 2007. Alignment-Based
rection for future work. We will perform addi- Discriminative String Similarity. In Proceedings of
the 45th Annual Meeting of the Association of Com-
tional experiments with (1) larger and different
putational Linguistics. pages 656–663.
(e.g. domain-specific) corpora and (2) other em-
bedding approaches to improve the semantic in- P. Bojanowski, E. Grave, A. Joulin, and T. Mikolov.
2017. Enriching Word Vectors with Subword Infor-
formation based on embedding distance. We are
mation. Transactions of the Association for Compu-
confident this will result in high-level quality em- tational Linguistics 5:135–146.
beddings for all candidate cognate pairs.
C. Brew and D. McKelvie. 1996. Word-pair extraction
Secondly, it would be interesting to perform for lexicography. In Proceedings of the 2nd Inter-
multi-class experiments, where a distinction is national Conference on New Methods in Language
made between cognates, false friends and non- Processing. pages 45–55.
related word pairs. To this end, a training and S. Castro, J. Bonanata, and A. Rosá. 2018. A High
evaluation corpus containing cognate candidates Coverage Method for Automatic False Friends De-
in context will be built and manually annotated. tection for Spanish and Portuguese. In Proceed-
ings of the Fifth Workshop on NLP for Similar
Finally, we plan to compile the corresponding Languages, Varieties and Dialects (VarDial 2018).
gold standard set for French-Dutch, which is also pages 29–36.
part of the Dutch Parallel Corpus. This will allow A. Ciobanu and L. Dinu. 2014. Automatic Detec-
an evaluation of our approach for a different lan- tion of Cognates Using Orthographic Alignment. In
609
52nd Annual Meeting of the Association for Compu- N. Ljubešić and D. Fišer. 2013. Identifying False
tational Linguistics, ACL 2014 - Proceedings of the Friends between Closely Related Languages. In
Conference. volume 2, pages 99–105. Proceedings of the 4th Biennial International Work-
shop on Balto-Slavic Natural Language Processing.
D. Crystal. 2008. A Dictionary of Linguistics and Pho- pages 69–77.
netics. The language library. Blackwell, 6th edition.
L. Macken, O. De Clercq, and H. Paulussen. 2011.
O. Frunza and D. Inkpen. 2007. A tool for detecting Dutch Parallel Corpus: A Balanced Copyright-
French-English cognates and false friends. In Actes Cleared Parallel Corpus. Meta 56(2):374–390.
de la 14me confrence sur le Traitement Automatique
I. D. Melamed. 1999. Bitext Maps and Alignment
des Langues Naturelles. Association pour le Traite-
via Pattern Recognition. Computational Linguistics
ment Automatique des Langues, pages 91–100.
25(1):107–130.
L. Gomes and J. G. Pereira Lopes. 2011. Measuring T. Mikolov, K. Chen, G. S. Corrado, and J. Dean.
Spelling Similarity for Cognate Identification. In 2013. Efficient Estimation of Word Representations
L. Antunes and H. S. Pinto, editors, Progress in Ar- in Vector Space. CoRR abs/1301.3781.
tificial Intelligence. Springer, pages 624–633.
G. A. Miller. 1995. WordNet: A Lexical Database for
S. T. Gries. 2004. Shouldn’t It Be Breakfunch? A English. Communications of the ACM 38(11):39–
Quantitative Analysis of Blend Structure in English. 41.
Linguistics 42(3):639–667.
R. Mitkov, V. Pekar, D. Blagoev, and A. Mulloni.
S. Hansen-Schirra, J. Nitzke, and K. Oster. 2017. Pre- 2007. Methods for extracting and classifying pairs
dicting cognate translation. In S. Hansen-Schirra, of cognates and false friends. Machine Translation
O. Czulo, and S. Hofmann, editors, Empirical mod- 21(1):29–53.
elling of translation and interpreting, Language Sci-
F.J. Och and H. Ney. 2003. A Systematic Comparison
ence Press, chapter 1, pages 3–22.
of Various Statistical Alignment Models. Computa-
tional Linguistics 29(1):19–51.
Z. S. Harris. 1954. Distributional Structure. WORD
10(2-3):146–162. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, A. Mller,
B. Hauer and G. Kondrak. 2011. Clustering Seman- J. Nothman, G. Louppe, P. Prettenhofer, R. Weiss,
tically Equivalent Words into Cognate Sets in Mul- V. Dubourg, J. Vanderplas, A. Passos, D. Courna-
tilingual Lists. In Proceedings of 5th International peau, M. Brucher, M. Perrot, and E. Duchesnay.
Joint Conference on Natural Language Processing. 2011. Scikit-learn: Machine Learning in Python.
pages 865–873. Journal of Machine Learning Research 12:2825–
2830.
W. Heeringa, J. Nerbonne, and P. Osenova. 2010.
Detecting contact effects in pronunciation. In T. Rama. 2016. Siamese convolutional networks
M. Norde, B. de Jonge, and C. Hasselblatt, editors, based on phonetic features for cognate identifica-
Language Contact. New Perspectives., Benjamins, tion. CoRR abs/1605.05172.
pages 131–154.
J. Schepens, T. Dijkstra, and F. Grootjen. 2012. Dis-
G. Kondrak. 2000. A New Algorithm for the Align- tributions of cognates in Europe as based on Leven-
ment of Phonetic Sequences. In Proceedings of the shtein distance. Bilingualism: Language and Cog-
1st North American Chapter of the Association for nition 15(1):157–166.
Computational Linguistics Conference. pages 288–
J. Schepens, K. Paterson, T. Dijkstra, F. Grootjen, and
295.
W. J. B. van Heuven. 2013. Cross-Language Distri-
butions of High Frequency and Phonetically Similar
G. Kondrak and B. Dorr. 2004. Identification of Con- Cognates. PLOS ONE 8:1–15.
fusable Drug Names: A New Approach and Evalua-
tion Methodology. In Proceedings of the 20th Inter- M. Shlesinger and B. Malkiel. 2005. Comparing
national Conference on Computational Linguistics. Modalities: Cognates as a Case in Point. Across
pages 952–958. Languages and Cultures 6(2):173–193.
S. Labat, L. Vandevoorde, and E. Lefever. 2019. Anno- S. L. Smith, D. H. P. Turban, S. Hamblin, and N. Y.
tation Guidelines for Labeling English-Dutch Cog- Hammerla. 2017. Offline bilingual word vectors, or-
nate Pairs, version 1.0. Technical report, Ghent Uni- thogonal transformations and the inverted softmax.
versity, LT3 15-01. CoRR abs/1702.03859.
610
Resolving Pronouns for a Resource-Poor Language, Malayalam
Using Resource-Rich Language, Tamil
611
Proceedings of Recent Advances in Natural Language Processing, pages 611–618,
Varna, Bulgaria, Sep 2–4 2019.
focusing on Dravidian family of languages. His- retained the Proto –Dravidian verbs. For example,
torically, all Dravidian languages have branched the word for “talk” in Malayalam is “samsarik-
out from a common root Proto-Dravidian. kuka” “to talk” which has root in Sanskrit,
Among the Dravidian languages, Tamil is the old- whereas the Tamil equivalent is “pesuu” “to talk”,
est language. Though there is similarity at the syn- the root in Pro-Dravidian.
tactic level, there is no similarity in lexical form The Syntactic Structure Level: There is lot of sim-
or at the script level among the languages. We are ilarity at the syntactic structure level between the
motivated by the observation that related lan- two languages. Since antecedent to anaphor has
guages tend to have similar word order and syn- dependency on the position of the noun, the struc-
tax, but they do not have similar script or orthog- tural similarity is a positive feature for our goal.
raphy. Hence words are not similar. The syntactic similarity at Sentence level, Case
maker level, pronominal distribution level are ex-
Tamil has the most resources at all levels of lin- plained with examples.
guistics, right from morphological analyser to dis- Case marker level: Both the languages have the
course parser and Malayalam has the least. About same number of cases and their distribution is
the similarity of the two languages we give in de- similar. In both the languages, nouns inflected
tail in section 2. with nominative or dative case become the subject
of the sentence (Dative subject is the peculiarity
The remainder of this article is organized as fol- of Indian languages). Accusative case denotes the
lows: Section 2 provides a detailed description of object.
the two languages, their linguistic similarity cred- Clausal sentences: The clause constructions in
its and differences, Section 3 presents the pro- both the languages follow the same rule. The
nominal resolution in Tamil. In Section 4 we in- clauses are formed by nonfinite verbs. The clauses
troduce our proposed approach on how the lan- do not have free word order and they have fixed
guage model for Tamil can be used for resolving positions. Order of embedding of the subordinate
Malayalam pronouns. Section 5 describes the da- clause is same in both the languages.
tasets used for evaluation, experiments and analy- Ex:1
sis of the results and the paper ends with the con- (Ma) [innale vanna(vbp) kutti ]/Sub-RP-cl
clusion in Section 6. {sita annu}/Main cl
2 How similar the languages Tamil and (Ta) [neRu vanta(vbp) pon ]/Sub- RP-cl
{sita aakum}/Maincl
Malayalam Are?
As mentioned earlier, both the languages belong [Yesterday came girl ]/subcl
{Sita is}/ Maincl
to the Dravidian family and are relatively free
word order languages, inflectional (rich in mor- (The girl who came yesterday is Sita)
phology), agglutinative and verb final. They have
Nominative and Dative subjects. The pronouns As can be seen from the above example, the basic
have inherent gender marking as in English and syntactic structure is the same in both the lan-
have the same lexical form “avan” “he”, “aval” guages. The above example is a two clause sen-
“she” and “atu” “it” both in Tamil and Malaya- tence with a relative participial clause and a main
lam. Though the pronouns have same lexical form clause. The relative participial clause is formed by
and meaning, it can be said that there is no lexical the nonfinite verb (vbp). Using the same example
similarity between the two languages. The simi- we can find the pronominal distribution.
larity between two languages can be at three lev- Ex: 2
els, a) writing script, b) the word forms and c) the (Ma) [innale vanna(vbp) avali (PRP)
syntactic structure. ]/Sub-RP-cl {sitai annu}/Main cl
(Ta) [neRu vanta(vbp) avali (PRP)
Script Level: The two languages have different ]/Sub-RP-cl {sitai aakum}/Maincl
writing form, though the base is from Grandha
[Yesterday came shei (PRP)
script. Hence no similarity at the script level. ]/subcl { Sitai is} / Maincl
The Word Level: There is no similarity at the lex- (The she who came yesterday is Sita)
ical level between the two languages. The San-
skritization of Malayalam contributed to have In the above example the pronoun “aval” “she”
more Sanskrit verbs in Malayalam whereas Tamil
occurs at the same position in both the languages
612
and the antecedent “sita” also occurs at the same Students(N) school(N)+dat go(V)+present+3pl
position as shown by co-indexing. Consider an- (Students are going to the school)
other example. 4b. avarkaLi veekamaaka natakkinranar.
They(PN) fast(ADV) walk(V)+present+3pl
Ex:3 (They are walking fast.)
(Ma). sithaai kadaikku pooyi. avali pazham Considering Ex 4a and Ex.4b, sentence Ex.4b has
Vaangicchu(Vpast) 3rd person plural pronoun ‘avarkaL’ as the subject
(Ta). sithaai kadaikku cenRaal. avali pazham and it refers to plural noun ‘maaNavarkaL’ which
Vaangkinaal(V,past,+F,+Sg). is the subject in Ex.4a.
Sita shop went. She fruit
bought Ex:5
(Sitai went to the shop. Shei bought fruit.) 5a. raamuvumi giitavumj nanparkaL.
In the above example there are two sentences and Raamu(N)+INC Gita(N)+INC friends(N)
pronoun is in one sentence and antecedent is in (Ramu and Gita are friends.)
another. Here you can see the distribution of the
pronoun “aval” and where the antecedent “sita” is 5b. avani ettaam vakuppil padikkiraan.
He(PN) eight(N) class(N) study(V)+present+3sm
occurring. Though Tamil has number, gender and
(He studies in eight standard.)
person agreement between subject and verb and
Malayalam does not have, this cannot be consid- 5c. avaLumj ettaam vakuppil padikkiaal.
ered as a grammatical feature which can be used She(PN) eight(N) class(N) study(V)+present+3sf
for identifying the antecedent of an anaphor. This (She also studies in eight standard.)
grammatical variation does not have an impact on
In Ex.5b, 3rd person masculine pronoun ‘avan’ oc-
the identification of pronoun and antecedent rela-
curs as the subject and it refers to the masculine
tions. From the above examples we can see that noun ‘raamu’, subject noun in Ex.5a. Similarly,
the two languages have the same syntactic struc-
3rd person feminine pronoun ‘avaL’ in Ex.5c re-
ture at the clause and sentence level. We are ex-
fers to feminine noun ‘giita’ in Ex.3a. ‘atu’,
ploiting this similarity between the two languages
which is a 3rd person neuter pronoun, will also oc-
to achieve our goal. We find that using this simi-
curs as genitive/possessive case marker. Consider
larity between the languages, the language model the following example.
of Tamil can be used to resolve pronouns in Mal-
3.1.1 Non-anaphoric Pronouns
ayalam. The pronouns can also occur as generic mentions
without having referent. In English it known
3 Pronoun Resolution in Tamil as‘pleonastic it’.
3.1 Pronouns in Tamil Ex:6.
atu oru malaikalam.
In this section, we analyse in detail the pronomi-
It(PN) one(Qc) rainy_season (N)
nal expressions in Tamil. Pronouns are the words (It was a rainy season.)
used as a substitution to nouns, that are already
mentioned or that is already known. There are In Ex.6, the 3rd person neuter pronoun ‘atu’ (it) do
pronouns which do not refer. Pronouns in Tamil not have a referent. Here ‘atu’ is equivalent to ple-
have person (1st, 2nd, 3rd person) and number (sin- onastic ‘it’ in English
gular, plural) distinction. Masculine, feminine and
neuter gender distinctions are clearly marked in 3.1.2 Corpus annotation
3rd person pronouns, whereas in 1st and 2nd person
pronouns there is no distinction of masculine, We collected 600 News articles from various
feminine and neuter gender. In this work we con- online Tamil News wires. The News articles are
sider only third person pronouns. Third person from Sports, Disaster and General News domains.
pronouns in Tamil have inherent gender and as in The anaphoric expressions are annotated along
English and they are “avan” he, “aval” she and with its antecedents using graphical tool,
“atu” it. In this work, we resolve 3rd person pro- PAlinkA, a highly customisable tool for Dis-
nouns. The distribution of pronouns in various course Annotation (Orasan, 2003) which we cus-
syntactic constructions is explained with exam- tomized for Tamil. We have used two tags
ples below. namely, MARKABLE and COREF. The corpus
Ex:4 used for training is 54,563 words which are anno-
4a. maaNavarkaLi paLLikku celkiranar. tated for anaphora –antecedent pairs and testing
613
corpus is 10,912 words. We have calculated the approaches using Centering theory for Hindi.
inter-annotator agreement which is the degree of Sobha et al., (2007) presented a salience factor
agreement among annotators. We have used Co- based with limited shallow parsing of text. Aki-
hen’s kappa as the agreement statistics. The kappa landeswari et al., (2013) used CRFs for resolution
coefficient is generally regarded as the statistics of third person pronoun. Ram et al., (2013) used
of choice for measuring agreement on ratings Tree CRFs for anaphora resolution for Tamil with
made on a nominal scale. We got a Kappa score features from dependency parsed text. In most of
of 0.87. The difference between the annotators the published works resolution of third person
were analysed and found the variation in annota- pronoun was considered and it is a non-trivial
tion. It occurred in the marking of antecedents for task.
pronominal. This is common in sentences with Pronoun resolution engine does the task of identi-
clausal inversion, and genitive drop. fying the antecedents of the pronouns. The Pro-
noun resolution is built using Conditional Ran-
3.2 Pronoun Resolution System dom Fields (CRFs) technique. Though CRFs is
Early works in anaphora resolution by Hobbs notable for sequence labelling task, we used this
(1978), Carbonell and Brown (1988), Rich and technique to classify the correct anaphor-anteced-
Luper Foy (1988) etc. were mentioned as ent pair from the possible candidate NP pairs by
knowledge intensive approach, where syntactic, presenting the features of the NP pair and by
semantic information, world knowledge and case avoiding the transition probability. While training
frames were used. Centering theory, a discourse we form positive pairs by pairing anaphoric pro-
based approach for anaphora resolution was pre- noun and correct antecedent NP and negative
sented by Grosz (1977), Joshi and Kuhn (1979). pairs by pairing anaphoric pronouns and other
Salience feature based approaches were presented NPs which match in person, number and gender
by Lappin and Leass (1994), Kennedy Boguraev (PNG) information with the anaphoric pronoun.
(1996) and Sobha et al., (2000). Indicator based, These positive and negative pairs are fed to the
knowledge poor method for anaphora resolution CRFs engine and the language model is gener-
methods were presented by Mitkov (1997, 1998). ated. While testing, when an anaphoric pronoun
One of the early works using machine learning occurs in the sentence, the noun phrases which
technique was Dagan Itai’s (1990) unsupervised match in PNG with the pronoun, that occur in the
approach based on co-occurrence words. With the preceding portion of the sentence and the four pre-
use of machine learning techniques researchers ceding sentences are collected and paired with the
work on anaphora resolution and noun phrase anaphoric pronoun and presented to CRFs engine
anaphora resolution simultaneously. The other to identify the correct anaphor-antecedent pair.
machine learning approaches for anaphora resolu- 3.2.1 Pre-processing
tion were the following. Aone and Bennett (1995), The input document is processed with a sentence
McCarty and Lahnert (1995), Soon et al., (2001), splitter and tokeniser to split the document into
Ng and Cardia (2002) had used decision tree sentences and the sentences into individual tokens
based classifier. Anaphora resolution using CRFs which include words, punctuation markers and
was presented by McCallum and Wellner (2003) symbols. The sentence split and tokenized docu-
for English, Li et al., (2008) for Chinese and ments are processed with Syntactic Processing
Sobha et al., (2011, 2013) for English and Tamil. modules. Syntactic processing modules include
In Indian languages anaphora resolution engines Morphological analyser, Part-of-Speech tagger
are demonstrated only in few languages such as and Chunker. These modules are developed in
Hindi, Bengali, Tamil, and Malayalam. Most of house.
the Indian languages do not have parser and other a) Morphological Analyser: Morphological
sophisticated pre-processing tools. The earliest analysis processes the word into component
work in Indian language, ‘Vasisth’ was a rule morphemes and assigning the correct
based multilingual anaphora resolution platform morpho-syntactic information. The Tamil
by Sobha and Patnaik (1998, 2000, 2002), where morphological Analyser is developed using
the morphological richness of Malayalam and paradigm based approach and implemented
Hindi were exploited without using full-parser. using Finite State Automata (FSA) (Vijay
The case marker information is used for identify- Sundar et.al 2010). The words are classified
ing subject, object, direct and in-direct object. according to their suffix formation and are
Prasad and Strube (2000), Uppalapu et al., (2009) marked as paradigms. The number of
and Dekwale et al., (2013) had presented different paradigms used in this system is Noun
614
paradigms: 32; Verb Paradigms: 37; The features for all possible candidate antecedent
Adjective Paradigms: 4; Adverb Paradigms: and pronoun pairs are obtained by pre-processing
1. A root word dictionary with 1, 52,590 root the input sentences with morphological analyser,
words is used for developing the POS tagger, and chunker. The features identified
morphological anlyser. The morphological can be classified as positional and syntactic fea-
analyser is tested with 12,923 words and the tures
system processed 12,596 words out of which Positional Features: The occurrence of the can-
it correctly tagged 12,305. The Precision is didate antecedent is noted in the same sentence
97.69% and Recall = 97.46%. MA returns all where the pronoun occurs or in the prior sentences
possible parse for a given word. or in prior four sentences from the current sen-
tence.
b) Part Of Speech Tagger (POS tagger): Part Syntactic Features: The syntactic arguments of
of Speech tagger disambiguates the multiple the candidate noun phrases in the sentence are a
parse given by the morphological analyser, key feature. The arguments of the noun phrases
using the context in which a word occurs. The such as subject, object, indirect object, are ob-
Part of speech tagger is developed using the tained from the case suffix affixed with the noun
machine learning technique Conditional phrase. As mentioned in section 2 the subject of a
Random fields (CRF++) (Sobha L, et.al sentence can be identified by the case marker it
2010). The features used for machine learning takes. We use morphological marking for the
is a set of linguistic suffix features along with above.
statistical suffixes and uses a window of 3 a) PoS tag and chunk tag of Candidate NP, case
words. We have used 4, 50,000 words, which marker tags of the noun.
are tagged using BIS POS tags. The system b) The suffixes which show the gender which
performs with recall 100% and Average gets attached to the verb.
Precision of 95.16%.
c) Noun and Verb Phrase Chunker: Chunking 3.3.2 Development of Tamil Language Model
is the task of grouping grammatically related
words into chunks such as noun phrase, verb We used 600 Tamil Newspaper articles for build-
phrase, adjectival phrase etc. The system is ing the language model. The preparation of the
developed using the machine learning training data is described below. The raw corpus
technique, Conditional Random is processed with sentence splitter and tokeniser.
fields(CRF++) (Sobha L et.al 2010). The fea- The tokenized corpus is then preprocessed with
tures used are the POS tag, Word and window shallow processing modules, namely, morpholog-
of 5 words. Training Corpus is 74,000 words. ical analyser, part-of-speech tagger and chunker.
The recall is 100%. Average Precision of The training data is prepared from this processed
92.00%. corpus. For each pronoun, the Noun phrase (NP)
preceding the pronouns and in the NPs in preced-
3.3 Pronoun Resolution Engine ing sentence till correct antecedent NP, which
match in Person, number and Gender (PNG) are
In both training and testing phase, the noun selected for training. The above features are used
phrases (NP) which match with the PNG of the for CRF for learning. The system was evaluated
pronoun are considered. The features are ex- with data from the web and the result is given be-
tracted from these NPs. In the training phase the low.
positive and negative pairs are marked and fed to Domain Testing Precision Recall
the ML engine for generating a language model. Corpus (%) (%)
In the testing phase these NPs with its features are (Words)
input to the language model to identify the ante- News 10,912 86.2 66.67
cedent of a pronoun. Here we have not taken the data
lexical item or the word as a feature. We have
used only the grammatical tags as feature. The Table 1: Pronominal Resolution (CRFs engine)
features selected represent the syntactic position
of the anaphor –antecedent occurrence. 4. Resolution of Pronouns in Malayalam Cor-
3.3.1 Features Selection pus using Tamil Language Model
The features required for machine learning are
identified from shallow parsed input sentences.
615
In this section, we present in detail how Malaya- 1. [determiner] [quantifier][intensifier] [classi-
lam is tested using Tamil language model. Here fier][adjective] {Head Noun}. Here Head
Malayalam data is pre-processed as per the re- noun is obligatory and others are all optional
quirement of the Tamil language model test data. 2. NN+NN combination
The test data required four grammatical infor-
mation, i) POS, ii) the case marker, iii) the number 3. NN with no suffix+NN with no suffix……
gender and person and iv)chunk information. In +NN with suffix or without suffix.
the introduction we have asked three questions on
how to use a language model in source language Using the above rules we identified the noun
be used for testing a target language. The three chunks in Malayalam. The above discussed pre-
questions are dealt one by one below. processing gives an accuracy of 66% for POS and
1. Can the language model developed for suffix tagging and 63% for chunking.
one language be used for analyzing the 2. How the lexical form difference can be
other language? resolved in using the language model?
In this study we have used the language model of The second question we asked is about the lexical
the source language Tamil. The features used to form or words which are not similar in both the
develop this language model are POS tag, Chunk languages and how this can be resolved. The anal-
tag and the case/ suffix tags. The word form was ysis of Tamil has shown that the syntactic struc-
not considered as a feature. Since these are the ture of the language has more prominence over the
features used for learning, the test data also should words in the resolution of anaphors. Hence we
have these information. As said earlier, Malaya- have taken the syntactic features and did not take
lam is not a resource rich language and it does not word as a feature. The system learned only the
have pre-processing engines such as POS tagger, structure patterns.
Chunker and morphological analysers with high 3. How to overcome the challenges of script
accuracy. Hence we developed a very rudimen- variation?
tary preprocessing systems which can give the Since word feature is not considered the script do
POS tag, case/suffix tags and chunk tags. not pose any challenges. Still to have the same
The POS information: The POS and suffix infor- script we converted the two languages into one
mation are assigned to the corpus using a root form the WX notation. This helped in having the
word dictionary with part of speech information same representation of the languages.
and a suffix dictionary.
The dictionary has the root words which include 4. Testing with Malayalam Data
all types of pronouns and contains nearly 66,000 We selected 300 articles from Malayalam Wik-
root words. The grammatical information (POS) ipedia, which were on different genre and size.
such as noun, verb, adjective, pronoun and num- The 300 Malayalam documents from Wikipedia
ber gender person (PNG) information are given has 7600 sentences with 3660 3rd person pro-
for a word in the dictionary. nouns. The distribution of the pronouns is pre-
The suffix dictionary contains all possible suf- sented in Table 2.
fixes which a root word can take. The suffix in- Number of
cludes the changes in sandhi when added to the Pronoun Occurrences with
root word. The suffixes are of two types, i) that its Inflected forms
which gets attached to nouns called the case suf- avan (3rd person masculine
1120
fixes and 2) that which gets attached to verbs singular)
rd
called the TAM (Tense, Aspect, and Modal). Us- aval (3 person feminine
840
singular)
ing this suffix dictionary we can identify the POS
avar (3rd person honorific) 420
of the word even if the word is not present in the
athu (3rd person nuetor
root word dictionary. The suffix dictionary has 1280
singular)
1,00,000 unique entries. Total 3660
The noun chunk information is given by a rule Table 2. Distribution of 3rd person pronouns in the
based chunker which works on three linguistic corpus
rules. The noun phrases alone are required for As discussed in the earlier section, the prepro-
anaphora resolution. Chunks are identified using cessing done for Tamil using syntactic module are
the basic linguistic rule for Noun phrases as given morphological information, POS and Chunking
below information. Hence the same pre-processing in-
formation is necessary for Malayalam data as
616
well. The documents are initially preprocessed Precisi Recal
with a sentence splitter and tokenizer. The to- Type of pronoun
on (%) l (%)
kenized documents are pre-processed using the avan (3rd person masculine
70.83 69.52
pos, suffix and chunking systems discussed above singular)
to enrich the text with syntactic information. avaL (3rd person feminine
69.34 68.56
For each pronoun, we identify the possible candi- singular)
date antecedents. Those noun phrases that occur avar/avarkal (3rd person
65.45 70.34
preceding the pronoun in the current sentences plural/honorofic)
and preceding four sentences, which match in the atu (3rd person nuetor singular) 56.67 65.67
person, number gender (PNG) with the select pro- Total 68.45 67.34
noun are identified as possible candidates. For Table 3: Evaluation Results
these possible candidates we extract the features 6. Conclusion
required for CRFs techniques as explained in the In this paper we explained a method to use high
previous section. After extraction of features for resource language to resolve anaphors in less re-
selected candidate antecedents, the antecedent is source language. In this experiment the high re-
identified by using language model built using source language is Tamil and the less resource
Tamil data. The results are encouraging with 67% language is Malayalam. The results are encourag-
accuracy which is a respectable score. The errors ing. The model needs to be tested with more data
and evaluation is given in detail in the next sec- as future work.
tion.
References
5. Experiment and Discussion Carbonell J. G., and Brown R. D. 1988. Anaphora res-
olution: A multi- strategy approach. In: 12th
The experiment showed that the resolution of the International Conference on Computational Lin-
pronouns “avan” he and “aval” she is similar to guistics, 1988, pp. 96-101.
that of Tamil documents. The issues related to
split antecedent is not addressed and hence pro- Dagan I., and Itai. A. 1990. Automatic processing of
nouns which are referring to coordinated nouns large corpora for the resolution of anaphora refer-
ences. In: 13th conference on Computational lin-
were not resolved. The pronoun which is less re-
guistics, Vol. 3, Helsinki, Finland, pp.330-332.
solved is the third person neuter pronouns “atu”
compared to other pronouns. The third person Dakwale. P., Mujadia. V., Sharma. D.M. 2013. A Hy-
neuter pronoun usually has more number of pos- brid Approach for Anaphora Resolution in Hindi.
sible candidates, which leads to poor resolution. In: Proc of International Joint Conference on Natu-
Consider the following example. ral Language Processing, Nagoya, Japan, pp.977–
981.
Ex:7
avan joli ceytha jolikkarkku Li., F., Shi., S., Chen., Y., and Lv, X. 2008. Chinese
He(PRP) work(N) do(V)+past+RP worker(N)+pl+dat Pronominal Anaphora Resolution Based on Condi-
vellam kotuthu. tional Random Fields. In: International
water(N) give(V)+past Conference on Computer Science and Software En-
(He gave water to the workers who did the work.) gineering, Washington, DC, USA, pp. 731-734.
617
Parser. In: 16th International Conference on Com- Russell, B. 1919, “On Propositions: What They Are
putational Linguistics COLING’96, Copenhagen, and How They Mean,” Proceedings of the Aristote-
Denmark, pp. 113–118. lian Society, Supplementary Volume 2: 1–43; also
appearing in Collected Papers, Vol. 8
Lappin S., and Leass H. J. 1994. An algorithm for pro-
nominal anaphora resolution. Computational Lin- Senapati A., Garain U. 2013. GuiTAR-based Pronomi-
guistics 20 (4), pp. 535-561. nal Anaphora Resolution in Bengal. In: Proceedings
of 51st Annual Meeting of the Association for Com-
McCallum A., and Wellner. B. 2003. Toward condi-
putational Linguistics, Sofia, Bulgaria pp 126–130.
tional models of identity uncertainty with applica-
tion to proper noun coreference. In Proceedings of Sikdar U.K, Ekbal A., Saha S., Uryupina O., Poesio M.
the IJCAI Workshop on Information Integration on 2013. Adapting a State-of-the-art Anaphora Resolu-
the Web, pp. 79–84. tion System for Resource-poor Language. In pro-
ceedings of International Joint Conference on Natu-
McCarthy, J. F. and Lehnert, W. G. 1995. Using deci-
ral Language Processing, Nagoya, Japan pp 815–
sion trees for coreference resolution. In C. Mellish
821.
(Ed.), Fourteenth International Conference on Arti-
ficial Intelligence, pp. 1050-1055 Sobha L. and Patnaik B. N. 2000. Vasisth: An Anaph-
ora Resolution System for Indian Languages. In
Mitkov R. 1998. Robust pronoun resolution with lim-
Proceedings of International Conference on Ar-
ited knowledge. In: 17th International Conference
tificial and Computational Intelligence for Decision,
on Computational Linguistics (COLING’
Control and Automation in Engineering and Indus-
98/ACL’98), Montreal, Canada, pp. 869-875.
trial Applications, Monastir, Tunisia.
Mitkov, R. 1997. "Factors in anaphora resolution: they
Sobha L. and Patnaik,B.N. 2002. Vasisth: An anaphora
are not the only things that matter. A case study
resolution system for Malayalam and Hindi. In Pro-
based on two different approaches". In Proceedings
ceedings of Symposium on Translation Support
of the ACL'97/EACL'97 workshop on Operational
Systems.
factors in practical, robust anaphora resolution, Ma-
drid, Spain. Sobha L. 2007. Resolution of Pronominals in Tamil.
Computing Theory and Application, The IEEE
Ng V., and Cardie C. 2002. Improving machine learn-
Computer Society Press, Los Alamitos, CA, pp.
ing approaches to coreference resolution. In. 40th
475-79.
Annual Meeting of the Association for Computa-
tional Linguistics, pp. 104-111. Sobha L., Sivaji Bandyopadhyay, Vijay Sundar Ram
R., and Akilandeswari A. 2011. NLP Tool Contest
Prasad R., and Strube,M.,2000. Discourse Salience
@ICON2011 on Anaphora Resolution in Indian
and Pronoun Resolution in Hindi, Penn Working Pa-
Languages. In: Proceedings of ICON 2011.
pers in Linguistics, Vol 6.3, pp. 189-208.
Sobha Lalitha Devi and Pattabhi R K Rao. 2010. Hy-
Orasan, C. 2003, PALinkA: A highly customisable tool
brid Approach for POS Tagging for Relatively Free
for discourse annotation. SIGDIAL Workshop
Word Order Languages. In Proceedings of
2003: 39-43
Knowledge Sharing Event on Part-Of-Speech Tag-
Preslav Nakov Hwee Tou Ng 2012. Improving Statis- ging, CIIL, Mysore.
tical Machine Translation for a Resource-Poor Lan-
Vijay Sundar Ram and Sobha Lalitha Devi. 2010. Noun
guage Using Related Resource-Rich Languages;
Phrase Chunker Using Finite State Automata for an
Journal of Artificial Intelligence Research 44 (2012)
Agglutinative Language. In Proceedings of the
179-222;
Tamil Internet – 2010, Coimbatore, India, 218–
Ram, R.V.S. and Sobha Lalitha Devi. 2013."Pronominal 224.
Resolution in Tamil Using Tree CRFs", In Proceed-
Soon W. H., Ng, and Lim D. 2001. A machine learning
ings of 6th Language and Technology Conference,
approach to coreference resolution of noun phrases.
Human Language Technologies as a challenge for
Computational Linguistics 27 (4), pp.521-544.
Computer Science and Linguistics - 2013, Poznan,
Poland Taku Kudo. 2005. CRF++, an open source toolkit for
CRF, http://crfpp.sourceforge.net .
Rich, E. and LuperFoy S.,1988 An architecture for
anaphora resolution. In: Proceedings of the Second Uppalapu. B., and Sharma, D.M. 2009. Pronoun Reso-
Conference on Applied Natural Language Pro- lution For Hindi. In: Proceedings of 7th Discourse
cessing, Austin, Texas. Anaphora and Anaphor Resolution Colloquium
(DAARC 09), pp. 123-134.
618
Semantic Role Labeling with Pretrained Language Models for Known and
Unknown Predicates
Daniil Larionov Artem Shelmanov
FRC CSC RAS / Moscow, Russia Skoltech / Moscow, Russia
RUDN University / Moscow, Russia FRC CSC RAS / Moscow, Russia
dslarionov@isa.ru a.shelmanov@skoltech.ru
619
Proceedings of Recent Advances in Natural Language Processing, pages 619–628,
Varna, Bulgaria, Sep 2–4 2019.
The whole SRL process can be divided in four The developed models and the code are pub-
steps: predicate identification and identification lished online1 .
of its frame (disambiguation), argument extraction
(for each predicate), argument classification (or • We show that pretrained embeddings and lan-
labeling of arguments with semantic roles), and guage models can alleviate the problem of
global scoring that deals with linguistic constrains. annotation scarcity for predicates.
Predicate-argument structures in some notations
can be represented as two-level trees, rooted in • We conduct experiments that demonstrate the
predicates, with single tokens (nouns, adjectives, superiority of using embeddings generated
pronouns, proper names) as leaves that denote ar- by pretrained language models compared to
guments. We adopt this dependency-based nota- shallow embeddings like word2vec and Fast-
tion and treat the problem of semantic role label- Text.
ing as constructing such trees.
There are two main types of linguistic cor- The rest of the paper is structured as follows.
pora that are used for training models for SRL: Section 2 discusses the related work on seman-
FrameNet-like (Baker et al., 1998) and PropBank- tic role labeling for Russian and other languages.
like (Kingsbury and Palmer, 2002). The Russian- Section 3 describes the developed pipeline for se-
language resource that can be used for supervised mantic role labeling of Russian texts. Section 4
training is a FrameBank corpus (Lyashevskaya, presents the results of the experimental evaluation
2012; Lyashevskaya and Kashkin, 2015). The un- of the developed pipeline. Section 5 concludes and
derlying semantic model of this resource is close outlines the future work.
to the one FrameNet is based on. The biggest
difference from FrameNet besides semantic role 2 Related Work
inventory lies in the fact that FrameBank does The data-driven methods for semantic role label-
not group several verbs into frames but introduces ing originate from the work (Gildea and Juraf-
“frame” structures for each unique verb. The cor- sky, 2002), in which authors propose a statistical
pus contains partially annotated text samples with model based on various morpho-syntactic features
predicates, arguments, and their semantic roles. and train it on the FrameNet corpus. The release
A notable limitation of this resource is that there of the PropBank corpus sparked a notable inter-
are annotations only for a very limited set of pred- est in SRL among researchers. The consecutive
icates. In this work, we combat the problem of works and numerous shared tasks facilitated cre-
annotation scarcity by introducing two classifica- ation of elaborated machine learning models based
tion models that rely on different sets of features: on manually engineered lexico-syntactic features
one for “known” predicates that are present in the (Xue and Palmer, 2004; Punyakanok et al., 2005;
training set and one for “unknown” predicates that Pradhan et al., 2005).
are not seen in the training data. We show that
More recent works on semantic role label-
the model for “unknown” predicates can deal with
ing leverage deep neural networks (Collobert
the lack of annotation by using pretrained em-
et al., 2011) shifting from feature-engineering
beddings. We perform experiments with various
to architecture-engineering. Several notable ap-
types of embeddings including the ones generated
proaches suggest doing semantic role labeling in
by deep pretrained language models: word2vec
an end-to-end fashion relying only on raw low-
(Mikolov et al., 2013), FastText (Bojanowski et
level input consisted of characters or tokens and
al., 2017), ELMo (Peters et al., 2018), BERT (De-
well-known multilayer recurrent networks (He et
vlin et al., 2018), and show that embeddings gen-
al., 2017; Sahin and Steedman, 2018; Marcheg-
erated by deep pretrained language models are su-
giani et al., 2017). State-of-the-art approaches
perior to classical shallow embeddings for seman-
leverage multitask learning (Strubell et al., 2018)
tic role labeling in both cases of “known” and “un-
and self-attention techniques (Strubell et al., 2018;
known” predicates.
Tan et al., 2018). Several recent works also re-
The contribution of this paper is the following: port that although the end-to-end approaches have
• We present and evaluate the first full pipeline 1
https://github.com/IINemo/isanlp_srl_
for semantic role labeling of Russian texts. framebank/tree/master
620
managed to show comparable results, syntactic in- formed experiments on gold-standard morpholog-
formation still significantly helps semantic parsing ical features (POS tags and morphological char-
(He et al., 2018). acteristics), which does not reflect the real-world
scenario. In this work, we additionally suggest us-
It is worth noting a novel approach to creating
ing embeddings generated by deep pretrained lan-
annotated resources for semantic role labeling. In
guage models, train models on automatically gen-
(He et al., 2015; FitzGerald et al., 2018), instead
erated linguistic annotations (morphology / syn-
of annotating a corpora with a scheme grounded in
tactic trees), and provide the full pipeline for se-
elaborated linguistic theory, which requires highly
mantic role labeling including argument extrac-
qualified annotators, researchers suggest question-
tion. It is also worth noting the Frame-parser
answer driven approach to construction of anno-
project2 , however, it is in an early stage and only
tated resource based on crowd-sourcing. The re-
implements argument labeling using an SGD clas-
cently presented QA-SRL Bank 2.0 (FitzGerald et
sifier.
al., 2018) is a large-scale annotated dataset built by
non-experts. The construction of such a resource
3 Pipeline for Semantic Role Labeling
becomes possible due to simplicity of the anno-
tation scheme, which provides an ability to label The limitations of the FrameBank corpus do not
predicate-argument relationships using question- allow to use end-to-end / sequence labeling meth-
answer pairs. ods for SRL. Unlike PropBank, its text samples are
There is a number of works devoted to au- annotated only partially, so they are not suitable
tomatic semantic parsing of Russian texts. In for straightforward training of a supervised argu-
(Sokirko, 2001), a rule-based semantic parser is ment extractor or a combined pipeline. Therefore,
presented that converts a sentence into a seman- we split our pipeline into multiple stages, some of
tic graph. The work does not provide numeri- which leverage rule-based methods.
cal evaluation results, and the generated seman- The pipeline for semantic role labeling assumes
tic graph is substantially different from predicate- that input texts are preprocessed with a tokenizer, a
argument structures produced in SRL. In (Shel- sentence splitter, a POS-tagger, a lemmatizer, and
manov and Smirnov, 2014), authors present a a syntax parser that produces a dependency tree
rule-based semantic parser for Russian that re- in a Universal Dependencies format (Nivre et al.,
lies on a dictionary of predicates and a set of 2016). The SRL pipeline consists of the follow-
morpho-syntactic rules created by human experts. ing steps: predicate identification, argument ex-
They use this parser to automatically annotate traction, argument classification, and global scor-
representative corpus for supervised training of ing.
a transition-based labeler. In (Kuznetsov, 2015; In the predicate identification step, we mark all
Kuznetsov, 2016), an SVM-based labeling model verbs and some verb forms according to the given
is trained on FrameBank corpus. Authors rely on POS-tags of sentence tokens. We do not con-
feature-engineering approach and suggest to use sider verbal nouns as predicates since they are not
syntactic features and clusters of lexis. They also present in the FrameBank corpus. In the argument
implement integer linear programming inference extraction step, for each marked predicate, we try
as a post processing step. These works are based to detect its arguments within a sentence by ana-
on the pre-release version of the FrameBank cor- lyzing its syntax dependency tree with a number of
pus and do not provide code for data preparation, manually constructed rules. The arguments in the
modeling, and evaluation. They also do not con- pipeline are not spans as stated in CoNLL Shared
sider argument extraction and the problem with Tasks 2004, 2005, 2012 (Carreras and Màrquez,
labeling arguments of “unknown” predicates. In 2004; Carreras and Màrquez, 2005; Pradhan et al.,
(Shelmanov and Devyatkin, 2017), authors exper- 2012), but single tokens (nouns, proper names, or
iment with training a neural network models on the pronouns) as stated in CoNLL Shared Tasks 2008,
FrameBank corpus and suggest using word2vec 2009 (Surdeanu et al., 2008; Hajič et al., 2009).
embeddings for dealing with scarcity of predicate For the argument classification step, we train
annotations. However, they implement only an two neural models that predict roles of arguments
argument classification step but not the full SRL 2
https://github.com/lizaku/
pipeline. It is also worth noting that they per- frame-parsing
621
of “known” and “unknown” predicates accord- “known” predicates includes embeddings of argu-
ingly. We call a predicate “known” if it appears in ment and predicate lemmas, as well as the follow-
a training set within a labeled example, and “un- ing sparse lexical and morpho-syntactic features:
known” if it does not. During the inference we
choose a model by simply checking a presence of • Various types of morphological characteris-
a given predicate in a list of predicates appeared in tics of both an argument and a predicate
the training corpus. The result of model inference (case, valency, verb form, etc.)
is a set of probabilities for each semantic role in • Relative position of an argument in a sen-
the inventory (above a certain threshold). tence with respect to a predicate.
In the global scoring step, we enforce the fi-
nal predicate-argument structure to fulfill the im- • Preposition of an argument extracted from a
portant linguistic constraint: in a single predicate- syntax tree (including a complex preposition
argument structure, each semantic role can be as- as a single string).
signed only once, and each argument can have
• Name of a syntactic link that connects an ar-
only one role.
gument token to its parent in the syntax tree.
The whole pipeline is schematically depicted in
Figure 1. • Argument and predicate lemmas.
3.1 Argument Extraction We note that the predicate lemma is one of the
Base arguments are extracted from a syntax tree essential features for high-quality semantic role la-
of a sentence by rules that take into account POS- beling, since predicates express a situation in a
tags of tokens and direct syntax dependency links sentence and determine its meaning slots. Similar
rooted in predicates. Arguments are often also morpho-syntactic structures with different predi-
connected to predicates not directly but through cates can express different meanings. Therefore,
simple and complex prepositions that consist of the lack of annotation for a predicate in a train-
several words. Complex prepositions are de- ing set hits hard classifier confidence and over-
tected using the predefined list of triplets <PREP, all performance on examples with this “unknown”
NOUN, syntactic link>. Name of a syntactic link predicate. We combat this problem by introducing
is used to resolve ambiguity between noun phrases a second model that is trained without predicate
and complex prepositions. lemmas as features, so it should rely on predicate
We also take into account tokens that are not lemma word embeddings and other features in-
related to the predicate directly but are homoge- stead. This model performs worse than the model
neous to base arguments. The tokens that are for “known” predicates on seen predicates but it
linked to the base arguments with a conjunct rela- is also affected less by an absence of a predicate
tion (“conj”) are considered as extensions of base in a training corpus. This happens because predi-
arguments and are labeled with the same semantic cate lemma embeddings capture predicate seman-
role as the base argument. The list of predicates tics, and similarity between these embeddings can
is also expanded via adding the syntactic subjects help the model to guess meanings of “unknown”
and agents connected to the extracted arguments predicates.
with a nominal subject (“nsubj”) relation, as well In this work, we experiment with various types
as with “name” and “appos” in case it is a person of word embeddings obtained from shallow mod-
name or a title. The nominal modifier (“nmod”) els word2vec and FastText, as well as from pre-
relation is used for nominal dependents and often trained language models ELMo and BERT. Re-
indicates locations. The adverbial clause modifier cently, it has been shown that pretrained lan-
(“advcl”) helps to find the sequences of participle guage models provide substantially better gener-
clauses that have a common subject. alization to downstream models compared to shal-
low embeddings built by word2vec or FastText al-
3.2 Argument Classification gorithms. This happens because language mod-
For argument classification, we train two feed- els can take into account contexts of words, for
forward neural-network models: the model for which they generate an embedding, and capture
“known” predicates and the model for “unknown” much longer dependencies in texts. ELMo gen-
predicates. The feature set of the model for erates contextual word representations by using a
622
Figure 1: Semantic role labeling pipeline
stack of bidirectional LSTM layers that are trained duplicate argument labels, since a meaning slot of
to predict a following word on the basis of seen a situation can be filled by just a single participant
context. The output from each layer can be used as (for the core semantic roles) or a group of homo-
a set of features to downstream models. BERT is geneous arguments. In the argument classification
a masked language model that is trained to predict step, we use a neural network model to produce
masked words in a sequence given all other words a number of probabilities for a list of semantic
in the sequence. In addition, it is also trained roles and due to the linguistic constraint, we can-
to simultaneously predict whether two given sen- not greedily assign roles with maximum probabil-
tences are consecutive. BERT uses self-attention ity. In the global scoring step, we effectively pro-
encoder (Vaswani et al., 2017) that can be much duce the global optimal predicate-argument struc-
faster than LSTM and can capture longer depen- ture that fulfills the constraint by leveraging an in-
dencies across sequences. teger linear programming inference (Punyakanok
The feedforward neural network model for et al., 2004). Formally, it can be described in the
argument classification has three dense layers. following way. Let xij ∈ {0, 1} be a target vari-
Three inputs, namely embedding of a predicate, able and xij = 1 means an argument j has a se-
embedding of an argument, and sparse categori- mantic role i. Let p(i, j) be a probability of assign-
cal features are separately passed through the first ing a role i to an argument j estimated by a neural
piecewise layer with ReLU activation. Concate- network. Let n and m be the numbers of roles and
nated outputs of the first layer are then propa- arguments accordingly. The optimization problem
gated through another ReLU layer and the output formally:
layer with softmax activation. Before the activa-
m X
n
tion function, batch normalization is applied on X
argmax xij log p(i, j)
each hidden layer. The network is regularized with xij
j i
dropout. The output of this network is a vector of
probabilities for each semantic role in a given in- n
X
ventory. xij = 1, j = 1..m
i
3.3 Global Scoring
m
X
Semantic labels in a single predicate-argument xij = 1, i = 1..n
structure are not completely independent from j
each other. Moreover there is a certain linguis-
tic constraint that requires that there should be no xij ∈ {0, 1}, i = 1..n, j = 1..m
623
have at least 10 examples. The filtered dataset re-
Table 1: The merging scheme for mixed roles
Original (mixed) role Destination role tains 643 unique predicates (verbs). We also drop
agent – perceiver perceiver infrequent roles, for which the dataset contains
agent – sbj of mental state sbj of mental state less then 180 examples. The final corpus version
result / target result
location – patient location contains 52,751 examples for 44 unique semantic
speaker – sbj of psychol. sbj of psychol. state roles.
state
4.2 Embeddings and Pretrained Language
Models
The optimal solution to this problem is used
as the final assignment of semantic roles to argu- In our experiments, we use the following pub-
ments. licly available pretrained word embeddings and
language models:
4 Experiments
• Word2Vec: RusVectores5 (Kutuzov and Kuz-
4.1 Dataset and Preprocessing menko, 2017). Skip-gram model trained on
FrameBank contains annotated text samples with Russian Wikipedia, dimension: 300.
multiple contextual sentences. Each sentence con- • FastText: DeepPavlov6 (Burtsev et al., 2018).
sists of tokens with their morphological features. Skip-gram model trained on a mixed cor-
We follow preprocessing procedure from (Shel- pus of Russian Wikipedia and Lenta.ru news
manov and Devyatkin, 2017) to map annotations texts, dimension: 300.
to corresponding tokens. We also merged mixed
roles annotations from the original dataset into • ELMo: DeepPavlov. Pretrained on Rus-
their subsequent roles. See the merging scheme sian Wikipedia corpus, achieves perplexity of
in Table 1. 43.692, 2 layers, dimension: 1024.
Unlike (Shelmanov and Devyatkin, 2017), in
this work, we do not rely on gold-standard linguis- • BERT multilingual cased: released by the au-
tic annotations at all, since the goal of our work is thors (Devlin et al., 2018). Pretrained on 104
to develop the parser that can process raw texts. To languages, 12 encoder blocks, produces vec-
generate linguistic annotations for SRL input, we tors with 768 dimensions.
perform the following linguistic processing steps: • RuBERT: DeepPavlov. RuBert is an adap-
tation of BERT-multilingual with vocabulary
• Tokenization and sentence splitting are per-
enriched with Russian byte-pairs (Arkhipov
formed by NLTK library3 .
et al., 2019).
• Lemmatization, POS-tagging, and mor-
Both ELMo and FastText mitigate out-of-
phological analysis are done by MyStem
vocabulary problem, so we do not lose any predi-
(Segalovich, 2003).
cates and arguments, while there are some misses
• Syntax parsing is performed via UDPipe for word2vec. BERT models are built upon byte-
parser (Straka and Straková, 2017) with pair encoding, so we use only the first byte-pair
model trained on SynTagRus (Nivre et al., representation for each token as recommended by
2008). the authors. It is worth noting that although BERT
is a quite large model, it takes only 15 minutes
These steps are implemented using the IsaNLP on a single GTX 1080 Ti to encode all examples,
library4 . compared to 1.5 hours for ELMo.
The original corpus after preprocessing con-
tains examples for 803 predicates. However, for 4.3 Neural Network Hyperparameters
many predicates there are just few examples, and We performed hyperparameter tuning using ran-
some semantic roles are also rare. Therefore, we dom search on the task of argument labeling for
followed (Shelmanov and Devyatkin, 2017) and “known” predicates. The following parameters
filtered the dataset keeping only predicates that were selected and used in all further experiments:
3 5
https://www.nltk.org/ http://rusvectores.org/en/
4 6
https://github.com/IINemo/isanlp https://deeppavlov.ai/
624
categorical features layer hidden size – 400, em- 4.5 Results and Discussion
beddings projection layer hidden size – 100, con-
catenated vector layer size – 400, dropout – 0.3. In the argument extraction step, we achieve
74.48% precision, 85.12% recall, and 79.44% F1-
score. We note that many false positives are due
4.4 Experimental Setup
to the absence of non-core arguments in our eval-
uation set. These phrases that bear temporal, loca-
We evaluated the argument extraction step using
tive, and some other types of information are cor-
only the predicate-argument structures labeled in
rectly identified by our parser but are considered as
FrameBank and did not take into consideration
mistakes in the evaluation setup resulting in lower
any other structures in the corpus found by our
precision than it actually is. However, we see that
parser.
it is the only adequate way to assess extraction
For evaluation of argument labeling step, we quality with partially labeled data.
conducted two experiments using various token
representations and dataset splitting schemes. The results for the argument labeling step pre-
sented in Table 2 show that ELMo and BERT out-
In the first setup, we evaluate argument classi- perform all other approaches, including results in
fication step on full set of features and test vari- (Shelmanov and Devyatkin, 2017), although, un-
ous word representations. Lexical, morphological, like them, we do not rely on gold-standard mor-
and syntax features are encoded in one-hot man- phological features. In Table 4, we also report the
ner. Macro and micro F1 scores are calculated on a performance per semantic role of the model that
5-fold cross-validation. Evaluation results are pre- uses ELMo word representations.
sented in Table 2.
In many works, BERT outperforms ELMo by a
In the second setup, we evaluate the perfor-
significant margin. However, in our work, there is
mance of the argument classification models for
just an insignificant gap between ELMo and Ru-
“unknown” predicates. Thus, we divided the
BERT. This is probably due to BPE tokenization
dataset in two parts by leaving 80% of predicates
scheme of BERT, since we take encoded represen-
with their examples for training and 20% of pred-
tation only for the first subword unit of each token,
icates for testing. The sets of predicates in train-
with no fine-tuning, leaving a lot of information
ing and testing parts do not intersect. The division
about words unused.
of predicates was performed at random. For this
setup, there was no cross-validation. Instead, we In the second experimental setup, the gap be-
averaged results of 5 models trained with differ- tween RuBERT and ELMo is increased. In this
ent random seeds. In this experimental setup, we case, the model based on RuBERT shows worse
compare models for “unknown” predicates that do performance than all other approaches. However,
not take into account predicate lemma with vari- there is a certain improvement ∆1.5% of micro F1
ous types of token embeddings. score between ELMo and word2Vec-based mod-
els. It shows that representations generated by
To ensure the importance of introducing ad-
deep pretrained language models can restore se-
ditional model for “unknown” predicates in our
mantics of unseen predicates better than shallow
semantic role labeling pipeline and importance
models by additionally leveraging the context.
of predicate lemma feature, we also evaluate the
model for “known” predicates on the test set with The results of the model for “known” predicates
“unknown” predicates. The goal of this experi- even with the best word representations ELMo are
ments is to show that the model for “known” pred- expectedly low. The performance drop compared
icates overfits on predicate-lemma features and to the model for “unknown” predicates with the
performs worse than models trained specifically same embeddings is substantial: 10% micro-F1
for “unknown” predicates, since it is “blind” to and more than 8% macro-F1. It is worse than
its most important feature. We report the results any of other models except RuBERT. This proves
of the model for “known” predicates in this setup that predicate lemma is very important as a fea-
only with the embeddings that achieve the best ture and SRL pipeline should have two models to
score in the previous experiment. The test results process “known” and “unknown” predicates to al-
are presented in Table 3. leviate the domain shift.
625
Model Micro F1 Macro F1 Class Precision Recall F-score
agent (11.7%) 76.1 83.3 79.5
Plain Features Only 76.96 ± 0.67 73.63 ± 0.61 patient (10.2%) 85.1 88.7 86.9
Word2Vec UPOS 79.87 ± 0.34 76.70 ± 0.77 theme (6.9%) 84.6 71.6 77.6
sbj of psychol. state (6.2%) 86.7 83.9 85.2
FastText 80.60 ± 0.51 77.39 ± 0.36 goer (5.7%) 82.9 89.2 85.9
ELMo 83.42 ± 0.60 79.91 ± 0.40 cause (4.7%) 86.2 88.6 87.4
speaker (4.5%) 73.5 78.3 75.8
BERT-Multiling 79.04 ± 0.63 75.68 ± 0.72 location (4.1%) 87.4 82.5 84.9
RuBERT 83.12 ± 0.60 80.12 ± 0.62 content of action (3.6%) 89.1 83.8 86.3
content of thought (3.4%) 74.6 79.7 77.0
content of speech (3.4%) 75.9 69.5 72.6
Table 2: Performance of models in the experimen- final destination (3.4%) 70.3 52.0 59.8
result (2.8%) 63.5 54.0 58.4
tal setup with “known” predicates patient of motion (2.6%) 88.8 80.4 84.4
stimulus (2.4%) 85.1 72.2 78.1
cognizer (2.3%) 85.1 76.9 80.8
Model Micro F1 Macro F1 addressee (1.8%)
perceiver (1.7%)
75.7
90.5
79.1
79.0
77.4
84.3
ELMo (for known pred.) 45.51 ± 0.50 29.31 ± 0.82 counteragent (1.6%) 56.8 65.6 60.9
Word2Vec UPOS 53.97 ± 0.21 37.29 ± 0.74 effector (1.4%) 77.0 81.0 78.9
subject of social attitude (1.1%) 82.2 79.5 80.8
ELMo 55.50 ± 0.51 37.64 ± 0.41 initial point (1.1%) 76.0 80.4 78.1
FastText 49.37 ± 0.43 37.26 ± 0.29 topic of speech (1.0%) 58.3 81.5 68.0
manner (1.0%) 84.0 69.3 76.0
BERT-Multiling 31.81 ± 0.51 21.04 ± 0.13 recipient (1.0%) 82.3 68.0 74.5
RuBERT 43.68 ± 0.50 30.84 ± 0.55 goal (0.9%) 80.0 67.7 73.3
field (0.7%) 90.7 91.8 91.3
attribute (0.7%) 83.5 81.5 82.5
Table 3: Performance of models in the experimen- source of sound (0.7%) 73.7 69.5 71.6
behaver (0.6%) 84.8 84.4 84.6
tal setup with “unknown” predicates situation in focus (0.6%) 88.2 88.3 88.2
counteragent of social attitude (0.6%) 75.0 58.2 65.5
sbj of physiol. reaction (0.6%) 76.0 85.4 80.4
topic of thought (0.6%) 95.9 88.7 92.2
potential patient (0.5%) 89.3 90.9 90.1
5 Conclusion and Future Work status (0.5%)
patient of social attitude (0.5%)
89.0
86.1
78.4
76.2
83.3
80.8
standard (0.5%) 80.2 85.3 82.7
term (0.5%) 87.5 85.7 86.6
We presented and evaluated the first full pipeline attribute of action (0.5%) 92.5 71.2 80.4
for semantic role labeling of Russian texts. causer (0.4%)
initial possessor (0.4%)
72.6
83.7
65.2
73.5
68.7
78.3
The experiments with various types of embed- potential threat (0.4%)
path (0.3%)
73.6
90.3
82.7
80.0
77.9
84.9
dings showed that the pretrained language mod-
els ELMo and BERT substantially outperform the Table 4: The performance of the model based on
embeddings obtained with shallow algorithms like ELMo embeddings in the experimental setup with
word2vec and FastText. We also showed that pro- “known” predicates by semantic roles. The fre-
viding supplementary SRL model for “unknown” quencies of roles in the training corpus are pre-
predicates can alleviate the problem with annota- sented in parentheses
tion scarcity. We note, that in the case of pre-
dicting arguments for “uknown” predicates, the References
deep pretrained language model ELMo also out- Mikhail Arkhipov, Maria Trofimova, Yuri Kuratov, and
performed other types of embeddings. We publish Alexey Sorokin. 2019. Tuning multilingual trans-
the code of the pipeline that can be used to parse formers for language-specific named entity recogni-
raw Russian texts7 , we also publish the code for tion. In Proceedings of the 7th Workshop on Balto-
Slavic Natural Language Processing, pages 89–93.
model training and experimental evaluation.
In the future work, we are going to apply semi- Collin F Baker, Charles J Fillmore, and John B Lowe.
supervised and unsupervised algorithms to expand 1998. The berkeley framenet project. In Proceed-
ings of the 17th international conference on Compu-
the training data and improve the model perfor- tational linguistics-Volume 1, pages 86–90.
mance on out-of-domain data.
Emanuele Bastianelli, Giuseppe Castellucci, Danilo
Croce, and Roberto Basili. 2013. Textual inference
and meaning representation in human robot interac-
tion. In Proceedings of the Joint Symposium on Se-
mantic Processing. Textual Inference and Structures
Acknowledgments in Corpora, pages 65–69.
This paper is partially supported by Russian Foun- Piotr Bojanowski, Edouard Grave, Armand Joulin, and
dation for Basic Research (project No. 16-29- Tomas Mikolov. 2017. Enriching word vectors with
subword information. Transactions of the Associa-
12937, 17-07-01477). tion for Computational Linguistics, 5:135–146.
Mikhail Burtsev, Alexander Seliverstov, Rafael
7
https://github.com/IINemo/isanlp_srl_ Airapetyan, Mikhail Arkhipov, Dilyara Baymurz-
framebank/tree/master ina, Nickolay Bushkov, Olga Gureenkova, Taras
626
Khakhulin, Yuri Kuratov, Denis Kuznetsov, et al. Daniel Jurafsky and James H. Martin. 2009. Speech
2018. Deeppavlov: Open-source library for dia- and Language Processing (2Nd Edition). Prentice-
logue systems. In Proceedings of ACL 2018, System Hall, Inc.
Demonstrations, pages 122–127.
Paul Kingsbury and Martha Palmer. 2002. From tree-
Xavier Carreras and Lluı́s Màrquez. 2004. Intro- bank to propbank. In LREC, pages 1989–1993.
duction to the conll-2004 shared task: Semantic
role labeling. In Proceedings of the Eighth Confer- Andrey Kutuzov and Elizaveta Kuzmenko, 2017. We-
ence on Computational Natural Language Learning bVectors: A Toolkit for Building Web Interfaces for
(CoNLL-2004) at HLT-NAACL 2004. Vector Semantic Models, pages 155–161. Springer.
Xavier Carreras and Lluı́s Màrquez. 2005. Intro- Ilya Kuznetsov. 2015. Semantic role labeling for rus-
duction to the CoNLL-2005 shared task: Semantic sian language based on russian framebank. In Inter-
role labeling. In Proceedings of the Ninth Confer- national Conference on Analysis of Images, Social
ence on Computational Natural Language Learning Networks and Texts, pages 333–338. Springer.
(CoNLL-2005), pages 152–164.
Ilya Kuznetsov. 2016. Automatic semantic role la-
Ronan Collobert, Jason Weston, Léon Bottou, Michael belling in Russian language, PhD thesis (in Rus-
Karlen, Koray Kavukcuoglu, and Pavel Kuksa. sian). Ph.D. thesis, Higher School of Economics.
2011. Natural language processing (almost) from
scratch. Journal of machine learning research, Olga Lyashevskaya and Egor Kashkin. 2015. Frame-
12(Aug):2493–2537. bank: a database of russian lexical constructions. In
International Conference on Analysis of Images, So-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and cial Networks and Texts, pages 350–360.
Kristina Toutanova. 2018. BERT: Pre-training of
deep bidirectional transformers for language under- Olga Lyashevskaya. 2012. Dictionary of valencies
standing. arXiv preprint arXiv:1810.04805. meets corpus annotation: a case of russian frame-
bank. In Proceedings of the 15th EURALEX Inter-
Nicholas FitzGerald, Julian Michael, Luheng He, and national Congress, volume 15.
Luke Zettlemoyer. 2018. Large-scale qa-srl pars-
ing. In Proceedings of the 56th Annual Meeting Ana Marasović and Anette Frank. 2018. SRL4ORL:
of the Association for Computational Linguistics, Improving opinion role labeling using multi-task
pages 2051–2060. learning with semantic role labeling. In Proceed-
ings of the 2018 Conference of the North Ameri-
Daniel Gildea and Daniel Jurafsky. 2002. Automatic can Chapter of the Association for Computational
labeling of semantic roles. Computational linguis- Linguistics: Human Language Technologies, pages
tics, 28(3):245–288. 583–594.
Jan Hajič, Massimiliano Ciaramita, Richard Johans- Diego Marcheggiani, Anton Frolov, and Ivan Titov.
son, Daisuke Kawahara, Maria Antònia Martı́, Lluı́s 2017. A simple and accurate syntax-agnostic neural
Màrquez, Adam Meyers, Joakim Nivre, Sebastian model for dependency-based semantic role labeling.
Padó, Jan Štěpánek, et al. 2009. The conll-2009 In Proceedings of the 21st Conference on Computa-
shared task: Syntactic and semantic dependencies tional Natural Language Learning (CoNLL 2017),
in multiple languages. In Proceedings of the Thir- pages 411–420.
teenth Conference on Computational Natural Lan-
guage Learning: Shared Task, pages 1–18. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor-
rado, and Jeff Dean. 2013. Distributed representa-
Luheng He, Mike Lewis, and Luke Zettlemoyer. 2015. tions of words and phrases and their compositional-
Question-answer driven semantic role labeling: Us- ity. In Advances in neural information processing
ing natural language to annotate natural language. systems, pages 3111–3119.
In Proceedings of the 2015 conference on empirical
methods in natural language processing, pages 643– Joakim Nivre, Igor M Boguslavsky, and Leonid L
653. Iomdin. 2008. Parsing the syntagrus treebank of
russian. In Proceedings of the 22nd International
Luheng He, Kenton Lee, Mike Lewis, and Luke Zettle- Conference on Computational Linguistics, pages
moyer. 2017. Deep semantic role labeling: What 641–648.
works and whats next. In Proceedings of the 55th
Annual Meeting of the Association for Computa- Joakim Nivre, Marie-Catherine De Marneffe, Filip
tional Linguistics, pages 473–483. Ginter, Yoav Goldberg, Jan Hajic, Christopher D
Manning, Ryan McDonald, Slav Petrov, Sampo
Shexia He, Zuchao Li, Hai Zhao, and Hongxiao Bai. Pyysalo, Natalia Silveira, et al. 2016. Universal de-
2018. Syntax for semantic role labeling, to be, or pendencies v1: A multilingual treebank collection.
not to be. In Proceedings of the 56th Annual Meet- In Proceedings of the Tenth International Confer-
ing of the Association for Computational Linguis- ence on Language Resources and Evaluation (LREC
tics, pages 2061–2071. 2016), pages 1659–1666.
627
GS Osipov, IV Smirnov, and IA Tikhomirov. 2010. Chen Shi, Shujie Liu, Shuo Ren, Shi Feng, Mu Li,
Relational-situational method for text search and Ming Zhou, Xu Sun, and Houfeng Wang. 2016.
analysis and its applications. Scientific and Tech- Knowledge-based semantic embedding for machine
nical Information Processing, 37(6):432–437. translation. In Proceedings of the 54th Annual Meet-
ing of the Association for Computational Linguis-
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt tics, volume 1, pages 2245–2254.
Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. 2018. Deep contextualized word rep- Alexey Sokirko. 2001. A short description of dialing
resentations. In Proceedings of the 2018 Conference project.
of the North American Chapter of the Association
for Computational Linguistics: Human Language Milan Straka and Jana Straková. 2017. Tokenizing,
Technologies, pages 2227–2237. pos tagging, lemmatizing and parsing ud 2.0 with
udpipe. In Proceedings of the CoNLL 2017 Shared
Sameer Pradhan, Kadri Hacioglu, Wayne Ward, Task: Multilingual Parsing from Raw Text to Univer-
James H Martin, and Daniel Jurafsky. 2005. Seman- sal Dependencies, pages 88–99.
tic role chunking combining complementary syn-
tactic views. In Proceedings of the Ninth Confer- Emma Strubell, Patrick Verga, Daniel Andor,
ence on Computational Natural Language Learning, David Weiss, and Andrew McCallum. 2018.
pages 217–220. Linguistically-informed self-attention for semantic
role labeling. In Proceedings of the 2018 Confer-
Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, ence on Empirical Methods in Natural Language
Olga Uryupina, and Yuchen Zhang. 2012. Conll- Processing, pages 5027–5038.
2012 shared task: Modeling multilingual unre-
stricted coreference in ontonotes. In Joint Confer- Mihai Surdeanu, Richard Johansson, Adam Meyers,
ence on EMNLP and CoNLL-Shared Task, pages 1– Lluı́s Màrquez, and Joakim Nivre. 2008. The
40. conll-2008 shared task on joint parsing of syntactic
and semantic dependencies. In Proceedings of the
Vasin Punyakanok, Dan Roth, Wen-tau Yih, and Dav Twelfth Conference on Computational Natural Lan-
Zimak. 2004. Semantic role labeling via integer lin- guage Learning, pages 159–177.
ear programming inference. In Proceedings of Col-
Zhixing Tan, Mingxuan Wang, Jun Xie, Yidong Chen,
ing 2004, pages 1346–1352.
and Xiaodong Shi. 2018. Deep semantic role la-
Vasin Punyakanok, Dan Roth, and Wen-tau Yih. 2005. beling with self-attention. In Thirty-Second AAAI
The necessity of syntactic parsing for semantic role Conference on Artificial Intelligence.
labeling. In IJCAI, volume 5, pages 1117–1123.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Gozde Gul Sahin and Mark Steedman. 2018. Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Character-level models versus morphology in se- Kaiser, and Illia Polosukhin. 2017. Attention is all
mantic role labeling. In Proceedings of the 56th An- you need. In Advances in neural information pro-
nual Meeting of the Association for Computational cessing systems, pages 5998–6008.
Linguistics, pages 386–396. Nianwen Xue and Martha Palmer. 2004. Calibrating
features for semantic role labeling. In Proceedings
Ilya Segalovich. 2003. A fast morphological algorithm
of the 2004 Conference on Empirical Methods in
with unknown word guessing induced by a dictio-
Natural Language Processing.
nary for a web search engine. In MLMTA, pages
273–280. Citeseer.
628
A Structural Approach to Enhancing WordNet with Conceptual Frame
Semantics
629
Proceedings of Recent Advances in Natural Language Processing, pages 629–637,
Varna, Bulgaria, Sep 2–4 2019.
tual structures describing particular types of ob- The first one is Using (Is Used by ↔ Uses), a
jects, situations, or events along with their com- frame-to-frame relation defined as a relationship
ponents, called frame elements, or FEs (Baker between two frames where the first one makes ref-
et al., 1998; Baker and Ruppenhofer, 2002; Rup- erence in a very general kind of way to the struc-
penhofer et al., 2016). Depending on their sta- ture of a more abstract, schematic frame (Rup-
tus, FEs may be core, peripheral or extra-thematic, penhofer et al., 2016, p. 83); it may be viewed
cf. Ruppenhofer et al. (2016). For our purposes, as a kind of weak inheritance (Petruck, 2015)
we deal particularly with core FEs, which instanti- where only some of the FEs in the parent frame
ate conceptually necessary components of a frame, have a corresponding entity in the child frame,
and which in their particular configuration make and if such exist, they are more specific (Petruck
a frame unique and different from other frames. and de Melo, 2012). Thus, the frame Arrang-
To a lesser degree we touch upon peripheral FEs, ing Uses the frame Placing and the two frames
which mark notions such as Time, Place, Man- share the FEs Agent and Theme, while the more
ner, Means, among others and may be instantiated specific FE Configuration in the frame Arranging
in any semantically appropriate frame. A Lexi- corresponds to the more general FE Goal in the
cal Unit (LU) in FN is a pairing of a word with a frame Placing; the first frame is exemplified by the
meaning; each LU is associated with a frame de- synset {arrange:1, set up:5} which is a hyponym
scribing its conceptual semantics. of {put:1, set:1, place:1, pose:5} whose assigned
frame is the more general frame Placing.
2.2 Structural Properties of WN and FN We consider two more relations although
they align with the notion of inheritance only
FN frames are related into a netlike structure
marginally. Perspective (Is Perspectivized in ↔
through a number of frame-to-frame relations part
Perspective on) is defined as a relation which indi-
of which also provide a hierarchical organisation.
cates that a given situation viewed as neutral may
These relations are presented in detail in Leseva
be further specified by means of perspectivised
and Stoyanova (2019); below, we just sum up
frames representing different possible points-of-
those used in the procedures we propose.
view on this neutral situation (Ruppenhofer et al.,
Inheritance (Is Inherited by ↔ Inherits from) 2016, p. 82). Subframe (Has Subframe(s) ↔ Sub-
is the strongest and most prominent relation in frame of) is a relation between a complex frame re-
FN, which posits a relationship between a more ferring to sequences of states and transitions (each
general (parent) frame and a more specific (child) of which can itself be separately described as a
frame so that the child frame elaborates on the par- frame) and the frames denoting these states or
ent frame in such a way that each semantic fact transitions (Ruppenhofer et al., 2016, p. 83–84).
about the parent must correspond to an equally
These hierarchical relations are the basis for
or more specific fact about the child (Ruppen-
part of the procedures outlined in Section 4.
hofer et al., 2016, p. 81–82). By definition, In-
heritance corresponds to the relation of hyper-
2.3 Causativity in WN and FN
nymy/hyponymy in WordNet. In an ideal set-
ting, the hyponyms should be instantiations of the Causativity (Is Caused by ↔ Causative of) is a
hypernym’s frame or of a more specific frame systematic non-inheritance relation where one of
that inherits from the hypernym’s frame. For in- the frames represents the causative counterpart of
stance, the frame Execution Inherits from Killing the other, stative or inchoative, frame (Ruppen-
and is assigned to the WN synset {execute:1, put hofer et al., 2016, p. 85). Causativity corresponds
to death:1} – a hyponym of {kill:1}, which is as- straightforwardly to the WN relation causes, al-
signed the frame Killing. though this correspondence is exhibited in only
As the two resources have been developed in- a small number of cases (30 pairs) – the rela-
dependently, their relational structure is distinct, tion has not been implemented consistently neither
one of the major differences being that there are in FN, nor in WN even in clear-cut parts of the
other frame-to-frame relations that to various de- lexicon such as those described by the hypernym
grees embody the notion and features of inheri- trees with the roots {change:1, alter:1, modify:3}
tance; we use (some of) them in the definition of (’cause to change; make different; cause a trans-
the procedures proposed in Section 4.1. formation’) and {change:2} (’undergo a change;
630
become different in essence; losing one’s or its erage which are aimed at: (i) discovering exist-
original nature’). ing but unmapped relations between synset mem-
bers and FN frames; and (ii) transferring frames
3 Existing Mappings of WN and FN between synsets through relations of inheritance
Previous efforts at linking WN and FN include Shi derived from WN and FN.
and Mihalcea (2005), Baker and Fellbaum (2009),
WordFrameNet1 (Laparra and Rigau, 2009, 2010), 4 Enhancing WN Mappings to FN
MapNet2 (Tonelli and Pighin, 2009), and more en- As noted above, the proposed approach combines
hanced proposals, such as the system Semlink3 the features used in the direct mapping with the
(Palmer, 2009) which brings together WN, FN and structural properties of WN and FN – particu-
VerbNet with PropBank, and its follow-up Sem- larly, the inheritance relations existing between
link+ that brings in mapping to Ontonotes (Palmer hypernyms and hyponyms in WN and the inher-
et al., 2014). Some procedures for automatically itance (and other similar relations) that determine
extending the mapping, are presented by Leseva the hierarchical structure of FN. As shown in Le-
et al. (2018) and a more thorough overview may seva and Stoyanova (2019), although the relations
be found in Leseva and Stoyanova (2019). in the two resources have different number and
Whereas these efforts have resulted in the cre- scope, part of them are grounded in similar uni-
ation of databases of integrated semantic knowl- versal assumptions which leads to partial overlap,
edge, most of them deal with mapping of the units depending on their definition and the specificities
of the original resources to each other – FN LUs of the information in the resources.
and WN synset members (literals), LU definitions Apart from the correspondences between FN’s
and synset glosses, etc. Such a methodology is Inheritance and other relations and the WN hyper-
able to perform mapping in those cases where nymy relation, there are other systematic structural
there is a correspondence between LUs and liter- relations which can be applied for the purpose of
als with equivalent or close meaning, but would enhancing the resources. Notable examples are the
fail where such correspondence is missing. With Causativity relation between frames in FN and the
155,287 synonyms in 117,659 synsets and more causes relation defined between causative and sta-
than 246,577 relations, of which 91,631 are in- tive or inchoative verbs in WN (cf. Section 5).
stances of the hypernymy relation4 as compared
with 13,640 LUs and 1,875 frame-to-frame rela- 4.1 Expanding Mappings based on
tions5 in FN, the discrepancy in the size of the data Hierarchical Relations in WordNet
is reflected in the limited coverage of the map-
Our work relies on the assumption that in a tax-
pings between synsets and frames. To the best of
onomic structure such as WN subordinate nodes
our knowledge, no further checks and verification
inherit the properties of their superordinates, i.e. a
have been performed on the mappings, as well.
hyponym elaborates on the meaning of its hyper-
The approach that we propose in addition to
nym and shares its conceptual and linguistic prop-
the lexical mapping of units deals with explor-
erties. We propose that if a WN synset instantiates
ing and taking into account the relational structure
a particular FN frame, its hyponyms should (ide-
of the resources (especially the structure of WN),
ally) instantiate the same or a more specific frame
particularly the relation of inheritance which en-
which may or may not hold a(n) (inheritance) re-
sures the propagation of conceptual and linguistic
lation with the more general frame.
features down the trees. We employ features of
This assumption allows us to suggest that in the
the relational structure in the definition of proce-
cases where we are not able to assign a FN frame
dures for the augmentation of the mapping cov-
due to the fact that the coverage of the two re-
1
http://adimen.si.ehu.es/web/ sources is non-overlapping and/or other mapping
WordFrameNet
2 procedures fail, we may resort to assigning the
https://hlt-nlp.fbk.eu/technologies/
mapnet frame of a hypernym to its hyponyms; at worst,
3
https://verbs.colorado.edu/semlink/ the semantic representation will be too general.
4
https://wordnet.princeton.edu/ There are 14,103 verb synsets in WN, which,
documentation/wnstats7wn
5
https://framenet.icsi.berkeley.edu/ unlike nouns that all have a common root, are or-
fndrupal/current_status ganised in 566 separate trees. Initially, FN frames
631
have been assigned to a total of 4,306 synsets out relations – especially the hypernymy relation, as
of which 264 are root synsets. In order to improve well as the relations between synsets with a com-
the quality of the existing mappings and to expand mon hypernym (i.e., sister synsets); and (ii) LUs
the coverage we performed the following steps: from a particular FN frame (the verbs listed as
(i) manual verification of the frames assigned to instantiations of a given frame), the hierarchical
root synsets (resulting in 75 corrected mappings); frame-to-frame relations: Inheritance, Uses, Sub-
(ii) semi-automatic assignment of valid frames to frame, and Perspective, as well as the relation be-
selected root synsets with a large number of hy- tween two frames inheriting from the same frame
ponyms (27 roots); (iii) assignment of a hyper- (i.e., sister frames).
nym’s frame to its hyponyms in the cases where For a synset assigned a frame inherited from its
a hyponym is not directly mapped to FN frames, hypernym, we apply the following procedures:
thus obtaining an extended coverage of 13,226
(1) Literal–LU correspondence using FN re-
verb synsets with an assigned FN frame; (iv) defi-
lations: We check whether any of the synset liter-
nition of further procedures to the end of improv-
als appears as a LU in: (a) the assigned frame (to
ing the quality of this assignment (section 4.2).
confirm its validity); (b) more specific frames the
frame under discussion is linked to by means of
4.2 Selection of Frames based on the FN and
any of the considered frame-to-frame relations (to
the WN Structure
make it more precise); (c) the sister frames of the
We devised two types of procedures aimed at ob- assigned frame.
taining a more specific mapping: (i) procedures Example 2. Synset: eng-30-00540946-v
that make use of the conceptual and lexical infor- {extend:8; expand:4} ’expand the influence of’
mation and the relational structure in FN; (ii) pro- Assigned FR from hypernym: Cause change
cedures employing the conceptual and lexical in- Suggested FR from (1b): Change event duration
formation and the relational structure in WN. (LU: extend)
As noted above, the first step of assigning a FN Suggested FR from (1c): Cause expansion (LU:
frame to a WN synset is transferring the frame expand)
assigned to the synset’s direct or inherited hyper-
(2) Literal–LU correspondence using WN re-
nym. The frame so assigned may either appropri-
lations: We check whether any of the synset lit-
ately describe the conceptual structure of the liter-
erals appears as a LU in: (a) any of the frames
als in the synset, or it may provide a more general
assigned to its hyponyms; (b) any of the frames re-
description than an optimally informative one. We
lated to the frames in (a) through frame-to-frame
therefore view this as a default assignment on the
relations; and (c) any of the frames assigned to its
basis of which we try to elaborate to the end of
sister synsets.
discovering a more suitable or specific frame to
which to map the synset. When such a frame is Example 3. Synset: eng-30-00223374-v
found, we validate it manually and assign it to the {bolster:1; bolster up:1} ’support and strengthen’
hyponyms of the synset under discussion, overrid- Assigned FR from hypernym: Cause change
ing the more general frame as in Example 1. Suggested FR from (2c): Supporting (LU:
bolster)
Example 1. Synset: eng-30-00047945-v
{dress:6; clothe:1; enclothe:1; garment:1} (3) General literal–LU correspondence: We
’provide with clothes or put clothes on’ check whether any of the synset literals appears as
Assigned FR from hypernym: Undergo change a LU in any other frame in FN.
Suggested FR: Dressing (transferred automati- Example 4. Synset: eng-30-00544936-v
cally to 13 out of 15 hyponyms such as {corset:1} {exalt:1} ’raise in rank, character, or status’
’dress with a corset’, {vest:1} ’dress with a vest’, Assigned FR from hypernym: Cause change
{overdress:2} ’dress too warmly’) Suggested FR from (3): Judgment (LU: exalt)
Below, we describe the procedures proposed (4) Keywords: We use keywords (words con-
and how they make use of the relational structure tained in the FN frame name, plus their derivatives
of FN and WN and the following components of collected from WN through the eng derivative re-
the description in the two resources, in particular: lation), to identify synset literals and/or definitions
(i) WN literals (and synsets) and synset-to-synset containing these keywords as candidates to be as-
632
signed the frame in question. Procedure # 1-step # 2-step # 3+ step
Example 5. Synset: eng-30-00448864-v transfers transfers transfers
{clean out:1; clear out:1} ’empty completely’ (1) 516 231 121
Assigned FR from hypernym: Cause change (2) 460 41 17
Suggested FR from (4): Emptying (keyword (3) 1,701 859 145
empty found in gloss) (4) 1,088 612 27
(5) 1,175 526 202
(5) Direct similarity: We check the similarity (6) 1,009 417 194
between the gloss of a verb synset and FN LU
Unique 3,957 1,388 316
definitions (even when there is no correspondence
synsets
between literals and LUs) to identify candidate
frames for a given verb synset. We separate: (i) Table 1: Distribution of frames suggested for
suggested frames related to the one assigned from synsets with automatic frame assignments from
the hypernym, which are given higher priority; (ii) the hypernym (rows (1)-(6) include multiple sug-
unrelated suggestions. gestions for the same synset)
Example 6. Synset: eng-30-02514187-v
{gloss over:1; skate over:1; smooth over:1; slur
over:1; skimp over:1} ’treat hurriedly or avoid more specific or other possible frames have been
dealing with properly’ made for 5,661 synsets with automatically trans-
Assigned FR from hypernym: Intentionally act ferred hypernym frames – Table 1 shows the dis-
Suggested FR from (5): Avoiding (which In- tribution of the new suggestions in terms of the
herits from Intentionally act); similarity with the types of procedures that have been applied and the
definition of LU skirt.v ’avoid dealing with’ distance of the synset from the hypernym whose
frame has been inherited.
(6) Indirect similarity: We check the similar-
ity between the glosses of synsets derivationally 4.3 Discussion on Evaluation
related to the verb under discussion (as well as
These suggestions need to be manually verified as
the glosses of their hypernyms, which are consid-
so far no reliable fully automated verification pro-
ered their closest semantic generalisation) and FN
cedure has been established. Since the main ob-
LU definitions to identify candidate frames for the
jective is to discover, or suggest, a more precise
verb synset. We separate: (i) suggested frames
frame than the one assigned from the hypernym,
related to the one assigned from the hypernym,
which is not necessarily wrong but rather may be
which are given higher priority; (ii) unrelated sug-
too general, such evaluation needs to measure the
gestions.
degree of relevance as opposed to precision. Fur-
Example 7. Synset: eng-30-00831651-v
thermore, it will be highly dependent on the gran-
{warn:1} ’notify of danger, potential harm, risk’
ularity of the frames and their hierarchical organi-
Assigned frame from hypernym: Telling
sation. Designing such a measure and its automa-
Derivationally related synset: eng-30-
tisation, if at all achievable, is beyond the scope of
07224151-n {warning:1} ’a message informing
this work.
of danger’
Suggestions, although non-definitive, provide
Suggested FR from (6): Warning (Inherits from
useful pointers to candidate frames and thus are
Telling); similarity with the gloss of LU alert.n ’a
valuable in assisting the manual selection of
message to inform someone of danger; a warning’
frames. Only in 203 cases are there multiple sug-
Similarity in procedures (5) and (6) is calcu- gestions as a result of the procedures, out of which
lated as a cumulative measure based on coinciding in 177 cases 5 or more different frames are sug-
terms in the two definitions. Scores of similarity gested. There are 1,056 synsets for which a sug-
between two words are highest for full match and gestion is confirmed at least 2 times from the re-
lower when stemming is applied. Short words (up peated application of the same or different proce-
to length 3) are disregarded and longer words are dures, out of which 265 cases are confirmed 5 or
given more weight. The final score is normalised more times.
by the length (in words) of the definitions. Given the task, human judgment is indispens-
Through these steps 9,341 new suggestions of able, especially for frames assigned to synsets
633
higher in the tree as errors propagate down and Causative Non-causative
change:1; change:2;
may result in multiple wrong assignments.
From FN 241 251
5 Causativity and Inchoativity as a Direct hypernym 910 624
Systemic Structural Feature Indirect hypernym 577 469
Total 1,728 1,344
Another direction of expanding the mappings and General frame* 719 561
verifying the information in both FN and WN is in % 41.6% 41.7%
by employing systematic semantic relations such
as causativity. It is a non-hierarchical relation Table 2: Analysed data with respect to causativ-
that links stative (e.g. {lie:2} ’be lying, be pros- ity (* assignments of the most general frame
trate; be in a horizontal position’) or inchoative Cause change for the causative and Undergo
({lie down:1, lie:7} ’assume a reclining position’) change for the non-causative)
verbs with their causative ({lay:2, put down:2, re-
pose:5} ’put in a horizontal position’) counter-
choative verb in a causative tree or vice versa, or in
parts. The relation provides enhancement comple-
the FN data – missing or wrong relation between
mentary to the one using hierarchical relations de-
frames, undefined frames, etc.
scribed above and links in a systematic way large
parts of the lexicon. Sections 5.1 and 5.2 describe the procedures for
exploring pairs of causative–non-causative trees
A considerable part of causative and non-
and extracting information enabling the valida-
causative pairs are formed with the same root and
tion of assigned frames, as well as the increase of
are thus morphologically similar or identical, e.g.
the density of causativity relations within FN and
EN change – change; RO schimba – schimba; BG
WN. Section 5.3 deals with the formulation of new
promenyam – promenyam se, which makes them
causative or stative and/or inchoative frames.
easier to identify. Nevertheless, as noted above,
causativity is not consistently encoded in WN, and
5.1 Analysis and Consistency Checks
neither is it fully implemented in FN where we
have spotted a number of instances of inchoat- We have extracted two separate WN trees from
ive/stative or causative frames lacking a counter- two root synsets connected by the causes re-
part in the opposite domain. This means that the lation (see Table 2): (1) eng-30-00126264-
verbs instantiating them cannot be appropriately v {change:1; alter:1; modify:3}, assigned the
described in FN. Respectively, the mapping of lit- frame Cause change; and its corresponding
erals instantiating non-defined frames will result non-causative counterpart (2) eng-30-00109660-v
in failure of assignment or wrong assignment. {change:2}, assigned the frame Undergo change.
Causativity also has an important application The checks for consistency with regards to (i)-
in WN and FN data validation and expansion: (iv) above, included the following procedures:
exploring the assignment of frames from FN to (1) Identifying non-causative synsets in the
synsets enables us to check the consistency of as- causative tree and causative synsets in the non-
signments, by adopting the following logic: (i) causative tree. These mismatches are identified
in a tree whose root is a causative synset, all the by pattern matching of the gloss or by analysis
descendants must be assigned a causative frame; aimed at establishing whether the manually as-
(ii) in a tree with an inchoative/stative root all signed frame contradicts the position in the tree. 9
the descendants must be inchoative/stative; (iii) such cases have been found in the causative tree
the pairs of causative–non-causative synsets from (e.g., eng-30-00416880-v {even:6; even out:2}
corresponding trees should be connected to each ’become even or more even’). Pattern-matching
other through the WN causes relation in a consis- in the non-causative tree proved to be unreli-
tent way; (iv) the respective pair of causative–non- able. It identified 120 cases of ’make’ or ’cause’
causative frames assigned to such a pair of synsets in the gloss, but only a small number of them
should also be related via the Is Causative of rela- were causative synsets (e.g., eng-30-00330565-v
tion in FN. The opposite signals either wrong as- {break up:3; disperse:1; scatter:1} ’cause to sep-
signment of a frame or inconsistency either in the arate’). We propose moving each wrongly placed
WN data, that is – the encoding of a stative or in- synset (and the subtree rooting from it) to the rel-
634
evant tree and attaching it to its real hypernym. ing in 47 pairs of corresponding synsets – one in
Furthermore, there are synsets which com- each tree. A set of consistency checks showed that
bine the causative and the non-causative meaning there are no crossing relations (i.e., no instances
and thus, create inconsistency in the WN struc- where for a causative hypernym C1 and its hy-
ture. We identify such synsets by pattern match- ponym C2 , and a non-causative hypernym N1 and
ing of the gloss since they usually have glosses its hyponym N2 , C1 causes N2 and C2 causes N1 ).
such as ’make or become’, ’cause or become’, Further procedures were proposed to dis-
’cause or undergo’. There are 7 cases in the cover pairs of corresponding causative and non-
causative (e.g., eng-30-01253468-v {coarsen:2} causative synsets unrelated through the causative
’make or become coarse or coarser’) and 5 in relation. These are based on pattern matching of
the non-causative tree (e.g., eng-30-00280532- the definition and/or on measuring similarity, as
v {blacken:1; melanize:1} ’make or become well as on an analysis of the synsets position in the
black’). We propose that such synsets are split WN tree structure, the causative relations in which
into two and placed at the respective positions in their sisters, hypernym and hyponyms enter, and
the relevant trees. This is an optimal solution as the frames assigned to them. On the basis of these
these concepts are not necessarily expressed by the linguistic features we have identified 673 possible
same lexeme cross-linguistically, and such a split causative relations between pairs of synsets in the
improves the consistency of WN. two corresponding trees. After manual validation
(2) Identifying non-causative frames assigned they may be used to create a more dense structure
to synsets in the causative tree and causative of causative relations in WN, as well as to extend
frames assigned to synsets in the non-causative them to frame-to-frame relations in FN.
tree. A causative frame is identified based on:
keywords such as ’cause’ or ’make’ in its name 5.3 Suggesting New Frames
or its definition; Agent or Cause/Causer FEs in
its conceptual structure; its position as the first New frames are suggested where a suitable
member in a Is Causative of relation, etc. A non- causative or non-causative frame is not defined in
causative frame is identified based on: keywords FN to match its existing counterpart. The miss-
such as ’become’ or ’undergo’; lack of Agent or ing one is defined using the conceptual descrip-
Cause/Causer FEs in its structure; its position as tion of the available frame. Consider the synset
the second member of an Is Causative of relation. {age:3} ’make older’: we assign it the frame
Cause change and then try to acquire additional
We found 7 non-causative frames in the
classificatory information and, possibly, to find a
causative tree (e.g., eng-30-02190188-v
more specific frame by applying the remaining
{quieten:3; hush:5; quiet:9; quiesce:1; quiet
procedures. We confirm that the synset’s mean-
down:1; pipe down:1} ’become quiet or quieter’,
ing is causative through the keyword procedure
Frame: Becoming silent) and 61 causative frames
(cf. Section 4.2). Another mapping procedure
in the non-causative tree (e.g., eng-30-00339085-
suggests Aging as the corresponding frame. Ag-
v {crush:1} ’break into small pieces’, Frame:
ing is a non-causative frame denoting the mean-
Cause to fragment). These are clearly either er-
ing of an entity undergoing a particular kind of
rors in the frame assignment or wrongly encoded
change (see Example 8). Since Aging does not
synsets as discussed in (1).
have a causative counterpart in FN, we posit such
(3) Identifying synset pairs connected by the
a frame, Cause to age. The conceptual structure
causes relation in WN where the causative synset
of stative/inchoative and causative counterparts
is assigned a non-causative frame or vice versa.
is distinguished by the presence of a causative
Section 5.2 deals with the enrichment of the two
subevent in the latter (Van Valin Jr. and LaPolla,
resources with instances of the causative relation.
1997, p. 109) which is associated with a causative
(Agent or an Agent-like) participant (FE). Thus,
5.2 Densifying Causative Relations in WN
in the discussed example Cause to age is derived
and FN
from Aging by enriching the set of Aging’s FEs
The causative tree stemming from {change:1} and with the frame elements Cause and Agent. In ad-
the non-causative one stemming from {change:2} dition, we posit a Causative of relation between
were aligned using the WN causes relation, result- Cause to age and Aging. In general, causative
635
frames inherit from the abstract frame Transi- FEs, e.g. {drive:1} (Operate vehicle) as a hy-
tive action so we define an Inheritance relation be- ponym of {operate:3} (Operating a system) ap-
tween Transitive action and Cause to age. In such plies only to land vehicles while other verbs in the
a way the newly-defined relation is integrated into frame impose different restrictions on the FE Ve-
the FN relational structure. hicle;
Example 8. Frame: Cause to age – Profiling a different FE from the one profiled
Core frame elements: Agent/Cause; Entity by the hypernym, e.g. {rob:1} (Robbery) profiles
FN definition: An Agent or Cause causes an En- the Victim, while its hypernym {steal:1; rip off:2,
tity to undergo a change in age typically associated rip:4} (Theft) profiles the stolen Goods;
with some deterioration or change in state. – Inclusion/exclusion of a causative/agentive
Example synsets: {age:3} ’make older’ FEs corresponding to a causative subevent in
FN relation: Inherits from the respective pairs of frames, e.g. {break:5}
Frame: Transitive action (Cause to fragment) and {break:2, separate:10,
Core frame elements: Agent/Cause; Patient fall apart:4, come apart:1} (Breaking apart).
Frame definition: An Agent or Cause affects a Some of the types of specialisation are currently
Patient. being studied as a point of departure for defining
FN relation: Is Causative of more narrow-scope frames that would allow for
Frame: Aging more precise predictions about the selectional re-
Core frame elements: Entity strictions and the syntactic realisation of FEs.
Frame definition: An Entity is undergoing a
change in age typically associated with some 7 Future Work
deterioration or change in state. A further goal is to enrich FN by extending its lex-
Example synset: {senesce:1, age:2, get on:7, ical coverage on the basis of the expanded map-
mature:5, maturate:2} ’grow old or older’ ping to synsets. Verbs which do not have corre-
The domain of causativity provides an approach spondence among LUs (or no correspondence in a
at symmetricising large parts of the lexicon both given frame) but belong to synsets that have been
at a horizontal level (same level lexemes in a tax- successfully mapped to FN frames, will be sug-
onomical hierarchy) and in depth as the improve- gested as possible LUs to be included in the re-
ments in the higher levels of the lexicon influence spective frame(s).
the deeper levels as reflected in the procedure of A venue of ongoing research is to define precise
assigning relations by inheritance described in 4.1. selectional restrictions on FEs and to implement
them as semantic relations between a verb synset
6 Frame Specialisation and Relations and a set of noun synonyms that satisfy these re-
The observations on hierarchical relations, es- strictions. In such a way we will enrich WN with
pecially on the more populated ones, such as relations between verbs and nouns corresponding
Inheritance, Using and See also, shed light on to participants in their conceptual structure, partic-
the specialisation that takes place from parent ularly ones realised as arguments and adjuncts.
to child in the taxonomic (inheritance) hierar- The developed resource may have a consider-
chy. The changes in the causativity domain deal able impact on the development of methods for
with including/excluding FEs that correspond to identification of predicate-argument structure in
causative subevents in the event structure. The text, which in turn will facilitate the development
modifications that occur in the conceptual and se- of new methods for frame verification and consis-
mantic structure include, but are not limited to the tency checks on FN and WN.
following:
Acknowledgments
– Reducing the number of core frame ele-
ments by incorporating one of them in the verb’s This study has been carried out as part of the
meaning, e.g. {whip:4} incorporates the periph- project Towards a Semantic Network Enriched
eral FE Instrument (’whip’) of {strike:1} in the with a Variety of Semantic Relations funded by the
frame Cause harm assigned to both; National Science Fund of the Republic of Bulgaria
– Reducing the scope of the frame through im- under the Fundamental Scientific Research Pro-
posing more strict selectional restrictions on the gramme (Grant Agreement 10/3 of 14.12.2016).
636
References Miriam R. Petruck. 2015. The Com-
ponents of FrameNet. http://
Collin F. Baker. 2008. FrameNet, present and future. naacl.org/naacl-hlt-2015/
In Jonathan Webster, Nancy Ide, and Alex Chengyu tutorial-framenet-data/
Fang, editors, The First International Conference FNComponentsMRLP.pdf.
on Global Interoperability for Language Resources.
City University, City University, Hong Kong. Miriam R. Petruck and Gerard de Melo. 2012. Pre-
cedes: A semantic relation in FrameNet. In Pro-
Collin F. Baker and Christiane Fellbaum. 2009. Word- ceedings of the Workshop on Language Resources
Net and FrameNet as Complementary Resources for for Public Security Applications. pages 45–49.
Annotation. In Proceedings of the Third Linguistic
Annotation Workshop (ACL-IJCNLP ’09), Associa- Josef Ruppenhofer, Michael Ellsworth, Miriam R.
tion for Computational Linguistics, Stroudsburg, PA, Petruck, Christopher R. Johnson, C. F. Baker, and
USA. pages 125–129. Jan Scheffczyk. 2016. FrameNet II: Extended The-
ory and Practice. International Computer Science
Collin F. Baker, Charles J. Fillmore, and John B. Lowe. Institute, Berkeley, California.
1998. The Berkeley FrameNet Project. In COLING-
ACL ’98: Proceedings of the Conference. Montreal, Lei Shi and Rada Mihalcea. 2005. Putting Pieces To-
Canada. pages 86–90. gether: Combining FrameNet, VerbNet and Word-
Net for Robust Semantic Parsing. In A. Gel-
Collin F. Baker and Josef Ruppenhofer. 2002. bukh, editor, Computational Linguistics and Intelli-
FrameNet’s frames vs. levin’s verb classes. In gent Text Processing. CICLing 2005. Lecture Notes
J. Larson and M. Paster, editors, Proceedings of 28th in Computer Science, Springe, Berlin, Heidelbergr,
Annual Meeting of the Berkeley Linguistics Society. volume 3406.
pages 27–38.
Sara Tonelli and Daniele Pighin. 2009. New Features
Christiane Fellbaum. 1998. A Semantic Network of for Framenet – Wordnet Mapping. In Proceedings of
English Verbs . In Christiane Fellbaum, editor, the Thirteenth Conference on Computational Natu-
WordNet: An Electronic Lexical Database, MIT ral Language Learning (CoNLL’09), Boulder, USA.
Press, Cambridge, MA, pages 69–104.
Robert D. Van Valin Jr. and Randy J. LaPolla. 1997.
E. Laparra and G. Rigau. 2009. Integrating Word- Syntax: Structure, meaning and function . Cam-
Net and FrameNet using a knowledge-based Word bridge University Press.
Sense Disambiguation algorithm. In Proceedings of
Recent Advances in Natural Language Processing
(RANLP09), Borovets, Bulgaria. pages 208–213.
637
Compositional Hyponymy with Positive Operators
Martha Lewis
ILLC, University of Amsterdam
m.a.f.lewis@uva.nl
638
Proceedings of Recent Advances in Natural Language Processing, pages 638–647,
Varna, Bulgaria, Sep 2–4 2019.
network methods, as seen in Socher et al. (2013); in Kiela et al. (2015) is to incorporate other modes
Bowman et al. (2014), simpler element-wise com- of input into the representation of a word. Mea-
bination methods (Mitchell and Lapata, 2010), sures of entailment are based on the dispersion of
and tensor-based methods (Baroni and Zampar- a word representation, together with a similarity
elli, 2010; Coecke et al., 2010; Paperno et al., measure. All of these look at entailment at the
2014). Tensor-based methods operate by mod- word level. Related to the current work are the
elling words of different grammatical types in dif- ideas in Balkır (2014); Balkır et al. (2016). In this
ferent vector spaces, and viewing relational words work, the authors develop a graded form of en-
such as verbs and adjectives as linear maps that tailment based on von Neumann entropy and with
operate on their arguments. This allows methods links to the distributional inclusion hypotheses de-
from formal semantics to be more easily mapped veloped by Geffet and Dagan (2005). The au-
onto vector space representations, and thereby thors show how entailment at the word level car-
gives us mechanisms for composing words, in line ries through to entailment at the sentence level.
with their grammatical types, to form phrases and
sentences. More recent approaches involve specialising
We use an extension of the tensor-based ap- word vectors for entailment Vulić and Mrkšić
proach, based on the methods given in Piedeleu (2018), using non-Euclidean geometries Nickel
et al. (2015); Bankova et al. (2019); Balkır et al. and Kiela (2017); Nguyen et al. (2017); Le et al.
(2016). We represent nouns as positive operators, (2019), and using pattern-based hyponymy extrac-
which can be considered as representing collec- tion Roller et al. (2018); Le et al. (2019).
tions of vectors. Functional words like adjectives
and verbs are now represented as completely posi- Most approaches, however, provide only word-
tive maps, i.e. linear maps which preserve the pos- word hyponymy. To test hyponymy in a composi-
itivity of their arguments. These can be thought of tional setting, we refer to the dataset of Kartsaklis
as linear maps which take valid collections of vec- and Sadrzadeh (2016) where a number of sentence
tors to valid collections of vectors. and phrase-level hyponymy relationships are built
from WordNet (Miller, 1995)
2.1 Related Work
Another approach to detecting lexical entail-
Entailment is an important and thriving area of ment is via the identification of certain text pat-
research within distributional semantics. The terns which indicate a hyponym-hypernym rela-
PASCAL Recognising Textual Entailment Chal- tionship. Examples are: y such as x, x is a type of y,
lenge (Dagan et al., 2006) has attracted a large which allow us to pick out pairs (x, y) which stand
number of researchers in the area and generated a in the relation x is-a y. This approach was first
number of approaches. Previous lines of research outlined in Hearst (1992) and has been recently
on entailment for distributional semantics investi- used in Roller et al. (2018) to build vectors able to
gate the development of directed similarity mea- encode the required hierarchical relationships.
sures which can characterize entailment (Weeds
et al., 2004; Kotlerman et al., 2010; Lenci and In the current paper we provide methods for
Benotto, 2012). Geffet and Dagan (2005) intro- building word representations as positive opera-
duce a pair of distributional inclusion hypotheses, tors, using hierarchical information either from
where if a word v entails another word w, then human-curated sources such as WordNet, or un-
all the typical features of the word v will also oc- supervised methods such as using Hearst patterns.
cur with the word w. Conversely, if all the typi- We will show how these word representations can
cal features of v also occur with w, v is expected be composed, and how the hierarchical informa-
to entail w. Clarke (2009) defines a vector lattice tion percolates to the phrase level. Our contri-
for word vectors, and a notion of graded entail- bution is to provide a means of building hierar-
ment with the properties of a conditional proba- chically ordered word representations, that can be
bility. Rimell (2014) explores the limitations of composed into phrases and sentences. Previous
the distributional inclusion hypothesis by exam- work in this area has either concentrated on word-
ining the properties of those features that are not level hyponymy or phrase-level hyponymy. In this
shared between words. An interesting approach paper we combine the two in one framework.
639
3 Methods Example 1 (Nouns). Consider the noun pet, and
suppose we have three types of pet: a pug, a gold-
We model words as collections of vectors, as fol-
fish, and a tabby cat. We give these values in a
lows. For a given vector ~v ∈ V ,1 we can ‘lift’ this
distributional space spanned by the basis vectors
vector into the larger space V ⊗ V , by taking the −−→ −−−−−→ −−−−−→ −−−−→
{furry, domestic, working, aquatic} as follows:
outer product of the vector with itself. We use the
following notation: pug goldfish tabby
furry 3 0 5
v̄ := ~v~v > (1) domestic 4 5 5
working 0 0 0
When ~v is a unit vector, the resulting matrix v̄ is aquatic 0 6 0
a projection operator. Multiplying another vector We form the representation of the noun pet by
~x by v̄ projects ~x onto the one-dimensional sub- summing over the matrix representations of each
space spanned by ~v . A matrix of the form v̄ can be vector:
thought of as a collection of just one vector, giving
sharp, unambiguous information. JpetK = pug + goldfish + tabby
To represent collections of more than one vec- →> + − −→ −−→ −−−→ −−−→
=− →−
pug pug gfish gfish> + tabby tabby>
tor, we sum together their matrix representations,
resulting in another matrix: 34 37 0 0
37 66 0 30
=0 0 0 0
{~v , w,
~ ~x} 7→ v̄ + w̄ + x̄ ∈ V ⊗ V
= ~v~v > + w
~w~ > + ~x~x> ∈ V ⊗ V 0 30 0 36
Matrices M built in this way are called positive Each of the matrices pug, goldfish, and tabby
operators and have the following two properties: has just one non-zero eigenvalue, which is ||→ −v ||,
and corresponds to the normalised eigenvector
• ∀v ∈ V.h~v , M~v i ≥ 0 →
−v /||→
− v = −
v ||, for →
− → −
pug,
−−−→ −−−→
goldfish, and tabby re-
spectively.
• M is self-adjoint. The matrix JpetK, however, has three non-zero
eigenvalues of 100.52, 35.21, and 0.25, each cor-
If we additionally impose that M has trace 1, then
responding to a combination of the basis vectors
we can understand M as encoding a probability −−→ −−−−−→ −−−−→ −−−−−→
distribution over ~v ∈ V (Nielsen and Chuang, furry, domestic, aquatic. The basis vector working
2010). In the present work, we do not impose has an eigenvalue of 0, indicating that JpetK is or-
−−−−−→
this condition, instead viewing M as representing thogonal to the vector working.
a collection of vectors. 3.1 Ordering Positive Operators
The eigenvectors and eigenvalues of M can be
thought of as providing a summary of the infor- The set of positive operators on a vector space has
mation contained in M . A matrix of the form an ordering introduced by Löwner (1934). For
v̄ = ~v~v > will have one non-zero eigenvalue, cor- positive operators A and B, we define:
responding to the normalized eigenvector ~v /||~v ||.
A v B ⇐⇒ B − A is positive
When multiple vectors have been included in the
collection, the matrix M will have more than one In Bankova et al. (2019) the authors introduce a
non-zero eigenvalue, and these will represent the notion of graded hyponymy. The hyponymy rela-
weights for their corresponding eigenvectors. tion may be true up to some error term, as follows.
We take a kind of extensional stance. We con- If A v B, then B −A = D, where D is some pos-
sider words to be modelled as collections of their itive operator. If this does not hold, it is possible to
instances. To model a noun, we can consider add in some error term E so that A v B +E. This
the collection of nouns that are hyponyms of that is viewed as saying that A entails B to the extent
noun, and form the matrix representation corre- E. We wish to find the smallest such error term.
sponding to that collection. In Bankova et al. (2019), the error term was of
1
Throughout the paper, we assume that vector spaces are the form (1 − k)A and the scalar k ∈ [0, 1] gave
Rn a graded notion of hyponymy. The effect of this
640
scalar is to reduce the size of A until it ‘fits inside’ Now, since pug and tabby are both positive, and
B, giving a notion of graded hyponymy that says positivity is preserved under addition, we know
that A is a k-hyponym of B, A vk B if B − kA that JpetK − goldfish is also positive. Therefore,
is positive. under either of our graded measures, the extent to
One of the drawbacks of this measure is that which a goldfish is a pet is 1.
if the space spanned by eigenvectors of A, called Example 3 (Graded Hyponymy). Now suppose
Span(A), is not a subspace of Span(B), then the that we define JdogK = pug + collie, with −→ and
pug
value of k must be 0. We therefore consider two −−−→
collie defined as below:
new measures, which we now describe. If B − A
pug collie
is not positive, it is possible to make it positive
furry 3 3
by adding in a positive operator constructed in the
domestic 4 2
following manner.
working 0 2
1. Firstly diagonalize B − A, resulting in a real- aquatic 0 0
valued matrix, since B −A is real symmetric. Then to determine whether a dog is a pet, we cal-
culate:
2. Construct a matrix E by setting all positive
entries of B − A to 0 and changing the sign JpetK − JdogK
of all negative eigenvalues.
34 37 0 0 18 18 6 0
Then B − A + E will give us a positive matrix. 37 66 0 30
= − 18 20 4 0
This E is our error term. The size of E is bounded 0 0 0 0 6 4 4 0
above by the size of A, since certainly B − A + 0 30 0 36 0 0 0 0
A is positive. We propose two different measures
16 19 −6 0
related to this error term that give us a grading for 19 46 −4 30
hyponymy. =−6
−4 −4 0
The first measure is
P 0 30 0 36
λi
kBA = P i (2) The eigenvalues of JpetK − JdogK are 75.38, 24.39,
i |λi |
-5.77, 0, i.e., they are not all positive. This is be-
where λi is the ith eigenvalue of B − A and | · | in- cause the subspace corresponding to JdogK is not
dicates absolute value. This measures the propor- a subspace of JpetK, in particular because JdogK is
tions of positive and negative eigenvalues in the −−−−−→
not orthogonal to the basis vector working.
expression B − A. If all eigenvalues are nega- However, much of JdogK is included in JpetK.
tive, kBA = −1, and if all are positive, kBA = Using our graded measures given in equations (2)
1. This measure is symmetric in the sense that and (3), we can calculate that under kBA , dog is a
kBA = −kAB . hyponym of pet to the extent 0.89 and under kE ,
Secondly, we propose dog is a hyponym of pet to the extent 0.86.
||E||
kE = 1 − (3) 3.2 Composing Positive Matrices
||A||
To compose positive matrices, we combine the
where ||·|| denotes the Frobenius norm. This mea- methods outlined in Bankova et al. (2019) with the
sures the size of the error term as a proportion of type-lifting methods outlined in Kartsaklis et al.
the size of A. Since A = E in the worst case, this (2012) to lift word representations into a higher-
measure ranges from 0 when E = A to 1 when dimensional space. The impact of these methods
E = 0. are that we can define nouns and verbs within the
Example 2 (Full Hyponymy). Recall the example same distributional space, and then lift the verb
representations to a space that corresponds to a
JpetK = pug + goldfish + tabby completely positive map.
To determine whether a goldfish is a pet, we cal- We have choices about how to implement this
culate: type-lifting. One choice is composition, i.e. ma-
trix multiplication, of the two operators. This lat-
JpetK − goldfish = pug + tabby ter operation results in a matrix that is no longer
641
self-adjoint, and so Piedeleu (2014) suggests using Table 1: Performance of word matrices derived from
the non-commutative and non-associative opera- different word embeddings, using WordNet derived hy-
1 1 ponyms. Bold highlights the highest value of each row.
tor M22 M1 M22 in its place. This operator can be
thought of as a kind of subspace projection, where GloVe Count FastText
M1 is projected onto M2 . Piedeleu (2014) also MC 0.8441 0.7124 0.8223
notes that the pointwise multiplication of two pos- MEN 0.4836 0.2861 0.5090
itive operators is a completely positive map, giving RG 0.8460 0.6325 0.8387
us another choice for composition. SIMLEX-999 0.4426 0.2638 0.4272
Following Kartsaklis et al. (2012), this gives us SimVerb 0.3458 0.2158 0.3290
a method for building higher-level operators for WS-353 0.3001 0.1677 0.3033
verbs from lower-level operators. Firstly, we as- YP-130 0.5619 0.3465 0.5166
sume the noun space N ⊗ N to be equal to the
sentence space S ⊗ S, and refer to these both as
W ⊗ W . Given a representation of an intransitive vector evaluation package provided by Faruqui
verb JverbK ∈ W ⊗W , the effect of lifting the verb and Dyer (2014). We compared on the Rubin-
to a higher order space and then composing with a stein and Goodenough (Rubenstein and Goode-
noun JnounK ∈ W ⊗ W is to apply the Frobenius nough, 1965), WordSim353, (Finkelstein et al.,
multiplication to JnounK ⊗ JverbK. 2002), Miller and Charles (Miller and Charles,
For intransitive verbs we can therefore combine 1991), SimLex999 (Hill et al., 2015), SimVerb
the noun and the verb via three operations which (Gerz et al., 2016), the Yang and Powers dataset
we call Mult, MMult1 (for matrix multiplication), (Yang and Powers, 2006) and the MEN dataset
and MMult2: (Bruni et al., 2012). We did not inclue the rare
word dataset or the similarity/relatedness splits
Mult: Jn verbK = JnK JverbK (4) of WS-353. Word matrices derived from GloVe
1 1 vectors score best overall when using WordNet-
MMult1: Jn verbK = JnK JverbKJnK
2 2 (5)
1 1
derived hyponyms (table 1). This is also seen in
MMult2: Jn verbK = JverbK JnKJverbK
2 2 (6) the validation settings of the non-compositional
datasets. When using Hearst-derived hyponyms,
For transitive verbs there is one possibility for
the generated word matrices perform poorly, al-
pointwise multiplication of the operators, since
though GloVe vectors still score most highly.
this is both commutative and associative. For the
We compare our WordNet models to a sym-
second operation there are a number of composi-
bolic model called Symb. For this model we mark
tion orders. We will concentrate on two which re-
two words w1 and w2 as being in the hyponym-
flect the difference between composing the verb
hypernym relationship if w1 is found in the transi-
with the object first and composing with the sub-
tive closure of the hyponyms of w2 .
ject first. We therefore have:
Mult: Js v oK = JsK JvK JoK (7) 4.1 Nouns
1 1 1 1 In order to build positive matrices for nouns, we
MMultO: Js v oK = JsK 2 JoK 2 JvKJoK 2 JsK 2 (8)
1 1 1 1
use information about hypernymy relations. This
MMultS: Js v oK = JoK 2 JsK 2 JvKJsK 2 JoK 2 (9) information can be elicited using human built re-
sources such as WordNet Miller (1995), or us-
4 Experimental Setting
ing Hearst patterns Hearst (1992). These are pat-
We build representations of words as positive ma- terns like ‘y such as x’, which give markers for
trices, and selected from a number of alterna- hyponym-hypernym pairs (x, y). To collect the
tive embeddings including GloVe vectors (Pen- hyponyms of a given word w from WordNet, we
nington et al., 2014), FastText (Bojanowski et al., traverse the WordNet hierarchy and collect every
2017), and distributional vectors built from the word wi in the transitive closure of the hyponymy
concatenation of the UKWaC and Wacky cor- relation.
pora using PPMI and dimensionality reduction, For hyponyms generated by Hearst patterns,
all with 300 dimensions. To select the vec- we use the publicly available dataset described in
tor embeddings, we built word matrices as de- Roller et al. (2018), and refer the reader to that
scribed above and tested them using the word paper for details of its creation. The dataset con-
642
sists of a set of word pairs P = {(xi , yi )}i which and WaCky corpora (Ferraresi et al., 2008), and
are in a hyponym-hypernym relationship, together collect those arguments that appear at least 300
with the count w(x, y) of the number of times times in the corpus.
that relationship has been seen in the text. As de-
scribed in Le et al. (2019), the relationships thus 4.3 Building Matrices
extracted are both noisy and sparse, containing Finally, having collected instances of nouns and
cycles and inconsistencies. As one example of verbs, we take the vectors → −
w i corresponding to
this, the dataset contains the pair (rome, european each of these instances, take the outer product of
country). To mitigate these phenomena, we apply each with itself, and add these together, i.e.:
a ppmi weighting to the counts. The weighting is X
→
−
as described in Roller et al. (2018), specifically JwK = w i→
−
w>i ∈W ⊗W (10)
i
p(x, y) We have discarded weighting information. Words
ppmi(x, y) = max 0, log − ,
p (x)p+ (y) which have more instances are both more widely
P dispersed in terms of their eigenvalues, and also
where, letting W = (x,y)∈P , we have:
larger in terms of their matrix norm.
p(x, y) = w(x, y)/W 4.4 Datasets
X
p− (x) = w(x, y)/W We evaluate single word representations on the
(x,y)∈P non-compositional BLESS hyponymy subset (Ba-
X roni and Lenci, 2011), WBLESS (Weeds et al.,
+
p (x) = w(y, x)/W
2014), and BIBLESS (Kiela et al., 2015) datasets.
(y,x)∈P
We test our models using the hypernymy suite pro-
These equations give, respectively, the probability vided by Roller et al. (2018). The BLESS dataset
that the pair (x, y) is chosen from P, the proba- requires the model to infer the direction of a hy-
bility that x appears as a hyponym, and the proba- pernym pair. All pairs in the model are indeed
bility that x appears as a hypernym. This sets the in the hyponym-hypernym relationship, and the
weighting of various unwanted pairs, such as the model must identify that this is the case. WB-
aforementioned (rome, european country), to 0. LESS consists of a set of pairs which may be in
To collect the set of hyponyms of a noun, the hyponym-hypernym relationship, or unrelated.
we use only those hyponyms with a non-zero Each pair is assigned a value of 1 or 0 based on
ppmi weighting, and take one transitive step. So whether or not there is a hyponymy relationship.
the set of hyponyms of a given word x is the The software provided performs 1000 random it-
union of the sets {yi |ppmi(x, yi ) > 0}i and erations in which 2% of the data is used as a val-
{zij |ppmi(yi , zij ) > 0}ij . We limit to one tran- idation set to learn a classification threshold, and
sitive step to again mitigate the noisiness of the tests on the remainder of the data. Average ac-
dataset. curacy across all iterations is reported. The BIB-
LESS dataset assigns values of 1, 0, and -1 based
4.2 Verbs on whether the first word is a hyponym of the sec-
WordNet contains verb hyponymy relationships, ond, whether there is no relationship, or whether
and therefore we can use similar methods to ex- the second is a hyponym of the first. The soft-
tract lists of hyponyms. However, we cannot use ware firstly tunes a threshold using 2% of the data,
Hearst patterns to collate verb hyponymy relation- identifying whether a pair exhibits hypernymy in
ships. As a proxy, we represent verbs as col- either direction. Secondly, the relative comparison
lections of their arguments. The intuition behind of scores determines which direction is predicted.
this is that of the extensional approach in formal Again, the average accuracy over 1000 iterations
semantics, mapped to distributional semantics in is reported.
Grefenstette and Sadrzadeh (2011). We can think We further test on the compositional datasets
of both nouns and verbs as predicates, and con- from Kartsaklis and Sadrzadeh (2016). This is
sider the instances that the predicate applies to. a series of three datasets, covering simple in-
To collect the arguments of the verbs, we use the transitive sentences, transitive sentences, and verb
concatenation of the dependency parsed ukWaC phrases. The intransitive verb dataset consists of
643
paired sentences consisting of a subject and a verb.
Table 2: Accuracy on the variants of the BLESS
In half the cases the first sentence entails the sec- dataset. HyperVec figures are from Nguyen et al.
ond, and in the other half of cases, the order of the (2017), Hearst from Roller et al. (2018), HypeCones
sentences is reversed. For example, we have: from Le et al. (2019), LEAR from Vulić and Mrkšić
(2018). Entries tagged with WN use WordNet.
summer finish, season end, T
Model BLESS WBLESS BIBLESS
season end, summer finish, F HyperVec - WN 0.92 0.87 0.81
Hearst 0.96 0.87 0.85
HypeCones 0.94 0.90 0.87
The first sentence is marked as entailing, whereas LEAR - WN 0.96 0.92 0.88
the second is marked as not entailing. The dataset Symb - WN 0.91 0.93 0.91
kBA - WN 0.95 0.88 0.84
is created by selecting nouns and verbs from kE - WN 0.95 0.91 0.87
WordNet that stand in the correct relationship. kBA - Hearst 0.91 0.84 0.76
To test our models, we build the basic word kE - Hearst 0.91 0.86 0.80
representations as in equation (10). We then use
the compositional methods outlined in section 3.2
Table 3: Area under ROC curve on the KS2016 datasets
to create the sentence representations. We calcu- using kE and WordNet derived hyponyms. For the SV
late the graded entailment value between the com- and VO datasets, MMult1 refers to the model described
posed sentence representations, and in results re- in equation (5) and MMult2 refers to the model de-
port area under ROC curve for comparison with scribed in equation (6). For SVO, MMult1 refers to the
previous literature. In particular, we compare with model described in equation (8) and MMult2 refers to
the best model from Kartsaklis and Sadrzadeh the model described in equation (9). ∗ indicates statisti-
cally significantly higher than the previous best perfor-
(2016), which uses a metric based on the distribu-
mance Kartsaklis and Sadrzadeh (2016). + indicates
tional inclusion hypothesis, together with a tensor- significantly higher than the additive baseline.
based compositional model.
Model SV VO SVO
4.5 Significance Testing KS2016 best 0.84 0.82 0.86
Verb only 0.870∗ 0.944∗ 0.908∗
To test significance of our results, we use boot- Addition 0.941∗ 0.948∗ 0.972∗
Mult 0.975∗+ 0.981∗+ 0.978∗
strapping Efron (1979) to calculate 100 values of
MMult1 0.970∗+ 0.971∗+ 0.965∗
the test statistic (either accuracy or AUC) drawn MMult2 0.967∗+ 0.969∗+ 0.971∗
from the distribution implied by the data. We com-
pare with figures from the literature using a one-
sample t-test, and compare between models using Table 4: Area under ROC curve on the KS2016
a paired t-test. We apply the Bonferroni correction datasets, using kBA and WordNet derived hyponyms.
Refer to Table 3 for explanations.
to compensate for multiple model comparisons.
Model SV VO SVO
5 Results KS2016 best 0.84 0.82 0.86
Verb only 0.902∗ 0.967∗ 0.931∗
5.1 BLESS Variants Addition 0.970∗ 0.964∗ 0.978∗
Mult 0.974∗ 0.984∗+ 0.982∗
We present results on variants of the BLESS MMult1 0.987∗+ 0.985∗+ 0.995∗+
MMult2 0.987∗+ 0.985∗+ 0.995∗+
dataset in terms of accuracy, for comparison with
other models, presented in table 2. Our best per-
foming model is the WordNet based model with Table 5: Area under ROC curve on the KS2016
metric kE . Althoough this model does not out- datasets, using kE and Hearst-pattern derived hy-
perform the best supervised model (the differences ponyms. Refer to Table 3 for explanations.
in score are significant), the differences are fairly
Model SV VO SVO
minimal (0.01 accuracy). Our methods (and those KS2016 best 0.84 0.82 0.86
of others) outperform the symbolic baseline for the Verb only 0.714 0.808 0.716
BLESS dataset. Our WordNet-based model does Addition 0.877∗ 0.807 0.912∗
Mult 0.887∗ 0.808 0.864
outperform the earlier model HyperVec with sig- MMult1 0.902∗+ 0.824+ 0.883∗
nificance. Hearst-pattern based representations do MMult2 0.903∗+ 0.800 0.877∗
not perform so strongly.
644
Table 6: Area under ROC curve on the KS2016 ing oven. What cannot be remedied, however, is
datasets, using kBA Hearst-pattern derived hyponyms. where a term has no hyponyms in WordNet. For
Refer to Table 3 for explanations.
example, herbivore has no hyponyms in WordNet.
Model SV VO SVO This means that the WordNet-based representa-
KS2016 best 0.84 0.82 0.86 tions have no way of forming a wide representa-
Verb only 0.719 0.819 0.716
Addition 0.880∗ 0.811 0.909∗
tion of herbivore thatincludes any of its instances.
Mult 0.867∗ 0.792 0.843 As well as performing well on single-word hy-
MMult1 0.909∗+ 0.842∗+ 0.930∗+ ponymy datasets, the representations we build sit
MMult2 0.904∗+ 0.830∗+ 0.924∗+
within a compositional framework that allows us
to form phrases and sentences and to reason about
5.2 Compositional Datasets their entailment relationships. The WordNet-
based representations behaved particularly well on
On the KS2016 compositional datasets results are
this dataset, due to the fact that the dataset is built
reported in terms of area under ROC curve. Our
from WordNet. However, it is still an interesting
measures perform particularly well with WordNet
set of results in that our graded measures interact
derived hypernyms (Tables 3 and 4). This is likely
well with the compositional methods we have pro-
to be due to the fact that both the dataset and
posed. Note that the measures we propose result
our word representations were constructed from
in high baseline values to beat - i.e. for the verb-
WordNet, and hence the high performance is to be
only and addition models. Again, this is likely
expected. More interestingly, the word represen-
due to the construction of the dataset. The dataset
tations built using unsupervised methods also out-
is formed from upwardly-monotone contexts, so
perform previous best scores on this dataset, (ta-
computing entailment based only on the verb will
bles 5 and 6), based on a form of the distributional
still perform extremely well. Again, although this
inclusion hypothesis for tensor-based composition
is due to the construction of the dataset, is is inter-
(Kartsaklis and Sadrzadeh, 2016), despite not per-
esting to note that the measures and word repre-
forming so strongly on the single-word datasets.
sentations we developed can model this structure
6 Discussion and Further Work so well. Furthermore the Hearst-pattern derived
representations also outperformed previous work,
We have suggested a mechanism for building the indicating that these representations interact nicely
positive operators needed for the theory presented with compositionality.
in Bankova et al. (2019), together with two novel Similarities to our approach can be found in the
measures of graded hyponymy. We tested these notion of words being represented as Gaussians
representations and measures on a number of well- (Jameel and Schockaert, 2017; Vilnis and McCal-
known datasets, looking at similarity at the word lum, 2014). The positive operators we build have
level, hyponymy at the word level and one of the same structure as covariance matrices and, if
which gives hyponymy at the phrase and sen- appropriately normalized, are interpreted as rep-
tence level. The representations and the mea- resenting a probability distribution over vectors.
sures we have developed perform competitively on Representing words as Gaussians does not come
these datasets. We have used both human-curated with a given mechanism for composing words as
information and unsupervised methods to build we do. Exploring these connections is an area of
the word representations. Unsurprisingly, human- further work.
curated information gives better performance. Finally, a crucial extension to this whole ap-
A nice comparison is with the symbolic model. proach is to be able to model hyponymy, compo-
The fact that our WordNet-based models outper- sition, and their interaction in more contexts, for
form this baseline shows that the models we pro- example using the natural logic introduced in Bar-
pose can provide a ‘smoothed’ representation that wise and Cooper (1981).
allows hyponymy relationships to be inferred. For
example, one of the hyponymy relationships not Acknowledgements
picked up in WordNet is (oven, device). How-
ever, there are a number of other instances such as Martha Lewis gratefully acknowledges funding
(electric appliance, device) that are similar enough from NWO Veni grant ‘Metaphorical Meanings
to oven that device can be understood as includ- for Artificial Agents’
645
References B. Efron. 1979. Bootstrap methods: Another look at
the jackknife. Ann. Statist., 7(1):1–26.
Esma Balkır. 2014. Using density matrices in a compo-
sitional distributional model of meaning. Master’s Manaal Faruqui and Chris Dyer. 2014. Community
thesis, University of Oxford. evaluation and exchange of word vectors at word-
Esma Balkır, Mehrnoosh Sadrzadeh, and Bob Coecke. vectors.org. In Proc. of ACL: System Demonstra-
2016. Distributional sentence entailment using den- tions.
sity matrices. In Topics in Theoretical Computer
Science, pages 1–22. Springer. Adriano Ferraresi, Eros Zanchetta, Marco Baroni, and
Silvia Bernardini. 2008. Introducing and evaluating
Dea Bankova, Bob Coecke, Martha Lewis, and Dan ukwac, a very large web-derived corpus of english.
Marsden. 2019. Graded hyponymy for composi- In Proceedings of the 4th Web as Corpus Workshop
tional distributional semantics. Journal of Language (WAC-4) Can we beat Google, pages 47–54.
Modelling, 6(2):225–260.
Lev Finkelstein, Evgeniy Gabrilovich, Yossi Matias,
Marco Baroni and Alessandro Lenci. 2011. How we Ehud Rivlin, Zach Solan, Gadi Wolfman, and Ey-
blessed distributional semantic evaluation. In Pro- tan Ruppin. 2002. Placing search in context: The
ceedings of the GEMS 2011 Workshop on GEomet- concept revisited. ACM Transactions on informa-
rical Models of Natural Language Semantics, pages tion systems, 20(1):116–131.
1–10. Association for Computational Linguistics.
Marco Baroni and Roberto Zamparelli. 2010. Nouns John R Firth. 1957. A synopsis of linguistic theory,
are vectors, adjectives are matrices: Representing 1930-1955. Studies in linguistic analysis.
adjective-noun constructions in semantic space. In
Proceedings of the 2010 Conference on Empirical Maayan Geffet and Ido Dagan. 2005. The distribu-
Methods in Natural Language Processing, pages tional inclusion hypotheses and lexical entailment.
1183–1193. Association for Computational Linguis- In Proceedings of the 43rd Annual Meeting on ACL,
tics. pages 107–114. ACL.
Jon Barwise and Robin Cooper. 1981. Generalized Daniela Gerz, Ivan Vulić, Felix Hill, Roi Reichart, and
quantifiers and natural language. In Philosophy, lan- Anna Korhonen. 2016. SimVerb-3500: A Large-
guage, and artificial intelligence, pages 241–301. Scale Evaluation Set of Verb Similarity. In EMNLP.
Springer.
Edward Grefenstette and Mehrnoosh Sadrzadeh. 2011.
Piotr Bojanowski, Edouard Grave, Armand Joulin, and Experimental support for a categorical composi-
Tomas Mikolov. 2017. Enriching word vectors with tional distributional model of meaning. In Proceed-
subword information. Transactions of the Associa- ings of the Conference on Empirical Methods in
tion for Computational Linguistics, 5:135–146. Natural Language Processing, EMNLP ’11, pages
Samuel R Bowman, Christopher Potts, and Christo- 1394–1404, Stroudsburg, PA, USA. ACL.
pher D Manning. 2014. Recursive neural net-
works can learn logical semantics. arXiv preprint Zellig S Harris. 1954. Distributional structure. Word,
arXiv:1406.1827. 10(2-3):146–162.
Elia Bruni, Gemma Boleda, Marco Baroni, and Nam- Marti A Hearst. 1992. Automatic acquisition of hy-
Khanh Tran. 2012. Distributional semantics in tech- ponyms from large text corpora. In Proceedings of
nicolor. In Proceedings of the 50th Annual Meet- the 14th conference on Computational linguistics-
ing of the Association for Computational Linguis- Volume 2, pages 539–545. Association for Compu-
tics: Long Papers-Volume 1, pages 136–145. Asso- tational Linguistics.
ciation for Computational Linguistics.
Felix Hill, Roi Reichart, and Anna Korhonen. 2015.
Daoud Clarke. 2009. Context-theoretic semantics for Simlex-999: Evaluating semantic models with (gen-
natural language: an overview. In Proceedings of uine) similarity estimation. Computational Linguis-
the Workshop on Geometrical Models of Natural tics, 41(4):665–695.
Language Semantics, pages 112–119. ACL.
Bob Coecke, Mehrnoosh Sadrzadeh, and Stephen Shoaib Jameel and Steven Schockaert. 2017. Model-
Clark. 2010. Mathematical foundations for a ing context words as regions: An ordinal regression
compositional distributional model of meaning. approach to word embedding. In Proceedings of
arXiv:1003.4394. the 21st Conference on Computational Natural Lan-
guage Learning (CoNLL 2017), pages 123–133.
Ido Dagan, Oren Glickman, and Bernardo Magnini.
2006. The PASCAL recognising textual entailment Dimitri Kartsaklis and Mehrnoosh Sadrzadeh. 2016.
challenge. In Machine learning challenges. evalu- Distributional inclusion hypothesis for tensor-based
ating predictive uncertainty, visual object classifica- composition. In Proceedings of COLING 2016,
tion, and recognising tectual entailment, pages 177– the 26th International Conference on Computational
190. Springer. Linguistics: Technical Papers, pages 2849–2860.
646
Dimitri Kartsaklis, Mehrnoosh Sadrzadeh, and Stephen Robin Piedeleu. 2014. Ambiguity in categorical mod-
Pulman. 2012. A unified sentence space for categor- els of meaning. Master’s thesis, University of Ox-
ical distributional-compositional semantics: Theory ford.
and experiments. In In Proceedings of COLING:
Posters, pages 549–558. Robin Piedeleu, Dimitri Kartsaklis, Bob Coecke, and
Mehrnoosh Sadrzadeh. 2015. Open system categor-
Douwe Kiela, Laura Rimell, Ivan Vulic, and Stephen ical quantum semantics in natural language process-
Clark. 2015. Exploiting image generality for lexi- ing. arXiv:1502.00831.
cal entailment detection. In Proceedings of the 53rd
Annual Meeting of the ACL. ACL. Laura Rimell. 2014. Distributional lexical entailment
by topic coherence. In Proceedings of the 14th Con-
Lili Kotlerman, Ido Dagan, Idan Szpektor, and Maayan ference of the European Chapter of the Association
Zhitomirsky-Geffet. 2010. Directional distribu- for Computational Linguistics, pages 511–519.
tional similarity for lexical inference. Natural Lan-
guage Engineering, 16(04):359–389. Stephen Roller, Douwe Kiela, and Maximilian Nickel.
2018. Hearst patterns revisited: Automatic hy-
Matt Le, Stephen Roller, Laetitia Papaxanthos, Douwe pernym detection from large text corpora. arXiv
Kiela, and Maximilian Nickel. 2019. Inferring con- preprint arXiv:1806.03191.
cept hierarchies from text corpora via hyperbolic
embeddings. arXiv preprint arXiv:1902.00913. Herbert Rubenstein and John B Goodenough. 1965.
Contextual correlates of synonymy. Communica-
Alessandro Lenci and Giulia Benotto. 2012. Identify-
tions of the ACM, 8(10):627–633.
ing hypernyms in distributional semantic spaces. In
Proceedings of the First Joint Conference on Lexical Richard Socher, Alex Perelygin, Jean Wu, Jason
and Computational Semantics, pages 75–79. ACL. Chuang, Christopher D Manning, Andrew Ng, and
Karl Löwner. 1934. Über monotone Matrixfunktionen. Christopher Potts. 2013. Recursive deep models
Mathematische Zeitschrift, 38(1):177–216. for semantic compositionality over a sentiment tree-
bank. In Proceedings of the 2013 conference on
George A. Miller. 1995. Wordnet: A lexical database empirical methods in natural language processing,
for english. Commun. ACM, 38(11):39–41. pages 1631–1642.
George A Miller and Walter G Charles. 1991. Contex- Luke Vilnis and Andrew McCallum. 2014. Word rep-
tual correlates of semantic similarity. Language and resentations via gaussian embedding. arXiv preprint
cognitive processes, 6(1):1–28. arXiv:1412.6623.
Jeff Mitchell and Mirella Lapata. 2010. Composition Ivan Vulić and Nikola Mrkšić. 2018. Specialising word
in distributional models of semantics. Cognitive sci- vectors for lexical entailment. In Proceedings of the
ence, 34(8):1388–1429. 2018 Conference of the North American Chapter of
Kim Anh Nguyen, Maximilian Köper, Sabine Schulte the Association for Computational Linguistics: Hu-
im Walde, and Ngoc Thang Vu. 2017. Hierarchi- man Language Technologies, Volume 1 (Long Pa-
cal embeddings for hypernymy detection and direc- pers), volume 1, pages 1134–1145.
tionality. In Proceedings of the 2017 Conference on Julie Weeds, Daoud Clarke, Jeremy Reffin, David Weir,
Empirical Methods in Natural Language Process- and Bill Keller. 2014. Learning to distinguish hyper-
ing, pages 233–243. nyms and co-hyponyms. In Proceedings of COL-
Maximillian Nickel and Douwe Kiela. 2017. Poincaré ING 2014, the 25th International Conference on
embeddings for learning hierarchical representa- Computational Linguistics: Technical Papers, pages
tions. In Advances in neural information processing 2249–2259. Dublin City University and Association
systems, pages 6338–6347. for Computational Linguistics.
Michael A Nielsen and Isaac L Chuang. 2010. Quan- Julie Weeds, David Weir, and Diana McCarthy. 2004.
tum Computation and Quantum Information. Cam- Characterising measures of lexical distributional
bridge University Press. similarity. In Proceedings of the 20th interna-
tional conference on Computational Linguistics,
Denis Paperno, Marco Baroni, et al. 2014. A prac- page 1015. ACL.
tical and linguistically-motivated approach to com-
positional distributional semantics. In Proceedings Dongqiang Yang and David M. W. Powers. 2006. Verb
of the 52nd Annual Meeting of the Association for similarity on the taxonomy of wordnet. In In the 3rd
Computational Linguistics (Volume 1: Long Pa- International WordNet Conference (GWC-06), Jeju
pers), volume 1, pages 90–99. Island, Korea.
Jeffrey Pennington, Richard Socher, and Christo-
pher D. Manning. 2014. Glove: Global vectors for
word representation. In Empirical Methods in Nat-
ural Language Processing (EMNLP), pages 1532–
1543.
647
The Impact of Semantic Linguistic Features in Relation Extraction:
A Logical Relational Learning Approach
648
Proceedings of Recent Advances in Natural Language Processing, pages 648–654,
Varna, Bulgaria, Sep 2–4 2019.
stantial body of deep linguistic knowledge, in com- tokenization, passing for shallow analysis, and
bination with an expressive inductive learning finishing with more advanced semantic analysis.
technique, can generate effective RE models.
649
analysis (syntactic parsing) identifies the struc- (Niles and Pease, 2003), and WordNet Domain hi-
tural relationships holding between words at the erarchy (Bentivoli et al., 2004). Such semantic re-
sentence level. The final semantic analysis links sources offer a variety of semantic relations in-
words to lexical semantic resources, including cluding synonyms and hyponyms from WordNet,
WordNet (Fellbaum, 1998), SUMO ontology and additional semantic relations between verbs
and their arguments considered here as predicates.
1 4
https://stanfordnlp.github.io/CoreNLP https://github.com/clearnlp/clearnlp
2 5
http://opennlp.apache.org https://sourceforge.net/projects/supersensetag
3 https://sourceforge.net/projects/jwordnet 6 http://www.adampease.org/OP
650
synonyms, domain sense), and semantic roles of EQ3. How well do the rules generalize among dif-
the verbs with their arguments. ferent datasets: either in the same domain or in dis-
Structural features consist of the structural ele- tinct domains?
ments connecting all the above features according
to our relational model. They denote (i) the se- 4.2 Dataset and Evaluation Measures
quencing of tokens preserving the token order in Three publicly available RE datasets from the
the input sentence; (ii) the part-whole relation be- newswire and the biomedical domain containing
tween tokens and the chunk containing them; (iii) binary relations were selected for analysis:
the sequencing of chunks is represented by edges reACE 2004/2005. The reACE 2004/2005 datasets
between their head tokens; and (iv) the grammati- introduced by Hachey et al. (2011) are the result of
cal dependency between two tokens in a sentence several transformation steps (refactoring, prepro-
according to the typed dependencies between cessing, and reannotation) for normalizing the two
words. original ACE datasets so that they adhere to a com-
The relational representation of all the above mon notion of relation that is more intuitive and
types of features is straightforward: a unary pred- simpler: relation instance denotes a predicate over
icate in Prolog denotes identifiers, e.g., token(id), two arguments, where the arguments represent
while binary predicates correspond to attribute– concepts in the real world.
value pairs and relations, e.g., rel(arg1, arg2). Table 2 shows the distributions of the relation
NLP Subtask Tool or Resource types in reACE 2005 dataset, whereas Table 3
Tokenization shows some examples of them.
Sentence Splitter Stanford CoreNLP
POS reACE 2005 - Relation Types Freq
Lemmatization Employment 228
Chunking OpenNLP Chunker Membership 36
NER Stanford CoreNLP Located 280
Morphological Analysis Citizen-Resident-Religion-Ethnic 39
Gazetteer Look-up ad hoc programs Business 16
Pronoun Normalization Family 42
Geographical 119
Syntactic Parsing - Dependency Stanford CoreNLP
Subsidiary 47
Worde Sense Disambiguation Sense Learner
WordNet Synsets (synonyms and WordNet 3.0 Table 2: reACE 2005 relation types.
hypernyms) Relation type Example phrases
Similar words Lin ́s database business (John, superiors) John´s superiors…
SRL with Propbank/VerbNet ClearNLP Employ (Investors, “Wall Street”) Investors on Wall
Selectional Preferences SuperSense Tagger Street…
Semantic mapping to Domains WordNet domains
citizen (voters, Missouri) Some Missouri vot-
Semantic mapping to SUMO Ad hoc program
ers…
Table 1. Complete pipeline of the Deep NLP Table 3. Examples of reACE 2005 relations.
tools component.
IEPA. The Interaction Extraction Performance As-
4 Experiments sessment (IEPA) corpus (Ding et al.,2002) is a bio-
This section reports the results of experiments per- medical dataset comprising 303 abstracts retrieved
formed on three benchmark datasets from two dis- by ten queries suggested by domain experts to the
tinct domains (newswire and biomedical). PUBMED repository. An interaction between two
terms, i.e., a specific pair of co-occurring chemicals
4.1 Experimental Questions
in the IEPA corpus, was defined as a direct or indi-
We investigate the effectiveness of the proposed se- rect influence of one on the quantity or activity of
mantic linguistic features used in the induction of the other (Ding et al., 2002). Examples of interac-
the relation extraction rules by the our ILP system. tions between terms A and B are "A increased B",
More precisely, we want to answer the following and "A activated C, and C activated B".
experimental questions (EQ): Evaluation Measures. The classical IR measures
EQ1. Do the features present a complementary of Precision P, Recall R, and F1-measure (Baeza-
contribution to the performance results? Yates and Ribeiro-Neto, 1999) were used for meas-
EQ2. What is the impact of the semantic linguistic uring the effectiveness (impact) of the proposed en-
features on the final induced set of extraction hanced features on the RE task.
rules?
651
4.3 Experimental Protocol On the Impact of Semantic Linguistic Features.
We employed five-fold cross-validation which al- EQ1 can be positively answered because by incre-
lows both the maximal use of the available training mentally incorporating new groups of features to
data on all the datasets used in the experiments. the baseline, that contributed to the improvement
Moreover, preliminary experiments were per- of the scores for all datasets. Indeed, the perfor-
formed for determining the optimal learning pa- mance improves as more features are used, starting
rameters according to the criteria of achieving high with the F-measure of 77.77 and reaching 81.80 for
accuracy and preventing model overfitting. The the reACE 2004 dataset. Analogously for the
best parameter setting for the ILP-based learning reACE 2005, the best overall F1 performance
component found were: evalfn = coverage, i (71.86%) may indicate that this dataset is more dif-
(depth) = 3, minpos = 5, and noise = 0.2. ficult than the reACE 2004. One possible explana-
tion is that, in the reACE 2005 dataset, some rela-
4.4 Results tions (particularly Business) are very poorly repre-
sented with only 16 positive examples, which ham-
Table 4 summarizes the results of using several
pers the overall score. More importantly, the over-
combinations of features on reACE 2004/2005 and
all F1 scores suggest that the proposed four groups
IEPA datasets, while Tab. 5 conveniently displays
of features have both a positive and complementary
the difference in performance between each pair of
impact on the overall F1 scores for all the datasets
corresponding lines indexed by the column id.
evaluated.
Starting from Line 2 in Tab. 4, a given group of
Concerning EQ2, one can notice that including
features are incrementally added to the baseline
semantic features into the RE process improves av-
(Line 1) which, in turn, includes the following
erage performance in terms of F1 measure for all
group of features: lexical, syntactic and structural
datasets. In fact, the boost in F1 measure was 4%
features, i.e., syntactical dependencies (Dep),
in average for the reACE datasets, while for the
chunk information (Chunk), POS tagging, and
IEPA dataset, the improvement was more than 3%.
other chunk related features. The baseline setting
However, the impact on both P and R scores were
corresponds to all the features that do not take into
unbalanced for the reACE corpora, since the se-
account the semantic features (i.e., lexical seman-
mantic features contributed relatively more in re-
tics and mapping to semantic resources). The other
call than in precision. This contrast with the results
lines (Line 2-4) in Tab. 4 integrate other groups of
on the IEPA corpus that were very balanced. On the
features (NER, Corpus types) to the baseline: NER
one hand, the highest difference in performance
denotes recognized named entities whereas Corpus
was achieved on the reACE 2005 corpus, as the se-
types features denote the golden standard annota-
mantic features improved P in almost 12%. On the
tions already available in the given corpus. For in-
other hand, for two other combinations in this da-
stance, the reACE datasets provide named entities
taset (Line 5 and Line 7), adding semantic features
such as Organizations and Person, while the IEPA
to the learner in fact slightly hampered precision.
corpus only assigns the label protein to each term
Such impact on both P and R were expected since
denoting a given protein. The last group of features
the effect of adding semantic features to the learner
(semantic), denotes the semantic features compris-
could not only improve R over P, but also provide
ing SRL, synonyms/hypernyms, and mapping of
to it an extended layer of categorization of all
words to WordNet, WordNet Domains, SUMO on-
terms. Contrastingly, for the IEPA corpus, the use
tology, similar words, and selectional preferences.
of semantic features slightly increased precision
The missing entries in IEPA column are due to the
more than recall. After inspecting the final induced
fact that typical named entities are useless in the
extraction rules, we found that this was mainly due
IEPA biomedical corpus, and therefore, they were
to the semantic role labeling features. Actually,
not considered.
many verbs denoting the interaction between two
5 Discussion on Experimental Questions proteins terms in IEPA corpus were correctly anno-
This section discusses both the impact of semantic tated along with the roles of its arguments. As a re-
linguistic features on RE, and related aspects on sult, the ILP-based learner is more precise when a
domain adaptability. given verb has semantic role features attached to it.
652
Table 4. Results on reACE 2004/2005 and IEPA datasets.
653
References Mihalcea R. and E. Faruque. 2004. SenseLearner: Min-
imally supervised word sense disambiguation for all
Amal Zouap. 2011. An Overview of Shallow and Deep words in open text”. In Proceedings of ACL/SIG-
Natural Language Processing for Ontology Learn- LEX Senseval-3, Barcelona, Spain.
ing. In W. Wong, W. Liu, & M. Bennamoun (Eds.),
Ontology Learning and Knowledge Discovery Us- Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean.
ing the Web: Challenges and Recent Advances. Her- 2013. Efficient Estimation of Word Representations
shey, PA: IGI Global, pages 16-37. in Vector Space. In Proceedings of Workshop at
ICLR.
Baeza-Yates, R., Ribeiro-Neto, B., 1999. Modern In-
formation Retrieval. Addison-Wesley, Boston. Miwa, M., & Bansal, M. 2016. End-to-End Relation
Extraction using LSTMs on Sequences and Tree
Bentivogli L., Forner P., Magnini B. and Pianta E. Structures. In Proceedings of the 54th Annual Meet-
2004. Revising WordNet Domains Hierarchy: Se- ing of the Association for Computational Linguis-
mantics, Coverage, and Balancing. In COLING tics, Berlin, Germany, pages 1105–1116.
Workshop on Multilingual Linguistic Resources,
Geneva, Switzerland, pp. 101-108. Muggleton S. 1991. “Inductive Logic Programming”
New Generation Computing 8 (4): 29.
Björne, J., Salakoski, T., 2015. TEES 22: Biomedical
event extraction for diverse corpora. BMC Bioinfor- Nguyen, T.H., Grishman, R. 2015. Relation extraction:
matics 16 (Suppl 16), S4, PMC Web. Perspective from convolutional neural networks. In:
In Proceedings of the 1st Workshop on Vector Space
Choi S.P, S. Lee., H. Jung, and S. Song. 2013. An In- Modeling for Natural Language Processing. pages
tensive Case Study on Kernel-based Relation Ex- 39–48.
traction. In Proceedings of Multimedia Tools and
Applications, Springer, US, 2013, pp. 1 -27. Niles, I. and Pease, A. Linking Lexicons and Ontolo-
gies: Mapping WordNet to the Suggested Upper
Ciaramita M., Y. Altun. 2005. Named-entity recogni- Merged Ontology. 2003. In Proceedings of the 2003
tion in novel domains with external lexical International Conference on Information and
knowledge. Adv. in Structured Learning for Text and Knowledge Engineering (IKE 03), Las Vegas, Ne-
Speech Processing Workshop (NIPS). vada, June 23-26.
Ciaramita M., Yasemin Altun. 2006. Broad-Coverage Pyysalo S. 2008. A dependency parsing approach to bi-
Sense Disambiguation and Information Extraction omedical text mining, Ph.D. thesis. Department of
with a Supersense Sequence Tagger. EMNLP, pages Information Technology, University of Turku, 2008.
594- 602.
Luc De Raedt. 2008. Logical and Relational Learning:
Ding J., Berleant, D., Nettleton, D., and Wurtele, E. From ILP to MRDM (Cognitive Technologies).
2002. Mining MEDLINE: abstracts, sentences, or Springer-Verlag, Berlin, Heidelberg.
phrases? In Proc. of the Pacific Symposium on Bio-
computing, 326-337. Santos J. 2010. Efficient Learning and Evaluation of
Complex Concepts in Inductive Logic Program-
Fellbaum, C.D., 1998. WordNet – An Electronic Lexi- ming. Ph.D. Thesis, Imperial College.
cal Database. Language, Speech and Communica-
tion. MIT Press.
Hachey H., C. Grover, and R. Tobin. 2011. Datasets for
Generic Relation Extraction. Journal of Natural
Language Engineering, Cambridge University
Press.
Lin D., Zhao S., Qin L., Zhou M. 2003. Identifying
Synonyms among Distributionally Similar Words.
IJCAI 2003: pages 1492-1493.
Lima R., B. Espinasse, F. Freitas. 2017. OntoILPER:
an Ontology- and Inductive Logic Programming-
based Relations from Text, in: Knowledge And In-
formation System (KAIS) Journal, Springer London.
Lima, R., Espinasse, B. Freitas, F. 2019. A logic-based
relational learning approach to relation extraction:
The OntoILPER system. Journal of Engineering Ap-
plications of Artificial Intelligence. Volume 78,
pages 142-157.
654
Detecting Anorexia in Spanish Tweets
655
Proceedings of Recent Advances in Natural Language Processing, pages 655–663,
Varna, Bulgaria, Sep 2–4 2019.
toring and surveillance. Specifically, Natural Lan- cial media data for health informatics. Further-
guage Processing (NLP), also known as Language more, eRisk (Losada et al., 2017) is a challenging
Technologies (LT) can be used to generate sys- workshop focused on mental health disorders and
tems for early anorexia detection. One of the main it has been held from 2017 in the framework of
problems is the lack of resources to train systems the well-known international conferences CLEF1 .
and more if we focus on a language other than En- eRisk explores the evaluation methodology, effec-
glish. tiveness metrics and practical applications (partic-
The main goal of this paper is to develop a sys- ularly those related to health and safety) of early
tem for the automatic detection of anorexia in tex- risk detection on the Internet. The different tasks
tual information. For this, we first generated a cor- proposed include depression and anorexia detec-
pus with tweets written in Spanish including both tion.
people talking about anorexia and people talking Concerning to mental health, we can find some
about healthy food habits. The corpus is called interesting papers studying NLP techniques for
SAD (Spanish Anorexia Detection). Using the treating textual information. (Rahman et al., 2018)
SAD corpus, we have developed different models review several studies focused on detecting men-
based on Machine Learning approaches that inte- tal health using and analyzing the information ex-
grate several linguistic features. We have analyzed tracted from social networks. After analyzing sev-
the results and compared the different approaches. eral methods, machine learning algorithms, lan-
The rest of the paper is structured as follows: guages and sources of information, the authors
In Section 2 we comment on some related stud- conclude that machine learning is the most fre-
ies. The SAD corpus is described in Section 3, quently used method used for mental health de-
and present some interesting statistics. The differ- tection, with Support Vector Machine (SVM) pre-
ent machine learning approaches and the results senting the best results. In addition, the study
obtained are shown in Section 4. Finally, the anal- shows that Twitter is the major data source from
ysis of errors is conducted in Section 5 and con- social networks and English is the main language
clusions are presented in Section 6. studied in the different papers. In (Prieto et al.,
2014) four different health conditions are classi-
2 Related Work fied using machine learning methods over a cor-
pus of tweets extracted by applying a set of crafted
The detection of mental health issues using tex- regular expressions. They integrate some relevant
tual information is a recent task mainly inspired by features in order to improve the final system. In
the massive of use and access to social networks. addition, this is one of the few papers which center
People have become accustomed to using social on languages other than English. Specifically, the
networks to express all kinds of opinions, feelings authors work on Spanish and Portuguese tweets
and emotions. This valuable information can be and the results indicate that the approach is a fea-
captured and treated by an automatic system to sible option for tracking health information on so-
learn how people with some health problems use cial networks.
language to express the frustration, depression or Regarding eating disorders, we can also find
bad feelings. In this way, NLP can help to build some recent studies. For example, (De Choud-
systems to detect early on health problems such hury, 2015) focuses on detecting anorexia on the
as eating disorders, depression or suicidal tenden- social network Tumblr using different affective,
cies. social, cognitive, and linguistic features. They
Although this task is relatively new, some chal- also analyze the clinical implications of detecting
lenging workshops and shared tasks related to the anorexia related content on social media. (Chan-
detection of health conditions have been proposed cellor et al., 2016a) use Instagram in order to study
in recent years. For example, Social Media Min- the eating disorders community and propose a sta-
ing for Health Applications (SMM4H) is a work- tistical model combining topic modeling and clini-
shop and share task that has been held since 2016 cal annotations. Finally, (Wang et al., 2017) center
(Sarker et al., 2016) and continues every year. on Twitter generating a corpus by collecting eat-
The main goal is to attract researchers interested ing disorders and non-eating disorders data. Then
in automatic methods for the collection, extrac-
tion, representation, analysis, and validation of so- 1
http://www.clef-initiative.eu/
656
they train a SVM classifier, obtaining promising our method can easily be adapted to other lan-
results. The high performance achieved suggests guages since the Twitter API allows specification
that it is feasible to design automatic text analy- of the language of the posts retrieved.
sis tools that give early warnings of signs of eating In order to obtain enough tweets, we had to
disorders. However, this study only focuses on En- download messages from past years, more con-
glish and it is important to prove that the systems cretely, in a date range of February 2014 to March
can also be applied to other languages. Thus, in 2019.
this paper we create a Spanish corpus from Twit- To make the corpus more interesting, we used
ter with information concerning of anorexia and as a query different hashtags related to food, nutri-
non-anorexia data. Then we apply several ma- tion, diet and healthy living in a converse way to
chine learning algorithms in order to demonstrate anorexia. We collected data referring to anorexia
the feasibility of implementing systems to auto- using as query the hashtag #anaymia on Twitter.
matically detect sings of anorexia in Spanish mes- In addition, we collected three sets of reference
sages written on social networks. data as negative samples using the hashtag #real-
food #comidareal and #fitness.
3 SAD Corpus
Label 1 (anorexia) has been assigned to tweets
Anorexia and bulimia are two of the most worri- that satisfy the query #anaymia, label 0 (control)
some eating disorders, affecting adolescents and for the other cases. Different messages are shown
young people the most. ”Ana y mia” are the names in Table 1 and in Table 2 we can see the English
used on the web pages that promote anorexia and translation.
bulimia to identify themselves. ”Ana” is anorexia
and ”mia” is bulimia. But it is not a recent phe-
nomenon, it began to become popular on the In-
Twitter
ternet in 2004 (Campos Rodrı́guez, 2007). Today,
they have millions more pages and loyal followers,
and the Internet has connected thousands of peo-
ple with eating disorders. For this reason there are
currently several studies of this disease (Moess-
#anaymia #comidareal #fitness #realfood
(4000 tweets) (4020 tweets) (4009 tweets) (4000 tweets)
657
Tweet Label
Solo quiero llegar a mi casa a comer csm, no puedo más con esta hambre. Pero el hambre
1
es belleza entrando a tu cuerpo.
La comida de hoy es ligera, muy ligera. Alcachofas al horno, simplemente llevan ajo,
aceite, perejil y sal. Mmmm. #masendocrino #alcachofas #dietasana #dietamediterranea 0
#aove #aceitedeoliva #hungry #cocinaespañola #comidacasera #foodpic #banquetesv
Tweet Tag
I just want to get home to eat, I can no longer cope with this hunger. But hunger is beauty
1
coming into your body.
Today’s food is light, very light. Baked artichokes, simply with garlic, oil, parsley and
salt. Mmmm. #masendocrino #alcachofas #dietasana #dietamediterranea #aove #oaceit- 0
edeoliva #hungry #cocinaespañola #comidacasera #foodpic #banquetesv
3.2 Data Filtering tweets being taken from each hashtag (#comi-
Secondly, the extracted data is very noisy, so the dareal, #fitness and #realfood), in this way, the
set requires thorough cleaning before any analysis corpus contains 2707 tweets annotated as positive
can be carried out. The language used by Twitter (anorexia) and 3000 tweets annotated as negative
users contains some attributes that we had to re- (control). Figure 1 shows the number of tweets
move to provide useful information for the classi- downloaded and how the collection decrements at
fication process. This filtering that was performed: each step.
658
Total Anorexia Control
Number of tweet 5707 2707 3000
Number of different users 2585 1120 1466
Number of total words 122798 43179 79619
Number of different words 24635 8761 18515
Average of tweet words 21.52% 15.95% 26.54%
Number of total stop words 30619 13118 17501
Number of different stop words 207 183 185
Average of tweet stop words 5.37% 4.85% 5.83%
library with the module es core news sm3 . spaCy one of the most important steps because it should
is a free open-source library for NLP in Python. help improve the performance of the classifier and
Table 4 shows the statistics obtained, and in it speed up the classification process. Online texts
we can see relevant information on verbs, nouns, usually contain lots a great deal of noise and un-
adjectives and adverbs. We found special interest informative parts which increases the dimension-
in the high number of verbs and nouns used in an- ality of the problem and hence makes the classifi-
notated tweets without anorexia. cation more difficult. For this reason, we applied
We wanted to obtain some statistics about the pre-processing techniques in order to prepare the
mood of users and how they express themselves data for the text classification. In particular, we
through social networks. To obtain this informa- preprocessed the tweets of the SAD Dataset fol-
tion we used the resource iSOL (Molina-González lowing these steps: The tweets were tokenized us-
et al., 2013). This resource has a list of opinion ing NLTK TweetTokenizer4 and all hashtags were
indicator words in Spanish independent of the do- removed.
main. The list consists of 2,509 positive words and Features in the context of text classification are
5,626 negative. The results are described in Table the words, terms or phrases that express the opin-
5. This table shows that users with anorexia prob- ion of the author. They have a higher impact on
lems use more negative language than users with- the orientation of the text. There are several ways
out anorexia. The same happens in the opposite to assess the importance of each feature by attach-
case, whereby the tweets annotated as controlled ing a certain weight to it in the text. We use the
are written with more positive words. most popular: The Term Frequency Inverse Doc-
Finally, Table 6 shows some data about the use ument Frequency scheme (TF-IDF). Specifically,
of hashtags in the messages collected. We can using this scheme each tweet is represented as a
observe that the number of hashtags used in con- vector of unigrams.
trolled tweets is much higher than on the contrary,
and consequently there is also more variety of 4.2 Machine Learning Algorithms
hashtags in messages annotated without anorexia. Machine learning techniques are popular in the bi-
nary classification. For this reason, we decide to
4 Experiments and Results employ different machine learning algorithms in
In this section, we describe different experiments order to classify the corpus in anorexic and non
we carried out to test the validity of the SAD cor- anorexic tweets. The algorithms are Support Vec-
pus. In particular, we trained several classifiers tor Machine (SVM), Naive Bayes (NB), Random
based on machine learning. Forest (RF), Multilayer Perceptron (MLP), Logis-
tic Regression (LR) and Decision Tree (DT).
4.1 Pre-Processing
4.3 Results
Pre-processing the data is the process of clean-
ing and preparing the text for classification. It is In this subsection, we report and discuss the per-
3
formances of our systems on the Spanish anorexic
https://github.com/explosion/
4
spacy-models/releases/tag/es_core_news_ https://www.nltk.org/api/nltk.
sm-2.1.0 tokenize.html
659
Total Anorexia Control
Adjectives in corpus 15332 3996 11336
Nouns in corpus 28594 8536 20058
Verbs in corpus 13647 5592 8055
Adverbs in corpus 5326 2518 2808
Number of different adjectives in corpus 4786 1493 3638
Number of different nouns in corpus 7326 2769 5449
Number of different verbs in corpus 4990 2342 3256
Number of different adverbs in corpus 622 296 455
Average adjectives in tweet 2.69% 1.48% 3.78%
Average nouns in tweet 5.01% 3.15% 6.69%
Average verb in tweet 2.39% 2.07% 2.69%
Average adverbs in tweet 0.93% 0.93% 0.94%
660
Total Anorexia Control
Hashtags in corpus 25133 5037 20096
Different hashtags in corpus 7479 1282 6341
Average hashtags in tweet 4.40% 1.86% 6.70%
Table 7: Results obtained by different classifiers on the SAD corpus (10-fold cross validation).
fused when, for example, the user talks about sport classification systems do not work properly. Our
in general. On the other hand, if we focus on false next goal will be to apply other techniques (such
negatives, we see that one of the problems is the as irony detection or sentiment analysis) in cases
irony in the tweet and another of the problems is where textual information is poor or where rhetor-
when the user is transmitting a negative emotion ical figures such as irony and sarcasm are used.
but does not say the cause.
Acknowledgments
This work has been partially supported by Fondo
Europeo de Desarrollo Regional (FEDER),
LIVING-LANG project (RTI2018-094653-B-
C21) and REDES project (TIN2015-65136-C2-1-
R) from the Spanish Government.
References
Jon Arcelus, Alex J Mitchell, Jackie Wales, and Søren
Nielsen. 2011. Mortality rates in patients with
anorexia nervosa and other eating disorders: a meta-
analysis of 36 studies. Archives of general psychia-
try 68(7):724–731.
Figure 2: Confusion matrix.
Ghelmar Astoveza, Randolph Jay P Obias, Roi Jed L
Palcon, Ramon L Rodriguez, Bernie S Fabito, and
Manolito V Octaviano. 2018. Suicidal behavior de-
6 Conclusion tection on twitter using neural network. In TEN-
CON 2018-2018 IEEE Region 10 Conference. IEEE,
This article presents a new corpus in Spanish for pages 0657–0662.
detecting anorexia in social network messages.
Belén G Bermejo, Luis Ángel Saúl, and Cristina Je-
Several systems are also developed to test the per-
naro. 2011. La anorexia y la bulimia en la red. ana
formance of this task with different classifiers. y mia dos malas compañı́as para las jóvenes de hoy
The results obtained show that the performance [the anorexia and bulimia on the web: Ana and mia
is very similar in all systems, although SVM and two “bad company” for youth today]. Acción psi-
MLP are the only ones that obtain accuracy above cológica 8(1):71–84.
0.9. José Miguel Campos Rodrı́guez. 2007. Anorexia, bu-
Error analysis reveals that there are cases where limia e internet. aproximación al fenómeno pro-ana
661
Tweet True label Predicted
”El fı́sico no importa” 1 0
Time to listen to music to relax and forget about the shitty world I live in 1 0
Hello @IKEASpain my 9 month old baby eats solid food, and has never
0 1
eaten a baby food and I think it’s unfair that you charge him for the food
Buttock Exercise Routine 0 1
y mı́a desde la teorı́a subcultural. Frenia. Revista de uation Forum for European Languages. Springer,
Historia de la Psiquiatrı́a 7(1):127–144. pages 346–360.
Stevie Chancellor, Zhiyuan Lin, Erica L Goodman, Markus Moessner, Johannes Feldhege, Markus Wolf,
Stephanie Zerwas, and Munmun De Choudhury. and Stephanie Bauer. 2018. Analyzing big data in
2016a. Quantifying and predicting mental illness social media: Text and network analyses of an eat-
severity in online pro-eating disorder communities. ing disorder forum. International Journal of Eating
In Proceedings of the 19th ACM Conference on Disorders 51(7):656–667.
Computer-Supported Cooperative Work & Social
Computing. ACM, pages 1171–1184. M Dolores Molina-González, Eugenio Martı́nez-
Cámara, Marı́a-Teresa Martı́n-Valdivia, and José M
Stevie Chancellor, Jessica Annette Pater, Trustin Clear, Perea-Ortega. 2013. Semantic orientation for polar-
Eric Gilbert, and Munmun De Choudhury. 2016b. ity classification in spanish reviews. Expert Systems
# thyghgapp: Instagram content moderation and with Applications 40(18):7250–7257.
lexical variation in pro-eating disorder communi-
ties. In Proceedings of the 19th ACM Conference Bridianne O’Dea, Stephen Wan, Philip J Batterham,
on Computer-Supported Cooperative Work & Social Alison L Calear, Cecile Paris, and Helen Chris-
Computing. ACM, pages 1201–1213. tensen. 2015. Detecting suicidality on twitter. In-
ternet Interventions 2:183–188.
Munmun De Choudhury. 2015. Anorexia on tumblr: A
characterization study. In Proceedings of the 5th in- F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
ternational conference on digital health 2015. ACM, B. Thirion, O. Grisel, M. Blondel, P. Pretten-
pages 43–50. hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Pas-
sos, D. Cournapeau, M. Brucher, M. Perrot, and
Huijie Lin, Jia Jia, Jiezhong Qiu, Yongfeng Zhang, E. Duchesnay. 2011. Scikit-learn: Machine learning
Guangyao Shen, Lexing Xie, Jie Tang, Ling Feng, in Python. Journal of Machine Learning Research
and Tat-Seng Chua. 2017. Detecting stress based 12:2825–2830.
on social interactions in social networks. IEEE
Transactions on Knowledge and Data Engineering Vı́ctor M Prieto, Sérgio Matos, Manuel Álvarez, Fidel
29(9):1820–1833. Cacheda, and José Luı́s Oliveira. 2014. Twitter: A
good place to detect health conditions. PLoS ONE
David E Losada, Fabio Crestani, and Javier Parapar. 9:e86191.
2017. erisk 2017: Clef lab on early risk prediction
on the internet: experimental foundations. In Inter- Rohizah Abd Rahman, Khairuddin Omar, Shahrul Az-
national Conference of the Cross-Language Eval- man Mohd Noah, and Mohd Shahrul Nizam Mohd
662
Danuri. 2018. A survey on mental health detec-
tion in online social network. International Journal
on Advanced Science, Engineering and Information
Technology 8(4-2):1431–1436.
Abeed Sarker, Azadeh Nikfarjam, and Graciela Gon-
zalez. 2016. Social media mining shared task work-
shop. In Biocomputing 2016: Proceedings of the
Pacific Symposium. World Scientific, pages 581–
592.
Judy Hanwen Shen and Frank Rudzicz. 2017. De-
tecting anxiety through reddit. In Proceedings of
the Fourth Workshop on Computational Linguistics
and Clinical Psychology—From Linguistic Signal to
Clinical Reality. pages 58–65.
Mike Thelwall. 2017. Tensistrength: Stress and re-
laxation magnitude detection for social media texts.
Information Processing & Management 53(1):106–
121.
Sho Tsugawa, Yusuke Kikuchi, Fumio Kishino, Ko-
suke Nakajima, Yuichi Itoh, and Hiroyuki Ohsaki.
2015. Recognizing depression from twitter activ-
ity. In Proceedings of the 33rd annual ACM confer-
ence on human factors in computing systems. ACM,
pages 3187–3196.
Theo Vos, Ryan M Barber, Brad Bell, Amelia Bertozzi-
Villa, Stan Biryukov, Ian Bolliger, Fiona Charlson,
Adrian Davis, Louisa Degenhardt, Daniel Dicker,
et al. 2015. Global, regional, and national inci-
dence, prevalence, and years lived with disability for
301 acute and chronic diseases and injuries in 188
countries, 1990–2013: a systematic analysis for the
global burden of disease study 2013. The Lancet
386(9995):743–800.
Tao Wang, Markus Brede, Antonella Ianni, and Em-
manouil Mentzakis. 2017. Detecting and charac-
terizing eating-disorder communities on social me-
dia. In Proceedings of the Tenth ACM International
Conference on Web Search and Data Mining. ACM,
pages 91–100.
663
A Type-Theoretical Reduction of Morphological, Syntactic and Semantic
Compositionality to a Single Level of Description
Erkki Luuk
Institute of Computer Science, University of Tartu, J. Liivi 2, Tartu 50409, Estonia
erkkil@gmail.com
664
Proceedings of Recent Advances in Natural Language Processing, pages 664–673,
Varna, Bulgaria, Sep 2–4 2019.
is available. We can model a (reasonably large) 3 A Note on Selectional Restrictions
fragment of NL in a suitable type system to make
More or less overlooked in Montagovian (Mon-
it (fail to) type check in exactly the same ways
tague, 2002) and categorial (Lambek, 1958) tra-
as (hypothetical) NL expressions pass (and fail)
ditions, selectional restrictions have recently be-
acceptability tests. Call this principle “the cor-
come a focus of intense research in modern type
respondence criterion”. Arguably, this possibil-
theories (Asher, 2014; Luo, 2010; Bekki and
ity, the ultimate test for any type-driven linguistic
Asher, 2013). The essence of the semantic (and
theory, has not been fully explored, while several
perhaps even more logical than linguistic) phe-
significant steps in this direction have been taken.
nomenon of selectional restrictions is prescribing
Chatzikyriakidis and Luo (2014b, 2016) describe
types for a relation’s arguments.
(among other things) a use of proof assistants for
There is an important difference between (1) ar-
testing the logical soundness of linguistic theo-
guments belonging to types and (2) relations im-
ries, while Grammatical Framework (GF, Ranta
posing types on their arguments. While an argu-
(2004)), a statically typed programming language
ment can clearly belong to different types3 , a re-
for writing NL grammars, gets closest to a type-
lation should not impose different types on its kth
theoretical implementation of NL. However, GF
argument, for a fixed k. Modeling selectional re-
is a high level tool, mathematically opaque to
strictions along with grammar is an important mo-
the end user, and quite specialized. Because it
tivation for lump types, described below (in sec-
is geared towards writing NL grammars, it does
tion 5.1.1).
not offer a selection of different mathematical for-
malisms to work with, being thus unsuitable for 4 Introducing NLC
a general (low level) modeling of NL and theo-
ries thereof. In addition, it does not concern itself Call the formalism we are considering NLC. It is a
with modeling compositional semantics, although type system for modeling NL syntax, morphology
a FrameNet API for GF has been proposed and and compositional semantics — briefly, composi-
partly implemented (Gruzitis et al., 2012; Gruzitis tionality in NL. The basic unit of NLC is a func-
and Dannélls, 2017). tion of a small (usually ≤ 3) arity. The expressions
A powerful feature of typed theories (especially of NLC are functions, function applications, func-
of the richly typed2 ones — Chatzikyriakidis and tion types (Π-types), lump types, terms of lump
Luo (2014b)) is that they allow to capture not types, and universes (types of types). Elementary
only grammatical but also semantic acceptabil- terms of NLC are functions. This is possible if
ity. The paper shows that a combination of func- we interpret “proper arguments” as nullary func-
tions and rich typing makes it possible to use a tions (functions that take no arguments). So func-
single type system for modeling the core of NL tions are divided into proper arguments and proper
morphology, syntax and compositional semantics, functions (the latter being functions that take argu-
thus questioning the soundness of the theoretical ments). For a fixed NL k, let T k be a proper type
distinction between these different NL “layers” variable of NLC, where “proper type” refers to a
(and partly eradicating their even more theoreti- type that is not a universe, and Mk the set of mor-
cal “connections” such as the syntax-semantics in- phemes4 . Then we have the rule:
terface). Because of a wide selection of mathe- a ∈ Mk
matical formalisms in a richly typed setting, com- ATV-Intro,
a : Tk
bined with a relatively “instant” compile time type
checking, proof assistants (e.g. Coq, Agda, Lean) for generating atomic terms of NLC and introduc-
are suitable tools for such work (cf. Chatzikyr- ing type variables for them. The rule ATV-Intro
iakidis and Luo (2016)). As shown below, im- says that all morphemes have types in NLC (tech-
plementing polymorphic functions with a subclass nically: “if a is a morpheme of language k then a
of compound types (called lump types) allows to has type T k ”).
partly collapse different “levels” of NL (morphol- Some morphemes (e.g. stems) occur only in
ogy, syntax, and compositional semantics). an argument position (i.e. are proper arguments),
3
E.g. a book is a physical and informational object.
2 4
Rich typing coincides mainly (although perhaps not ex- Morphemes are smallest signs (form-meaning corre-
clusively) with dependent and/or polymorphic types. spondences) in language.
665
while others (e.g. plural markers) are proper func- A. Thus, we have same-type-reference without
tions. More generally, a function or function ap- self-reference.
plication is a parsimonious interpretation of mor-
phemes, words, phrases and sentences in NL. Sen- 5.1 Polymorphism and Lump Types
tences, multimorphemic words and phrases are The complexity of NLC goes well beyond regu-
function applications. This amounts to a rigor- lar function types. In considering a NL expression
ous interpretation of the more general “principle type-theoretically, one is frequently inclined to as-
of compositionality”, as it is known at least since sign it to more than one type. Confining our anal-
Frege5 . Broadly speaking, there are only two ways ysis to only the linguistically relevant features, we
to explain the emergence of compositional mean- may want to type e.g. stone as a flexible, physi-
ing: by specifying a relation together with its (a) cal object, word in nominative case, etc. A way
arguments or (b) type. The first corresponds to — corresponding to polymorphism — to go about
e.g. function application and the second to func- this is to define coercions to (i.e. coercive subtyp-
tion declaration. ing for) all the types we need. In fact, we have
In a functional type system, generating complex three possibilities: either we (1) lose some type
terms and introducing type variables for them is information, type stone (2) polymorphically or (3)
straightforward (rule CTV-Intro, with Tik ranging with a lump type.
over proper types): As the nominal and verbal readings of stone pre-
clude each other, a polymorphism is required if
we want to encode them both (cf. footnote 7).
e1 : T1k , ..., em : Tm
k a : T1k → ... → Tm
k → Tk
m+1 In many other cases, however, a lump type may
k
a(e1 , ..., em ) : Tm+1 be preferred. Thus, NLC features types for mor-
phemes, function types, lump types and polymor-
where a(e1 , ..., em ) is an application and T1k → phism.
k
... → Tm+1 the usual (right-associative) function
type. Since CTV-Intro is the standard function 5.1.1 Lump Types
type elimination rule, the function type introduc- While possibility (3) is new, the superclass of
tion rule is derivable from CTV-Intro. lump types, compound types have been used for
NL modeling in the form of multi-field record
5 NLC: The Types types (Cooper, 2005; Ranta, 2004; Luo, 2011;
Chatzikyriakidis and Luo, 2014a). Also, some
Since grammatical (and semantic) categories have kind of polymorphism (e.g. by subtyping —
a limited, finite number of members, we need Luo (2010)) is frequently thought to be necessary.
some atomic types with limited membership. Let However, the use of compound types has been so
U denote the top-level universe of NLC. We use far confined to record types only, i.e. not properly
axioms of the form S : U and T : S, where S may generalized6 . A compound type is a type which
be a universe, for introducing atomic type con- is a syntactic compound of multiple types or their
stants corresponding to linguistic categories like terms. Normally, the compounded types are differ-
stem, case, nominative, noun, verb, etc. For proper ent; in the degenerate case, they are the same. Ex-
functions, we need function types (Π-types). Be- amples (or implementations) of compound types
sides this, we need only polymorphism and lump include Σ-, Π-, Cartesian product and multi-field
types, both of which can be (in various ways) im- record types.
plemented with function types.
We defined types as “categories of semantic
Since a term of type A may contain another value” but, as the example of stone shows, for NL
term of type A (or in case of a function type, expressions the value covers not only linguistic
take another term of the type as an argument), we semantics but also the meanings of syntactic and
have sufficient complexity without recursion (self-
6
reference, which we do not need). For example, In many (most?) programming languages that support
them, the notion of “compound type” (or “compound data
a sentence A containing another sentence B does type”) is synonymous with a multi-field record type. This is
not imply recursion unless A = B or B references not the way it is used here. While a record type can be defined
as a (mathematically more fundamental) Σ-, Π- or Cartesian
5
The principle is more general because it holds also for product type (Constable, 2003), I have never seen it defined
interpretations of formal languages. as a function application.
666
morphological categories (we will return to this The linguistic categories not formalized in
point in section 9). As compared to (1) and (2), Supplement B are gerunds, participles, auxiliary
packing an expression’s meaning into a lump type verbs, interrogatives, numerals, negation, mass/-
allows to do away with both the loss of informa- count distinction and unspecified selectional re-
tion and typing complexity. Of course, the lump strictions (and possibly others). These are omit-
type itself will be complex but this will, hopefully, ted not because of a special difficulty formalizing
present less problems than alternatives (1) and (2). them would pose but because the formalized frag-
As a bonus, the underlying linguistic model will ment is sufficiently expressive (and representative
simplify on account of reducing compositional se- of NL) to make the points of utility and feasibility
mantics and parts of morphology and syntax to a of NL formalization with the combination of com-
single level of description. Below is the rule for pound and polymorphic types. The formalization
lump type introduction (LT-Intro): has been done in the proof assistant Coq (ver. 8.9),
making use of its features like Ltac programming,
B : Tk c0 : C0k , ..., cn+1 : Cn+1
k
B 7→ c0 , ..., B 7→ cn+1
custom notations, etc. Besides showing the use
of lump (viz. application) types in NL modeling,
B: C0k ..Cn+1
k
Supplement B should fulfill the abovementioned
where T k is a proper type variable, B a term “correspondence criterion”.
constant and C0k , ..., Cn+1
k type constants in NLC,
5.1.2 Polymorphism
x 7→ y a function interpreting x as y 7 , and
C0k ..Cn+1
k the notation for a lump type (compris- Besides lump types, there is some use for poly-
ing types C0k through Cn+1 k , i.e. there must be at morphism as well — if not for any other reason,
least two). LT-Intro is formalism-agnostic — the then because NL expressions may be underspeci-
exact mathematical structure used for lumping is fied. E.g. sleep and stone are flexible stems that
irrelevant. In particular, as shown in Supplement can function both as nouns and verbs. As a verb,
A8 , we can implement lump types for NL as 1) sleep selects for a specific argument, say, a sen-
record types, 2) function applications, 3) Cartesian tient entity (only higher animals can sleep — for
product types, or 4) Π-types. In languages that trees, stones and bacteria it is not an option). As
have them (e.g. TypeScript, Flow...), it is natural a noun, it is quite similar to many others: a stem,
to encode lump types as intersection types. Lump a flexible in singular, an event, etc. Since verbs
types are defined as compound types that satisfy are functions, sleep’s type must be a function type,
LT-Intro (i.e. we are not interested in empty lump but since it also functions as a noun, a polymor-
types). phism is desirable. The alternative, defining two
Supplement B9 proceeds to formalize a diverse distinct sleeps, one noun and one verb, would be
fragment of NL with function applications. The redundant and inelegant — esp. in a program-
fragment comprises stems, nouns, verbs, flexi- ming language, where (differently from NL) they
bles, proper names, pronouns, XPs10 , adjectives, would have to be formally distinct (e.g. sleep and
sentential, adjectival and generic adverbs, deter- sleep0) even in the absence of any discriminating
667
ber of terms, while in NL the number is fixed, very 7 Comparisons
limited, and usually known in advance11 . The only
counterexamples to this rule are phrases, clauses In section 2 we compared NLC with related typed
and complex words. So the additional layer of approaches. This section takes a broader (al-
formalization is used for downgrading the over- beit still related) perspective, comparing NLC
powerful simple types to something on which type with Combinatory Categorial Grammar and Head-
inference would work. In Coq, the main devices Driven Phrase Structure Grammar.
for such downgrading are inductive types, “canon- 7.1 Combinatory Categorial Grammar
ical structures”, and type classes. Our formaliza-
tion uses them all, relying most heavily on canon- A feature of NLC, Combinatory Categorial Gram-
ical structures (essentially, canonical records of a mar (CCG) (Steedman, 2000) and some other cat-
record type). egorial formalisms is that they tend to do away
with syntax: “...syntactic structure is merely the
6 Truth-Functionality characterization of the process of constructing a
logical form, rather than a representational level...”
So far, the semantics developed here is not truth-
— Steedman (2000), p. xi12 . The main differ-
functional, i.e. it is type- but not model-theoretic.
ences are as follows. CCG is a categorial formal-
As type theory is ‘semantic’ by definition, it is
ism, NLC not. CCG has complex (combinatorial)
clearly sufficient for a semantical cast of seman-
syntactic types of the form X\Y and X/Y in-
tics without a recourse to model theory. This is ev-
stead of lump types with (what is conventionally
ident in programming language semantics, where
viewed as) morphological, syntactic and semantic
the role of model theory is marginal as compared
information. Another difference is CCG’s (by de-
to that of type theory. Traditionally, in natural lan-
fault limited) support for word order. (In particu-
guage semantics the opposite is true, as sentential
lar, CCG does not handle concatenations (of terms
semantics is usually construed as model-theoretic
of types) c(X/Z, Y, Z) and c(Z, Y, X\Z), where
even in a type-theoretical setting (e.g. Chatzikyr-
Y is nonempty13 .) Also, in CCG, NL construc-
iakidis and Luo (2015, 2016)). The obvious rea-
tions of sentence level and below can have multi-
son for this is that the (prevailing, i.e. Montago-
ple structures independently of (what is tradition-
vian) tradition is model-theoretic. For this reason,
ally called) interpretational ambiguity.
an optional truth-functionality module has been
added to the implementation. Degenerate models 7.2 Head-Driven Phrase Structure Grammar
(where all NLC sentences (S) are uniformly true,
It is also useful to compare NLC with a more ded-
false or undecidable) can be specified trivially by
icated syntactic formalism such as Head-Driven
subtyping, e.g.
Phrase Structure Grammar (HPSG). As a mature
Parameter s_prop:> S -> Prop.
(* all S-s undecidable *)
formalism that has been implemented for several
languages (Pollard and Sag, 1994; DELPH-IN,
and with only a little effort non-trivial models can 2019), HPSG is currently implementation-wise
be specified, too (with subtyping and a special no- much superior to NLC (which has been imple-
tation matching NLC constructions with appropri- mented only for a fragment of English14 ), so it
ate values). Below is an example from Supplement is appropriate to compare only formalisms. We
B: start with similarities. Both formalisms model
Check $(PRES walk john): Prop. (*False*) parts of semantics, syntax and morphology, but
Fail Check $(PAST walk stone). (*type mismatch*) HPSG’s scope is much wider, as it covers also
Check $(PRES sleep john). (*True*)
Check $(PRES sleep (PL boy)). (*False*) 12
However, as described below, CCG still features syntac-
Check $(PAST sleep (-s boy)). (*True*)
tic types.
Fail Check $sleep. (*type inference fail*) 13
I am not sure whether such concatenations exist in NLs
with fixed word order, but a decision to rule them out by de-
(*a trivial proof that "boys
fault is arbitrary. However, perhaps it would be feasible to
don’t sleep" and "john sleeps"*)
introduce a special rule for accommodating Y in this case.
Theorem pres: (∼ ($ (PRES sleep (-s boy)))) /\ 14
($ (PRES sleep john)). Proof. firstorder. Qed. Structurally, the fragment is quite universal. In fact, with
a slightly more general notation one can approximate a Uni-
versal Grammar (a statement that will make more sense after
11
The latter will depend on your linguistic theory, as differ- reading section 9 and recalling that NL semantics is univer-
ent theories posit different categories and members for them. sal).
668
lexicon and word and morpheme orders. Both (* ..(also w/ optional arguments omitted): *)
formalisms are compositional in that the mean- Check and (PAST walk john) (PAST sleep john).
My experience of implementing NLC is quite lim- (* We can stack adverbs and adjectives, and
ited, as I have so far tried to implement it only in use adjectival and adverbial connectives: *)
Check all ((and madly madly) red (red
one programming language and have implemented
[and blue limbed [-s john]])). (* All madly
at best a half of NL in terms of its general (or ty- and madly red, red, blue and limbed Johns *)
pological) category structure. Below is a test of
In Coq, _ is a placeholder for any admissible term
an implementation of NLC. The test is by type-
or type. A switch in the file the code is taken from
checking possible NL(C) expressions. The code
allows to choose between Strict and Lax modes,
(from Supplement B) is generously commented
respecting and ignoring selectional restrictions, re-
and should be self-explanatory.
spectively. The notation [...] interfaces with the
Check PAST throw john. (* "John threw"
type checks -- but not as a sentence: *)
current mode. The example omits all technical de-
Fail Check PAST throw john: S. (* "At the hut" tails like type definitions, etc. These are not instru-
can be the 3rd argument of "throw": *) mental to NLC, as the type system — i.e. one cap-
Check PAST throw john (-s stone)
(at (the hut)). (*..but not the 2nd one:*) turing the morphological, syntantic and semantic
Fail Check PAST throw john (at (the hut)). compositionality of NL with lump types as faith-
(* "In a hut" cannot be fully as possible — can be implemented in sev-
an argument of "throw": *)
Fail Check PAST throw john eral ways (cf. Supplement A) and different pro-
(-s stone) (in (a hut)). gramming languages. The implementation uses
(* ..but can be a sentence modifier: *) only a tiny subset of Coq’s features, and its main
Check in (a hut) (PAST throw john (-s stone)).
(* ..and so can "at every hut" and functionality, theorem proving, is entirely optional
sentential adverbs like "however": *) here. As I am not at all convinced that Coq is
Check at (every hut) (PAST throw john
(-s stone)): S.
the best language for implementing NLC, I en-
Check however (PAST throw john (-s stone)): S. courage the interested reader to experiment with a
(* Connectives cannot range over a programming language of their choice. That being
sentential and nominal argument: *)
Fail Check and (PAST throw john (a stone))
said, statically typed programming languages with
john. (* ..but can range over nominal: *) a sufficiently complex type system and advanced
Check and (every john) (all (the (-s boy))). type inference have some advantages for this kind
(* ..or sentential arguments: *)
Check and (PAST walk (-s boy) of work (e.g. in terms of rigor and the similarity
(to (all (-s hut)))) (PAST sleep john). of implemented formulas to NL expressions).
669
9 Implications and morphology, respectively; the compositional-
ity of words, phrases, morphemes and clauses per-
The driving force behind NLC has been to corre-
tains to semantics. As a desirable consequence,
spond to NL as closely as possible. Since ontology
we could continue using the existing general the-
(or world knowledge) interfaces with the compo-
ory of natural language with only a few termino-
sitional semantics of NL, it is desirable to formal-
logical changes. But (how) would (‡) be viable?
ize some of it in the form of selectional restric-
A possible justification would make at least two
tions. We have collapsed syntactic, morphological
arguments. First, from the theoretical side, mor-
and semantic compositionality to a single level —
phological, syntactic and semantic composition-
to that of the type system. In effect, some types
ality all refer to certain (parts of) knowledge —
have become syntactic, but the syntax has only two
namely, about morphology, syntax and world, re-
rules: functionality (CTV-Intro) and lumping (LT-
spectively. The only way to have knowledge is
Intro). In sum, the paper (and the underlying for-
by way of meaning, which, given the above, is
malization) have shown that:
clearly linguistic. This consideration roots our
(†) A feature of natural language — viz. morpho- enterprise in linguistic semantics. Second, from
logical, syntactic and semantic compositional- the formalization side, we are using type theory,
ity — can be reduced to a single level of de- which is a theory of semantics (broadly defined15
scription. — cf. (i)). This argument formally corroborates
the claim that NLC models only natural language
It is not clear what (†) means, so let us try to ex- semantics.
plore it further, by (temporarily) assuming that (†) Of course, the fact that NL can be modeled this
posits a new level of description — call it compo- way does not entail that this is the way it works in
sitionality — which, moreover, would have to in- the brain16 . So far, our argument has been solely
terface with lexical semantics and (what is left of) about modeling: It is more parsimonious to model
morphology and syntax. This would be not only compositionality in a functional type system than
theoretically unheard-of (which would be only a e.g. with syntax trees or phrase structure rewrite
mild objection) but would have the undesirable rules, since the latter cannot, neither separately
consequence of complicating the general frame- nor when combined, account for all composition-
work of linguistic theory. However, it would have ality. The only advantage of the rewrite rules and
some positive outcomes as well, namely “elimi- syntax trees over the type-theoretical modeling is
nating” compositional semantics and simplifying that they allow, in principle, to capture word order.
morphology and syntax proper. The general the- However, not all syntactic theories support linear
ory of natural language would become more com- order preserving trees (the Chomskian transforma-
plex while three subtheories (morphology, syntax tional grammar being a case in point — Chom-
and semantics) would simplify. sky (1965, 1981)). Secondly, a word order rule
Depending on one’s outlook on the general the- is, differently from compositionality, not a linguis-
ory of natural language, this might seem like a tic universal (there are many languages with flex-
path worth pursuing. However, below I will argue ible word order — Dryer (2013)). Incidentally,
that it is not the only one. The alternative would this also means that not all natural languages have
be to assume that: syntax.
One thing that seems to emerge from the liter-
(‡) There’s nothing “morphological” or “syntac- ature on language processing is the role of syntax
tic” about morphological and syntactic com- as guiding semantic interpretation, or (more figu-
positionality — it is all just semantic compo- ratively) serving semantics (Kempson et al., 2001;
sitionality. Morrill, 2010; Christiansen and Chater, 2016).
Some authors have explicitly argued against syn-
Clearly, (†) and (‡) are not mutually exclusive
— in fact, (‡) is just a more radical version of 15
Historically, the semantics of mathematics, more re-
(†) (and incidentally, also subsumes (†)). (‡) just cently also the semantics of programming languages.
16
conflates the hypothetical new level of descrip- Incidentally, there is little sense in trying to make a case
of “how language works in the brain”, as there is no con-
tion of (†) with compositional semantics. Word sensus on this among psycho- and neurolinguists (Chater and
and (subword) morpheme order pertain to syntax Christiansen, 2016).
670
tax as a separate representational level of linguis- NLC is generic in at least two respects: 1. It is
tic structure (Pulman, 1985; Steedman, 2000). To- applicable to all NLs, and 2. It allows for generic
pographical patterns of brain activation to nouns modeling of morphological, syntactic and seman-
and verbs are driven not by lexical (grammatical) tic compositionality. Besides functions, applica-
class but by semantics and word meaning (Mose- tions and their types, NLC features polymorphic
ley and Pulvermüller, 2014). A cognitive architec- and lump types. The latter are compound types
ture with a multi-level (syntactic, semantic, mor- satisfying LT-Intro. Compound types are types
phological, etc.) NL processing usually requires which are syntactic compounds of multiple types
positing at least as many memory buffers for it or their terms (Σ-, Π- and Cartesian product types
(Levinson, 2016), while our short-term memory are examples of compound types). At its core,
(obviously recruited in e.g. dialogues) is very lim- NLC is a simple system for an integrated model-
ited (Cowan, 2001). These pieces of evidence ing of the morphological, syntactic and semantic
from language studies also corroborate (‡). compositionality of NL with lump types.
In sum, we hypothesize that combinatorial The paper also presents an implementation of
(im)possibilities in syntax and morphology are NLC in Coq (which, unfortunately, is not quite
better analyzed as belonging to the domain of as simple, which may be Coq’s fault). The main
compositional semantics. Moreover, the conjec- goal of the implementation was to formalize a rea-
ture that the traditional boundaries between the sonably diverse fragment of NL in NLC, with for-
levels of description reflect more of a sociologi- malized NLC expressions type checking and fail-
cal (a division of labor among linguists) than a lin- ing to type check in exactly the same ways that
guistic fact is not too bold. NL expressions pass and fail their acceptability
If the hypothesis is correct, nothing remains in tests. Aside from this goal’s feasibility, the im-
syntax except word order17 . Since word order plementation shows several things: (1) the via-
can label (now already semantic) constituents, as bility and simplicity of NLC for modeling NL
in JohnSUB loves M aryOBJ , a limited form of compositionality, (2) the utility of lump and poly-
syntax-semantics interface is also expected. This morphic types in NL modeling, and most impor-
accords with the view of the function of syntax tantly, (3) the possibility of reducing morpholog-
as serving semantics, viz. in interpretation dis- ical, syntactic and semantic compositionality to a
ambiguation, which is the primary function of single level of description. In discussion we have
word order constraints (as witnessed in the exam- tried to identify this level as semantic composi-
ple above). Syntax-semantics interface is also an tionality — an interpretation which, besides be-
appropriate level for phenomena like anaphora and ing supported by results from language process-
ellipsis. Likewise, morphology proper retains only ing (Pulman, 1985; Steedman, 2000; Moseley and
(subword) morpheme order and fusion, while mor- Pulvermüller, 2014), has interesting implications
phophonological (i.e. morphology-phonology in- on NL structure and modeling. In particular, it
terface) phenomena like e.g. sandhi must also be may reduce syntax and morphology to word and
accounted for. As a result, syntax and morphology morpheme orders, respectively (with the syntax-
emerge as by-products of a contingent conforma- semantics and phonology-morphology interfaces
tion to the serial channel over which language is reducing correspondingly), with NL architecture
processed. taking on a rather different look. This has also im-
plications on linguistic typology, as syntax would,
10 Conclusion much like morphology before it (Muansuwan,
2002; Grandi and Montermini, 2005; Klamer,
If we interpret proper arguments as nullary func-
2005), cease to be a logically necessary compo-
tions, all NL expressions up to the sentence level
nent of NL.
(i.e. morphemes, words, phrases, clauses and sen-
tences) can be interpreted as functions and func-
Acknowledgments
tion applications. The paper presents NLC, a
generic functional type system (i.e. one consist- I thank Jason Gross, Hendrik Luuk, Erik Palmgren
ing primarily of functions and their applications). and Enrico Tassi for their advice. This work has
17
If we do not posit lexicon as a separate level, lexical cat- been supported by IUT20-56 and European Re-
egories also pertain to morphology or syntax. gional Development Fund through CEES.
671
References on language. Behavioral and Brain Sciences 39:e62.
https://doi.org/10.1017/S0140525X1500031X.
Nicholas Asher. 2014. Selectional restrictions, types
and categories. Journal of Applied Logic 12(1):75– Robert L. Constable. 2003. Recent results in type
87. https://doi.org/10.1016/j.jal.2013.08.002. theory and their relationship to Automath. In
Fairouz D. Kamareddine, editor, Thirty Five Years
Daisuke Bekki and Nicholas Asher. 2013. Logical pol-
of Automating Mathematics, Springer Netherlands,
ysemy and subtyping. In Yoichi Motomura, Alastair
Dordrecht, pages 37–48.
Butler, and Daisuke Bekki, editors, New Frontiers in
Artificial Intelligence, Springer, Berlin, Heidelberg,
Robin Cooper. 2005. Records and record
pages 17–24.
types in semantic theory. Journal of
Andrew Carnie. 2012. Syntax: A Genera- Logic and Computation 15(2):99–112.
tive Introduction. Wiley-Blackwell, Malden, https://doi.org/10.1093/logcom/exi004.
MA, 3rd edition. https://www.wiley.com/en-
ee/Syntax:+A+Generative+Introduction,+3rd+Edition- Nelson Cowan. 2001. The magical number
p-9780470655313. 4 in short-term memory: a reconsideration
of mental storage capacity. Behavioral and
Nick Chater and Morten H. Christiansen. Brain Sciences 24(1):87–114; discussion 114–85.
2016. Squeezing through the Now-or- https://doi.org/10.1017/S0140525X01003922.
Never bottleneck: reconnecting language
processing, acquisition, change, and struc- DELPH-IN. 2019. The DELPH-IN collaboration. Ac-
ture. Behavioral and Brain Sciences 39:e91. cessed 18.03.2019. http://www.delph-in.net.
https://doi.org/10.1017/S0140525X15001235.
Matthew S. Dryer. 2013. Order of Subject, Object and
Stergios Chatzikyriakidis and Zhaohui Luo. 2014a. Verb. In Matthew S. Dryer and Martin Haspelmath,
Natural language inference in Coq. Journal of editors, The World Atlas of Language Structures On-
Logic, Language and Information 23(4):441–480. line, Max Planck Institute for Evolutionary Anthro-
https://doi.org/10.1007/s10849-014-9208-x. pology, Leipzig. http://wals.info/chapter/81.
Stergios Chatzikyriakidis and Zhaohui Luo. 2014b. Nicola Grandi and Fabio Montermini. 2005. Prefix-
Natural language reasoning using proof-assistant suffix neutrality in evaluative morphology. In Geert
technology: Rich typing and beyond. In Booij, Emiliano Guevara, Angela Ralli, Salvatore
Proceedings of the EACL 2014 Workshop on Sgroi, and Sergio Scalise, editors, On-line Pro-
Type Theory and Natural Language Semantics ceedings of the Fourth Mediterranean Morphol-
(TTNLS). Association for Computational Lin- ogy Meeting (MMM4), Catania, 21-23 Septem-
guistics, Gothenburg, Sweden, pages 37–45. ber 2003. Università degli Studi di Bologna.
http://www.aclweb.org/anthology/W14-1405. https://geertbooij.files.wordpress.com/2014/02/
mmm4-proceedings.pdf.
Stergios Chatzikyriakidis and Zhaohui Luo. 2015.
Individuation criteria, dot-types and copredica- N. Gruzitis and D. Dannélls. 2017. A multilingual
tion: A view from modern type theories. In FrameNet-based grammar and lexicon for controlled
Proceedings of the 14th Meeting on the Math- natural language. Lang Resources & Evaluation
ematics of Language (MoL 2015). Association 51(1):37–66. https://doi.org/10.1007/s10579-015-
for Computational Linguistics, pages 39–50. 9321-8.
https://doi.org/10.3115/v1/W15-2304.
Normunds Gruzitis, Peteris Paikens, and Guntis
Stergios Chatzikyriakidis and Zhaohui Luo. 2016.
Barzdins. 2012. FrameNet resource grammar li-
Proof assistants for natural language seman-
brary for GF. In Tobias Kuhn and Norbert E. Fuchs,
tics. In Maxime Amblard, Philippe de Groote,
editors, Controlled Natural Language, Springer,
Sylvain Pogodalla, and Christian Retoré, ed-
Berlin, Heidelberg, pages 121–137.
itors, Logical Aspects of Computational Lin-
guistics. Celebrating 20 Years of LACL (1996–
Ruth Kempson, Wilfried Meyer-Viol, and Dov Gabbay.
2016). Springer, Berlin, Heidelberg, pages 85–98.
2001. Dynamic Syntax: The Flow of Language Un-
http://www.cs.rhul.ac.uk/ zhaohui/LACL16PA.pdf.
derstanding. Blackwell, Oxford.
Noam Chomsky. 1965. Aspects of the The-
ory of Syntax. The MIT Press, Cambridge. Marian Klamer. 2005. Explaining structural and se-
http://www.amazon.com/Aspects-Theory-Syntax- mantic asymmetries in morphological typology. In
Noam-Chomsky/dp/0262530074. Geert Booij, Emiliano Guevara, Angela Ralli, Sal-
vatore Sgroi, and Sergio Scalise, editors, On-line
Noam Chomsky. 1981. Lectures on Government and Proceedings of the Fourth Mediterranean Morphol-
Binding. Foris, Dordrecht. ogy Meeting (MMM4), Catania, 21-23 Septem-
ber 2003. Università degli Studi di Bologna.
Morten H. Christiansen and Nick Chater. 2016. The https://geertbooij.files.wordpress.com/2014/02/
Now-or-Never bottleneck: a fundamental constraint mmm4-proceedings.pdf.
672
Joachim Lambek. 1958. The mathematics of sentence Aarne Ranta. 2004. Grammatical Framework: a
structure. The American Mathematical Monthly type-theoretical grammar formalism. The Jour-
65(3):154–170. nal of Functional Programming 14(2):145–189.
https://doi.org/10.1017/S0956796803004738.
Stephen C. Levinson. 2016. “Process and per-
ish” or multiple buffers with push-down Mark Steedman. 2000. The Syntactic Process. MIT
stacks? Behavioral and Brain Sciences 39:e81. Press, Cambridge, MA, USA.
https://doi.org/10.1017/S0140525X15000862.
Zhaohui Luo. 2010. Type-theoretical semantics with
coercive subtyping. In Semantics and Linguistic
Theory. Vancouver, volume 20, pages 38–56.
Zhaohui Luo. 2011. Contextual analysis of word
meanings in type-theoretical semantics. In Sylvain
Pogodalla and Jean-Philippe Prost, editors, Logi-
cal Aspects of Computational Linguistics. Springer
Berlin Heidelberg, Berlin, Heidelberg, pages 159–
174.
Zhaohui Luo. 2014. Formal semantics in modern type
theories: is it model-theoretic, proof-theoretic, or
both? In Nicholas Asher and Sergei Soloviev, ed-
itors, Logical Aspects of Computational Linguistics
2014 (LACL 2014), Springer, Berlin, Heidelberg,
number 8535 in LNCS, pages 177–188.
Erkki Luuk. 2010. Nouns, verbs and flexi-
bles: implications for typologies of word
classes. Language Sciences 32(3):349–365.
https://doi.org/10.1016/j.langsci.2009.02.001.
Richard Montague. 2002. The proper treatment of
quantification in ordinary English. In Paul Portner
and Barbara H. Partee, editors, Formal Semantics:
The Essential Readings, Blackwell, Oxford, pages
17–34.
Glyn Morrill. 2010. Categorial grammar: Logical syn-
tax, semantics, and processing. Oxford University
Press, Oxford.
Rachel L. Moseley and Friedemann Pulvermüller.
2014. Nouns, verbs, objects, actions, and abstrac-
tions: Local fMRI activity indexes semantics, not
lexical categories. Brain and Language 132:28 – 42.
https://doi.org/10.1016/j.bandl.2014.03.001.
Nuttanart Muansuwan. 2002. Verb Complexes
in Thai. Ph.D. thesis, University at Buf-
falo, The State University of New York.
https://arts-sciences.buffalo.edu/content/dam/arts-
sciences/linguistics/AlumniDissertations/Muansuwan
dissertation.pdf.
Carl Pollard and Ivan A. Sag. 1994. Head-Driven
Phrase Structure Grammar. University of Chicago
Press, Chicago.
Stephen G. Pulman. 1985. A parser that doesn’t.
In Proceedings of the 2nd European Meet-
ing of the Association for Computational
Linguistics, Geneva: ACL. pages 128–135.
https://www.aclweb.org/anthology/E85-1019.
Aarne Ranta. 1994. Type-theoretical grammar.
Clarendon Press, Oxford; New York.
673
v-trel: Vocabulary Trainer for Tracing Word Relations - An Implicit
Crowdsourcing Approach
Verena Lyding Christos T. Rodosthenous
Institute for Applied Linguistics, Open University of Cyprus
Eurac Research Bolzano/Bozen christos.rodosthenous@ouc.ac.cy
verena.lyding@eurac.edu
674
Proceedings of Recent Advances in Natural Language Processing, pages 674–683,
Varna, Bulgaria, Sep 2–4 2019.
generated vocabulary exercises and gaining con- lexical words with a specific meaning or formu-
tinuous input from the learners to improve LRs. laic sequences / lexical chunks where pedagogical
In the remainder of the paper, we first intro- relations are structured by a particular object rep-
duce the vocabulary trainer v-trel and describe its resentation or a part of a particular object (Ald-
different modules and their technical implementa- abe et al., 2015). As noted by Schmitt (2013), vo-
tion (Section 2). We describe how v-trel instan- cabulary learning is a complex phenomenon that
tiates a generic prototype architecture for crowd- may be explored not only from the aspects of form,
sourcing language resources (Rodosthenous et al., meaning and usage but also from a representation
2019) and discuss technical decisions taken during of different meanings in different contexts.
the implementation process. Second, we describe Vocabulary exercises based on words’ seman-
the first experiment that has been carried out with a tic relations are considered to be effective activ-
small crowd of advanced language learners2 (Sec- ities. Rosenbaum (2001) shows that background
tion 3) and discuss the results and their implica- knowledge, context and morphology are essential
tions on the potential of the proposed approach. in vocabulary instruction to enable the learner to
Third, we point out and discuss related work rel- understand and disambiguate word meanings ef-
evant to the presented approach (Section 4). Fi- fectively. The richness of acquired vocabulary de-
nally, we sum up preliminary conclusions and in- pends not only on the number of learned lexical
dicate directions for future work (Section 5). items but also on the ability to connect and share
semantic networks of similar concepts. Hadley
2 Vocabulary Trainer et al. (2018) argue that “word learning is not sim-
The vocabulary trainer builds on top of a prototype ply the process by which isolated object-label as-
architecture for crowdsourcing language resources sociations are added to the mental lexicon one by
(Rodosthenous et al., 2019). It implements an im- one but also involves the learning of interrelated
plicit crowdsourcing paradigm which follows the clusters of concepts, in which the knowledge of
idea that if a language resource (e.g., a lexicon) one concept supports the learning of another” (p.
can be used to generate language learning exer- 42).
cises, then the answers of learners to these exer- From the crowdsourcing perspective, learners
cises can be used to improve the resource, which in are used as crowdworkers to enhance the LR un-
turn will improve the quality and versatility of the derlying the vocabulary trainer, namely the com-
exercises generated (Rodosthenous et al., 2019).3 mon sense ontology ConceptNet4 (see Section
2.2). While using the vocabulary trainer for learn-
2.1 Motivation and Design ing word senses the learners are providing their
The vocabulary trainer delivers interactive vocab- knowledge of related words which is collected and
ulary exercises for learning word senses. The evaluated in order to validate and enhance the LR.
learner is asked to input words which are re- The vocabulary trainer is composed of four
lated to a given word by the semantic relation modules which are presented in the following sub-
RelatedTo, and will in the future be extended sections: 1) The exercise generation module that
to other relations such as PartOf, AtLocation retrieves words from ConceptNet and generates
etc. The learner input is collected and evaluated to exercise content (Section 2.2), 2) the exercise and
enhance the LRs that it is generated from. result storage dispatcher that ingests the previ-
From the language learning perspective, vo- ously created exercise content (Section 2.3), dis-
cabulary exercises play an important role in lan- patches it to the various learner interfaces and
guage learning (Nation and Hunston, 2013). Hul- handles the responses from the learners, 3) the
stijn (2013) notes that every word in a mental lex- evaluation module that is responsible for evaluat-
icon has formal as well as semantic associative ing if learners’ contributions are fit for expanding
features. Depending on the learner’s level of lan- the language resource and assign points to each
guage, vocabulary building may encompass single learner according to the response given (Section
2
2.4), and 4) the user interaction module, where
For the initial experiment we involved proficient non-
native speakers of English (see Section 3.2 for details).
users are presented with the exercises and submit
3
The related article discusses the paradigm in more de- their responses (Section 2.5). In Figure 1, a high-
tail and points out strategies to counter the risk of collecting
4
wrong or low-quality data from non-proficient learners. http://conceptnet.io/
675
level diagram of the vocabulary trainer’s architec- 2.3 The Exercise and Results Storage
ture is depicted along with the exchange of data Dispatcher Module
between the core modules of the system. Transactions between modules are handled using
Interested readers are invited to also browse the web services through API calls. Data are pre-
project repository 5 . sented in a JSON7 format that can be consumed
2.2 The Exercise Generation Module by any programmatically created interface. Speci-
fication of the API is available at the project repos-
The exercise generation module is responsible itory using the Swagger8 Opensource API man-
for content retrieval from language resources like agement tool. This abstraction layer allows the
ConceptNet (Speer et al., 2017) which is a large exploitation of the system from various interfaces
semantic network that describes general human without developers having to know its underly-
knowledge and how it is expressed in natural lan- ing functionality. Currently, the system offers web
guage. ConceptNet provides a large set of back- services for registering users, retrieving exercises
ground knowledge for different terms that not from the exercise generation module, checking
only describes them but also connects them with learners’ contributions, assigning points/awards,
other terms using relations such as RelatedTo, storing presented hints and showing a leaderboard.
AtLocation, PartOf, IsA, etc.
For the latter part, the evaluation module is em-
In its current version, the exercise genera-
ployed. The outcome of the evaluation module is
tion module is able to search ConceptNet for
used to update both the learner’s points and awards
terms connected with the relation RelatedTo,
and the knowledge base list of answers for that
AtLocation and PartOf and generate exer-
specific exercise.
cises using a template such as “Name one thing
The dispatcher is also directly connected with
that is <RELATION> <TERM>”. For instance, if
the database for storing/retrieving data in/from ta-
a search for knowledge that is RelatedTo the
bles and provide another abstraction layer for in-
term “dog” is initiated, ConceptNet will yield re-
formation handling workflows.
sults such as “bark”, “pet”, “animal”, etc. The
generation module processes these by removing 2.4 The Evaluation Module
stopwords, duplicates and terms that have a lan-
The evaluation module processes the learner’s an-
guage other than English. The relevant informa-
swers in order to both update the knowledge base
tion is stored in a database to be processed later by
with new words and to assign points and award
other modules along with the exercise data.
badges to the learners, which are transformed into
Searching ConceptNet is a straightforward pro-
feedback messages and leaderboards in the user
cess since the knowledge base offers a number of
interface.
APIs to query it. In our case we use the typi-
The evaluation module operates on pairs of ex-
cal search query6 to get relatedTo terms. The
ercise and result (see Figure 1) produced each time
search term is provided in canonical form in Con-
a user completes one exercise. It checks whether
ceptNet, e.g., /c/en/cat and offset is the num-
the user’s answers are known or new to the knowl-
ber of records to skip and show the next, as Con-
edge base and evaluates if new answers are valid
ceptNet API has a limit of 1000 results per call.
candidates to enhance it. According to the eval-
An example of a generated exercise is “Name
uation points received, badges are assigned to the
one thing that is related to dog”, where the learner
learners.
is expected to enter a word that exists in the re-
sults retrieved from the knowledge base. In cases More specifically, the evaluation module checks
where new words are entered by the learner, the for each user’s answer, whether it is part of the re-
evaluation module handles whether they should be sults set for that exercise or whether the answer is
added to the knowledge base or not while a spe- new. If the answer is part of the knowledge base,
cific user feedback strategy is used to account for the user receives one point. If the answer is new,
the unknown correctness of the new answers (see the user receives potential points, and the answer
Section 2.4). is stored as a candidate answer for that exercise
together with the user id. The feedback message
5
https://gitlab.com/crowdfest_task3
6 7
http://api.conceptnet.io<TERM>?other= https://www.json.org/
8
/c/en&limit=1000&offset=<OFFSET> https://swagger.io/
676
Figure 1: An overview of the vocabulary trainer architecture, depicting the main modules of the archi-
tecture, the data interchange between them and user interaction at the interface level.
to the learner informs him that the answer is new of unknown answers), they first receive potential
and that he is gaining potential points, which may points and get notified asynchronously, once the
transform into actual points if the answer is vali- unknown answer was confirmed by other learners.
dated by several users over time. Currently, the vocabulary trainer forces the user
Every time a predetermined number K of candi- to input a word in order to move to the next exer-
date answers has been accumulated, the evaluation cise. In order to support the learner in case he runs
process is triggered: out of ideas for related words on a given exercise,
the “I Don’t Know” feature provides functional-
• All new answers are ranked by their occur-
ity for requesting a “hint”. The hint provides the
rence frequency (i.e., how many mentions).
user with a correct related word, which is taken
• The top-ranked answer is selected (given that from the knowledge base. After reading the hint,
it was mentioned at least N times). the learner is free to input the hint word or any
other word that he/she deems fitting. Whenever
• All learners that gave the selected answer a learner uses that feature, there is an underlying
get informed that it was a correct answer mechanism that stores the presented hint in the
and receive two points (transforming poten- database.
tial points into actual points). Within both interfaces, learners can see their
• In addition, the learner who was the first to points and badges gained and for the Telegram
give the selected answer receives a badge. chatbot, they can also access a leaderboard.
• The knowledge base is updated with the new Chatbot interface The chatbot interface was
word for that exercise. implemented on Telegram, a very versatile mes-
saging system. Telegram is available both as a
• All occurrences of the selected word are re- native application for mobile phones and desktop
moved from the candidate list. computers for all operating systems and as a web
2.5 User Interaction Module and application. It enables the implementation of chat-
Prototypical User Interfaces (UI) bots via the Telegram Bot APIs9 . In the imple-
mented chatbot10 , users interact with the system
In the current version of the vocabulary trainer,
via a standard dialogue chat interface (see Figure
two interfaces are implemented: a chatbot on Tele-
2). Apart from textual input, the interface provides
gram and a web application. Both interfaces allow
a button area that changes during the dialogue flow
the learner to receive and complete exercises on
to simplify the interaction.
the RelatedTo relation. Learners get an imme-
diate response back on the correctness of their an- 9
https://core.telegram.org/bots/api
swer and when this is not possible (i.e., in cases 10
https://t.me/LingoGame_bot
677
Figure 3: A screenshot from the web interface.
678
150 New words New word Frequency Level
Average words
Existing words grass 15 basic
calf 6 advanced
100
meat 5 basic
cowboy 4 basic
50 farmer 4 basic
herd 4 advanced
horn 3 moderate
0
animal building clothes pasture 3 advanced
Number of words
100 medium
of English, with a high proficiency level.13 Each advanced
of these users received an email with a link to both
the Web interface and the Telegram chatbot and
50
was asked to try any of the two for more than ten
minutes within a period of two days. At the end
of this period, each user received a link to an on-
line questionnaire14 to provide feedback on both 0
animal building clothes
the interfaces and the presented questions.
To summarize, during the 2 days period we
managed to gather 4533 contributions to 26 ex-
ercises presented to the user in random order. Figure 5: Level of proficiency of new words by
The contributions were collected in 44 distinct exercise category.
user sessions, of which 17 were on the Telegram
chatbot and 27 on the Web interface. Presum- vided by exercise category as follows: 119 words
ably, the sessions mostly relate to unique users, on animals, 168 words on clothing, and 162 words
although we know of at least one user, who ac- on buildings. The lower number of new words
cessed both interfaces (see Section 3.4).15 We for the category animal relates to the higher num-
also captured 683 possible answers presented to ber of existing words in the knowledge base for
learners through the “I Don’t Know” feature. Out that category (see Figure 4). For example for the
of the 4533 contributions, 449 new words were term “cow”, 15 new words were gathered. Ta-
crowdsourced from the users based on the eval- ble 2 shows the 8 new words that were named
uation mechanism with the parameter settings de- three or more times, while seven words16 met the
scribed above (see Section 2.4 and 3.1). In terms minimum threshold of two mentions. This shows
of questionnaire feedback, we gathered 17 fully- that ConceptNet has empty spots even for basic
completed and 17 partially-completed responses. vocabulary like “grass” or “farmer”, which could
be gathered through the learners. Also, it shows
3.3 Results Analysis
that learners were able to propose advanced vo-
Characteristics of new words. Overall, the ex- cabulary such as “ruminate” or “pasture”.
periment allowed to gather 449 new words, di- A manual analysis of the proficiency level of all
13
We are aware that involving speakers with a high English
new words was carried out by an expert from lan-
proficiency implicates that the crowd does not represent gen- guage teaching. It showed that the vast majority
uine language learners. By being composed of non-native of new words is part of basic vocabulary (65% of
speakers, we however assume them to resemble advanced
learners to a degree that allows to draw meaningful conclu- all added words), while 32% are of moderate level
sions for the scope of this first evaluation. and only 3% belong to advanced vocabulary.
14
LimeSurvey, a Free Opensource online tool was locally
deployed: https://www.limesurvey.org 16
New words for term “cow” with two mentions only: bell,
15
For simplicity, we refer to user sessions as users. burger, methane, ox, ruminate, sheep and veal
679
40% Existing relations in ConceptNet. To get a
% of existing words better idea of the type and quality of words learn-
30% ers contributed regarding our initial LR, i.e., Con-
ceptNet, we queried what other relations might ex-
20% ists in ConceptNet between the search word used
for generating the exercise and the contributed
10% words other than the RelatedTo relation. From
the 26 exercises learners contributed, on average
0% 14.15 words have no other direct relation (not
animal building clothes bidirectional) with the original word from Con-
ceptNet, i.e., <SUBJECT> <RELATION> <CON-
TRIBUTED_WORD>. When searching also for
Figure 6: Average percentage of existing words, bidirectional relations, 100% of contributed words
provided by users, per exercise for each category. have such a relation between them in Concept-
Net, including RelatedTo, which can be taken
as indication for the appropriateness of the words
Interestingly, for the animal category the ra- added by learners. On average 4.54 new relations
tio between basic and moderate vocabulary is far (other than direct RelatedTo) were identified
more balanced than for the other two categories, between the contributed terms and the subject used
also more advanced vocabulary is found (see Fig- to generate the exercise.
ure 5). This could be explained by the higher num-
ber of existing words in the knowledge base for 3.4 Feedback Questionnaire
the category animal. Most basic words are proba- After the experiment, a feedback questionnaire
bly already part of ConceptNet, which implies that with six items was sent to all participants:
new words necessarily need to be more advanced.
Figure 6 shows the average percentage of exist- • Level of English: [A1/A2; B1/B2; C1/C2]
ing words (per exercise) that were entered by the • Interface used: [Chatbot; Web; both]
learners, divided by category. It shows that learn-
ers named less than 40% of the words present in • What did you notice regarding the UI? [open]
ConceptNet17 , while they can still gain knowledge
• What did you think of the questions? [open]
on more than 60% of the words, e.g. by requesting
hints through the “I don’t know” feature. • What did you like about this approach to a
Responses to “I Don’t Know” hints. Overall, vocabulary exercise? [open]
683 hints were provided to the learners by means
of the “I Don’t Know” button. In response, 365 • Any other comments? [open]
words were entered by the learners of which 331 Out of 34 users, 22 completed the closed ques-
were a direct repetition of the hint word. The tions and 9 to 14 also responded to the open ques-
lower number of entered words in relation to hints tions; 18 of 22 respondents indicated an advanced
is due to multiple “I Don’t Know” clicks before level of English; 9 users used the Chatbot, 12 used
entering a word (up to 17 clicks in a row). the Web interface, and one user used both inter-
In 34 cases, the word entered in response to the faces.
hint(s) was different from the hint(s), which indi- User interface. 9 respondents perceived the in-
cates that the hint activated the learner’s knowl- terface as clear, easy and pleasant to use, while 9
edge on related words. A look at the words shows users criticized unclear navigation and pop-ups in-
that: terrupting the workflow.
Questions. 9 users remarked that words re-
• 2 times a variation of the lemma was entered
peated too often and that the phrasing of the ques-
• 11 times an analogy of the hint was entered tion “name one thing” can be misleading in terms
of which word class is requested (3 respondents).
• 16 times a different word class was entered Approach to vocabulary exercise. 12 users
17
The lowest coverage were found for category animal, the evaluated positively the interactivity and simplic-
category with the biggest set of available words. ity of the exercise, the opportunity to learn new
680
words by means of the “I don’t know” function, proach (von Ahn and Dabbish, 2008). Among
and its effect to reactivate words and to incen- GWAP, in particular the JeuxDeMots game by
tivize brainstorming. Still, five users suggested Lafourcade (2007) shows several similarities to
that it is no real vocabulary exercise, that it would our approach. The game is designed for con-
be difficult for beginners and that the processing structing lexical data in French in a playful way.
of the answers and assignment of points/feedback In order to gain points two players have to agree
was unclear, too open-ended, and lacking negative on words related to given terms. Also, the more
feedback. recent TileAttack game by Madge et al. (2017)
Other comments. The criticism about the way builds on player agreements to gather annotations
to award points and the overall educative value for text segmentation tasks. In addition, Rodos-
were stressed further. thenous and Michael (2016) developed a platform
that combined automated reasoning with games
3.5 Discussion of the Experiment Results for acquisition of knowledge rules. Moreover,
The analysis of new words shows that our ap- in work of Guillaume et al. (2016) a game ti-
proach is promising for extending ConceptNet tled ZombiLingo was developed, for annotating
with meaningful words, in particular the more ad- dependencies in French data. In work of Cham-
vanced level of new words added for the category berlain et al. (2008), the Phrase Detectives game
animal indicates that relevant new terms can be is presented, where players contribute relation-
gathered in a progressive fashion (e.g., basic vo- ships between words and phrases, aiming to cre-
cabulary is added first). Given that this first ex- ate a resource that is rich in linguistic informa-
periment was carried out with advanced learners it tion. Yet another indirect form of crowdsourcing
needs to be evaluated to which extent similar pos- has used large collaborative knowledge bases like
itive results can be achieved with beginner and in- Wikipedia to create multilingual resources such
termediate learners. Also, the analysis indicates as YAGO3 (Mahdisoltani et al., 2015) and DB-
that ConceptNet is ample enough to propose a pedia (Lehmann et al., 2015). Recently, Meur-
wide set of new vocabulary to the user (in average ers et al. (2019) have proposed a web-based work-
more than 60% of the words per exercise). Fur- book, which can be integrated into classroom cur-
thermore, results suggest that this approach can ricula and offers instructive feedback to students
also identify new relations between terms in Con- while also gathering data on learning processes for
ceptNet. Learners managed to reproduce what was Second Language Acquistion (SLA) research.
already coded in a different manner in the resource Research works with explicit crowdsourcing
and thus improved its completeness. often employ Amazon Mechanical Turk to col-
User feedback verifies that the vocabulary lect data. For instance, Biemann (2013) created
trainer as a User Interface can be improved but the the Turk Bootstrap Word Sense Inventory of fre-
idea behind it is interesting and can be applied in quently used nouns in English and Ganbold et al.
different language learning scenarios. (2018) localized the WordNet in the Mongolian
language. Related to SLA, MacWhinney (2017)
4 Related Work proposes a collaborative platform for collecting
and sharing learner data from corpora, online tu-
Approaches for crowdsourcing language resources tors, and Web-based experimentation.
can be divided into two broad categories: 1) im- Our research presents an implicit crowdsourc-
plicit crowdsourcing, i.e., users carry out any ac- ing approach implemented as a vocabulary learn-
tivity of their interest while their data is crowd- ing application with open-ended questions for lan-
sourced as a byproduct, and 2) explicit crowd- guage resource augmentation using multiple user
sourcing, i.e., the crowd is actively engaging. interaction methods (i.e., chatbots and web apps).
The Duolingo platform (von Ahn, 2013)
presents a most similar approach to our work, 5 Conclusion and Future Work
based on language learning, as it generates lan-
guage exercises, allowing the crowdsourcing of In this paper, we presented the v-trel vocabulary
language-related data (i.e., translations). Other re- trainer for crowdsourcing language resources and
search work based on implicit crowdsourcing has delivering exercises to language learners. V-trel
utilized the Games With A Purpose (GWAP) ap- can be accessed through two interfaces, a Tele-
681
gram chatbot and a Web application. Acknowledgments
Moreover, we presented the results of an em-
This article is based upon work from COST Action
pirical evaluation and a user satisfaction survey
enetCollect (CA16105), supported by COST (Eu-
for the vocabulary trainer and provide a relevant
ropean Cooperation in Science and Technology).
discussion of these results. Feedback from users
The work presented in this paper was started dur-
and the analysis of the contributed words are taken
ing the Crowdfest meeting organized by the Ac-
into account for updating v-trel with new features
tion in January 2019 in Brussels.
such as a more attractive interface, and new ex-
ercise types18 . In addition, we aim at including
links to pictures and definitions or examples of References
use of the “hint” words, to support the learner not
Itziar Aldabe, Mikel Larrañaga, Montse Maritxalar,
only in refreshing their existing vocabulary, but Ana Arruarte, and Jon A. Elorriaga. 2015. Do-
also to acquire new words. These are first steps main module building from textbooks: Integrating
to strengthen the learning effect of the tool, while automatic exercise generation. In Cristina Conati,
in the midterm we foresee to intensify further the Neil Heffernan, Antonija Mitrovic, and M. Felisa
Verdejo, editors, Artificial Intelligence in Educa-
collaboration with language teaching experts in or- tion. Springer International Publishing, Cham, pages
der to tailor the offered exercises more closely to 521–524.
specific learning goals.
Chris Biemann. 2013. Creating a system for lexi-
Also, in future work we aim at implementing cal substitutions from scratch using crowdsourcing.
more strategies for safeguarding the quality of the Language Resources and Evaluation 47(1):97–122.
acquired data, e.g., control mechanisms for ac- https://doi.org/10.1007/s10579-012-9180-5.
tive misuse, further evaluation strategies for new Jon Chamberlain, Massimo Poesio, and Udo Kr-
words, which could also involve dynamic retrieval uschwitz. 2008. Phrase detectives: A web-based
of knowledge from ConceptNet or evaluation cy- collaborative annotation game. In Proceedings of
cles re-proposing new words in new exercises, and the International Conference on Semantic Systems
(I-Semantics’ 08). pages 42–49.
strengthened gamification elements. In addition,
we intend to evaluate the crowdsourced words and Florian Daniel, Pavel Kucherbaev, Cinzia Cappiello,
their difficulty levels in relation to established ref- Boualem Benatallah, and Mohammad Allahbakhsh.
2018. Quality control in crowdsourcing: A survey
erence data such as the English Vocabulary Pro- of quality attributes, assessment techniques, and as-
file19 or similar resources. surance actions. ACM Computing Surveys (CSUR)
As proposed in the feedback questionnaire, the 51(1):7.
educational value needs to be validated and im-
Amarsanaa Ganbold, Altangerel Chagnaa, and Gábor
proved further. Accordingly, a follow-up user Bella. 2018. Using crowd agreement for wordnet lo-
study with focus on the educational aspect is fore- calization. In Proceedings of the 11th International
seen and will be designed and carried out with au- Conference on Language Resources and Evaluation
thentic learners of different proficiency levels. (LREC-2018).
Last but not least, the vocabulary trainer will Bruno Guillaume, Karën Fort, and Nicolas Lefebvre.
be integrated into the setup of the Revita online 2016. Crowdsourcing complex language resources:
system for language learning (Katinskaia et al., Playing to annotate dependency syntax. In Inter-
national Conference on Computational Linguistics
2018) to reach a larger audience. In particu- (COLING).
lar, it can be implemented as a part of a test-
ing mode where crowdsourced questions do not Elizabeth Hadley, David Dickinson, Kathy Hirsh-
Pasek, and Roberta Golinkoff. 2018. Building se-
influence the learner’s final score. We will also mantic networks: The impact of a vocabulary in-
work towards integrating v-trel into Games With tervention on preschoolers’ depth of word knowl-
A Purpose. Previous work of Rodosthenous and edge. Reading Research Quarterly 54:41–61.
Michael (2016, 2014) suggests that crowdsourcing https://doi.org/10.1002/rrq.225.
and GWAPs, in particular, can be used to gather Jeff Howe. 2006. Crowdsourcing: A Definition.
background knowledge https://www.wired.com/2006/06/crowds/.
18
E.g., various new types of vocabulary exercises such as Jan H. Hulstijn. 2013. Incidental Learning
“fill the gap”, or “select all verbs among the given words” in Second Language Acquisition, Wiley-
19
https://www.englishprofile.org/ Blackwell, volume 5, pages 2632–2640).
wordlists https://doi.org/10.1002/9781405198431.wbeal0530.
682
Anisia Katinskaia, Javad Nouri, and Roman Yangar- Christos Rodosthenous, Verena Lyding, Alexander
ber. 2018. Revita: a language-learning platform at König, Jolita Horbacauskiene, Anisia Katinskaia,
the intersection of its and call. In Proceedings of Umair ul Hassan, Nicos Isaak, Federico Sangati, and
the 11th International Conference on Language Re- Lionel Nicolas. 2019. Designing a prototype archi-
sources and Evaluation (LREC-2018). tecture for crowdsourcing language resources. In
Thierry Declerck and John P. McCrae, editors, Pro-
Mathieu Lafourcade. 2007. Making people play for ceedings of the Poster Session of the 2nd Conference
Lexical Acquisition with the JeuxDeMots prototype. on Language, Data and Knowledge (LDK 2019).
In SNLP’07: 7th International Symposium on Natu- CEUR, pages 17–23.
ral Language Processing. Pattaya, Chonburi, Thai-
land, page 7. https://hal-lirmm.ccsd.cnrs.fr/lirmm- Christos Rodosthenous and Loizos Michael. 2014.
00200883. Gathering background knowledge for story
understanding through crowdsourcing. In Proceed-
Jens Lehmann, Robert Isele, Max Jakob, Anja ings of the 5th Workshop on Computational
Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Models of Narrative (CMN 2014). Schloss
Sebastian Hellmann, Mohamed Morsey, Patrick Dagstuhl–Leibniz-Zentrum fuer Informatik,
Van Kleef, Sören Auer, et al. 2015. Dbpedia–a Quebec, Canada, volume 41, pages 154–163.
large-scale, multilingual knowledge base extracted https://doi.org/10.4230/OASIcs.CMN.2014.154.
from wikipedia. Semantic Web 6(2):167–195.
Christos Rodosthenous and Loizos Michael. 2016. A
Verena Lyding, Lionel Nicolas, Branislav Bédi, and hybrid approach to commonsense knowledge ac-
Karën Fort. 2018. Introducing the european net- quisition. In Proceedings of the 8th European
work for combining language learning and crowd- Starting AI Researcher Symposium. pages 111–122.
sourcing techniques (enetcollect). In Peppi Taalas, https://doi.org/10.3233/978-1-61499-682-8-111.
Juha Jalkanen, Linda Bradley, and Sylvie Thouësny,
Catherine Rosenbaum. 2001. A word map for middle
editors, Future-proof CALL: language learning as
school: A tool for effective vocabulary instruction.
exploration and encounters – short papers from
Journal of Adolescent & Adult Literacy 45(1):44–
EUROCALL 2018, Research-publishing.net, pages
49.
176–181.
Norbert Schmitt. 2013. An Introduction to Ap-
Brian MacWhinney. 2017. A shared plat- plied Linguistics. Hodder Arnold Publication.
form for studying second language acquisi- Taylor & Francis. https://books.google.com.cy/
tion. Language Learning 67(S1):254–275. books?id=5gIvAgAAQBAJ.
https://doi.org/10.1111/lang.12220.
Robert Speer, Joshua Chin, and Catherine Havasi.
Chris Madge, Jon Chamberlain, Udo Kruschwitz, and 2017. Conceptnet 5.5: An open multilingual graph
Massimo Poesio. 2017. Experiment-driven devel- of general knowledge. In Proceeding of the 31st
opment of a gwap for marking segments in text. In AAAI Conference on Artificial Intelligence.
Extended Abstracts Publication of the Annual Sym-
posium on Computer-Human Interaction in Play. Luis von Ahn. 2013. Duolingo: learn a language for
ACM, pages 397–404. free while helping to translate the web. In Proceed-
ings of the 2013 international conference on Intelli-
Farzaneh Mahdisoltani, Joanna Asia Biega, and gent user interfaces. ACM, pages 1–2.
Fabian M. Suchanek. 2015. Yago3: A knowledge
base from multilingual wikipedias. In Proceedings Luis von Ahn and Laura Dabbish. 2008. De-
of the 7th Conference on Innovative Data Systems signing Games With a Purpose. Com-
Research (CIDR). munications of the ACM 51(8):57.
https://doi.org/10.1145/1378704.1378719.
Detmar Meurers, Kordula De Kuthy, Florian Nuxoll,
Björn Rudzewitz, and Ramon Ziai. 2019. Scaling
up intervention studies to investigate real-life for-
eign language learning in school. Annual Review of
Applied Linguistics 39.
683
Jointly Learning Author and Annotated Character N-gram Embeddings:
A Case Study in Literary Text
Suraj Maharjan? Deepthi Mave? Prasha Shrestha? Manuel Montes-y-Gómez†
Fabio A. González‡ Thamar Solorio?
?
Department of Computer Science, University of Houston
†
Instituto Nacional de Astrofısica Optica y Electronica, Puebla, Mexico
‡
Systems and Computer Engineering Department, Universidad Nacional de Colombia
{smaharjan2, dmave, pshrestha3, tsolorio}@uh.edu
mmontesg@ccc.inoep.mx, fagonzalezo@unal.edu.co
684
Proceedings of Recent Advances in Natural Language Processing, pages 684–692,
Varna, Bulgaria, Sep 2–4 2019.
Figure 1: Learning and using author and annotated character n-gram embeddings
685
the following conditional probability for the next- is similar to the above method of averaging the
in-sequence output character n-gram co : embeddings of annotated character n-grams, but
exp(yco ) weights each annotated n-grams embedding by
p(co |ci , . . . , ci+n , a) = P (1)
c∈V exp(yc ) their IDF scores before averaging. Mathemati-
cally, the resultant
P book vector r is represented
yc = W h(ci , . . . , ci+n ; C, a; A) + b (2) N
idf (ci ,B)∗emb(ci )
as r = i=1
N , where idf (.) is
where yc is the unnormalized log probability for the function that gets the IDF score for given
annotated character n-gram c in the vocabulary annotated character n-gram c. The idf (.) is
V , W and b are softmax parameters, and h is learned from the training data and is defined
|B|
either the concatenation or mean function applied as idf (c, B) = log |{d∈B:c∈d}| , where B is the
to character n-gram and author vectors from C collection of books, and d is an instance of book.
and A, respectively. Similar to Le and Mikolov
(2014), we call the concatenation method Dis- 3 Book Likability Prediction
tributed Memory Concatenation (DMC) and the Here we present results of using the author and
mean method Distributed Memory Mean (DMM). character based n-gram embeddings for the task
of predicting whether readers will like a book or
Phase II: Building book representations not. We use likeability as a proxy to measure the
We use the annotated character n-grams from success of a book. Narrowing down success as
Phase I to obtain the book’s representations. We a measure of readers ratings is not ideal. But it
concatenate this with the book author’s embed- gives us a practical starting approach to evaluate
ding and feed them as features to an SVM clas- our models.
sifier. Similar to Maharjan et al. (2017), we con-
sider the first 1k sentences from each book in the 3.1 Dataset
Goodreads corpus. We define the following three We experiment with the publicly available book
methods to obtain book representations: likability prediction dataset (Goodreads corpus)
Bag of annotated character n-grams (ACn): from Maharjan et al. (2017). They collected books
Similar to bag-of-word (BoW) approach, we gen- from Project Gutenberg3 . They then labeled the
erate the annotated character n-grams from books’ books into two categories: Successful and Un-
content. We then represent each book by a sparse successful, by using the average rating and the
vector and weight each annotated character n- total number of reviews received by the books
grams using their term frequency-inverse docu- on Goodreads4 . It consists of 1,003 books (654
ment frequency (TF-IDF) scores. The motivation Successful and 349 Unsuccessful) from 8 genres
behind using this representation is the success of downloaded from Project Gutenberg: Detective
stylistic analysis in the domain of books success Mystery, Drama, Fiction, Historical Fiction, Love
prediction, author attribution, and author profiling. Stories, Poetry, Science Fiction, and Short Stories.
Mean of Annotated character n-grams em-
beddings (Mean): Unlike the above method, here 3.2 Experimental Settings
we use the annotated character n-gram embed- We used the same stratified splits of 70:30 train-
dings to represent each book by a dense vector. ing to test as provided by Maharjan et al. (2017).
We generate the annotated character n-grams from We used the negative sampling (Mikolov et al.,
book content and look up their embeddings. A 2013) method to train 300-dimension embeddings
book is then represented by the mean of all an- for both authors and annotated character 3-grams.
notated character n-gram embeddings generated We filtered out 3-grams with frequency<2, set the
from its content. Mathematically, P the resulting window size to 5, configured the sample thresh-
N
emb(ci )
book vector r is represented as r = i=1N , old to 1e-5 for randomly downsampling higher-
where N is the total number of annotated character frequency 3-grams, and trained for 100 epochs.
n-grams for a book, and emb(.) is the function that We predicted success separately (Single task
gets the embeddings for a given annotated charac- (ST)) as well as simultaneously with genre (Multi-
ter n-gram c. task (MT)). We used linear kernel SVM and tuned
Inverse Document Frequency (IDF) 3
https://www.gutenberg.org/
Weighted Average (Weighted): This method 4
https://www.goodreads.com
686
Type Overlap Partial Non-Overlap
Features ST MT ST MT ST MT
Character 3-grams (Maharjan et al., 2017) 66.9 70.0 - - - -
All Typed n-grams (Maharjan et al., 2017) 66.3 69.1 - - - -
Annotated char-3gram(AC3) 66.8 69.8 71.1 67.8 68.9 70.5
Mean (DMM) 60.5 67.6 63.7 66.6 65.6 68.3
Mean (DMC) 62.8 70.0 63.5 70.1 65.1 67.7
Weighted (DMM) 56.5 65.6 65.6 69.9 66.7 69.0
Weighted (DMC) 65.3 66.4 64.8 67.2 60.0 63.5
AC3 + Author (DMM) 62.8 68.5 67.5 67.5 70.6 68.5
AC3 + Author (DMC) 69.3 68.7 67.7 68.3 71.3 70.0
Mean + Author (DMM) 62.6 70.3 68.2 66.6 71.9 66.8
Mean + Author (DMC) 69.0 69.2 68.3 67.0 71.5 71.1
Weighted + Author (DMM) 62.3 70.0 66.9 66.6 70.6 70.7
Weighted + Author (DMC) 71.1 73.8* 69.8 70.1 71.7 70.5
Table 2: Weighted F1-scores (%) using book representations with and without author embeddings under
three settings (Overlap, Partial, and Non-Overlap). *statistically significant at p < 0.02 (McNemar
significant test with and without author embeddings).
the C hyper-parameter through a grid search over predictions for that genre. We obtained a large
(1e{-4,. . . ,4}), using three-fold cross validation on negative Pearson correlation coefficient of -0.753
the training split. (p < 0.03) between distances and the number of
incorrect predictions, supporting our intuition.
3.3 Results
Table 2 shows the results for our methods under Annotated vs Plain Character n-grams: To val-
the Overlap, Partial, and Non-Overlap settings. idate the importance of annotating n-grams, we
Our first set of experiments tests book represen- trained embeddings using unannotated n-grams
tation methods (AC3, Mean, and Weighted) with- and performed likability prediction using the best
out author embeddings. The sparse feature rep- setup from before. This produced an F1-score of
resentation method AC3 (71.1%) performs bet- 69.9% (< 73.8%) illustrating the usefulness of
ter than embedding aggregation methods, Mean considering the functional behavior of character
(69.9%) and Weighted (70.1%), as the mean oper- n-grams. The performance further decreased to
ation likely removes important information. Here, 63.4% with the removal of author embeddings.
the Partial setting yields the best results. Authors as binary vectors: To further confirm
Our next set of experiments combine book rep- the advantage of learning author embeddings from
resentations with author embeddings and this im- external data, we replace author embeddings with
proves the results in most cases. We obtain the one-hot vectors indicating the book’s author. Us-
overall highest F1-score of 73.8% with Weighted ing the best setup in Table 2, this produced a score
setting under concatenation (DMC) method. This of 70.3% (< 73.8%), strengthening our intuition
result is statistically significant (p < 0.02) over that author embeddings capture style related infor-
the same setup but without authors’ embeddings. mation, which is relevant for likability prediction.
This result is also better than the results from the
state-of-the-art methods by Maharjan et al. (2017): Author Embeddings and Genre: We experimen-
Character 3-grams (70.0%) and All typed n-grams tally verify that author embeddings capture genre-
(69.1%). We also see that the DMC method yields specific information by using them to perform
consistently better results than the DMM method. genre classification. We used the best model, Au-
Author embeddings and correctness: We group thor (DMC), to automatically infer author vectors
authors by the genre of their books (in case of mul- for all books in the dataset and fed them to an
tiple genres, we pick the one with the most books). SVM classifier. We obtained F1-scores of 64.6%,
We then obtain representations for successful and 66.8%, and 64.0% with the Overlap, Partial, and
unsuccessful authors in that genre by averaging Non-Overlap settings respectively. These scores
the author vectors for these two classes. Our in- outperform a random baseline of 15.2%, showing
tuition was that if the distance between these rep- that author embeddings are also capturing more
resentations is small, there will be fewer correct general style traces related to the genre.
687
0.8
0.6
0.4
0.2
Weights
−0.2
−0.4
−0.6
Mid-word Prefix Suffix Whole-word
−0.8
And dog Car god red sea the thy
n-grams
Figure 3: Feature importance assigned by SVM to different character n-grams for the likability predic-
tion task.
4 Discriminative Annotated Character stance, the embeddings for sub (prefix, mid-word),
n-grams est (whole-word, suffix), ion (suffix, whole-word),
mid (prefix, suffix), and the (whole-word, suffix)
Figure 3 shows some of the top positively and neg-
lie far from each other. On the other hand, ful in
atively weighted annotated n-grams by the clas-
suffix and mid-word form are close together, since
sifier. We used the best performing AC3 model
the contexts for ful as a suffix (beautiful, careful)
from Table 2, AC3 under Partial setting, to extract
are similar to the contexts where it occurs mid-
the weights for the annotated character n-grams.
word (beautifully, carefully). In addition, we can
For each of the annotated n-grams, the figure also
also see a clear separation between prefix and suf-
plots the weights for all four positional variants.
fix n-grams with the mid-word and whole-word
The figure clearly shows that different forms of
n-grams occupying regions in between. The suf-
the same character n-gram have different contri-
fix and prefix n-grams mostly occupy the regions
butions towards the likability prediction of books.
above and below the zero line respectively. This
This important piece of information would have
figure visually demonstrates that learning separate
been lost if we had treated these different forms of
embeddings for n-grams with different functional
the n-grams as one. For instance, sea as a whole-
roles is important to preserve their semantics.
word has a different meaning than when it is used
as a prefix or a suffix. Accordingly, the classifier Dataset RG (65) WordSim (353) RW (2034)
has also weighted them differently. The whole- Without Annotation 16.21 4.56 16.54
Annotated 30.75 12.27* 20.02**
word form of the n-gram sea is weighted higher
than its other forms. This also holds for the case Table 3: Results for word similarity task show-
of thy and dog. During this analysis, we also found ing Spearman’s rank correlation (ρ × 100) with
that quotation marks and male honorific titles were similarity scores assigned by human annotators
highly weighted by the classifier, similar to what ∗p < 0.03, ∗ ∗ p < 0.0001
Maharjan et al. (2017) found. This most likely
points to the importance of dialogues and the pref- We also empirically show the advantage of
erence of male characters in these books. these embeddings through the word similarity task
using three standard datasets: RG65 (Rubenstein
5 Analysis of Annotated Char n-grams
and Goodenough, 1965), WordSim353 (Agirre
In Figure 4, we visualize the annotated charac- et al., 2009), and RW (Luong et al., 2013). Similar
ter n-gram embeddings by projecting them us- to FastText (Bojanowski et al., 2017), we repre-
ing PCA. For some n-grams, the embeddings of sent a word as an average of the embeddings of its
different annotations are indeed distinct. For in- character n-grams. We create two representations
688
Figure 4: Projection of annotated char n-gram into 2D space using Principal Component Analysis (PCA)
for each word: one using plain character n-grams the book representations (Mean or Weighted), we
and another with annotated character n-grams. Ta- obtained the highest mean accuracy of 86.67%.
ble 3 shows the results for the word similarity task When we added in the inferred author embeddings
with these two different approaches. The Spear- (directly getting author embeddings would reveal
man’s rank correlation coefficient between human the author, so we infer author embeddings simi-
annotations and word vectors composed of our an- lar to Le and Mikolov (2014) using only the book
notated n-gram embeddings are higher than the content without revealing the actual author of the
same obtained from plain n-grams. The difference book), the accuracy improved to 95% (∼ 10%
between the two proposed methods is statistically above Char 3-gram), showing that our approach
significant for WordSim353 (p < 0.03) and highly not only works for likability prediction but also for
statistically significant for RW2034 (p < 0.0001). AA.
These results demonstrate that annotated charac-
ter n-gram embeddings are also good at producing 7 Related Work
high-quality word representations and they might
Sapkota et al. (2015) sub-grouped character n-
be capturing semantics at some level.
grams according to grammatical classes, like af-
6 Authorship Attribution fixes, lexical content, and stylistic classes, like
beg-punct and mid-punct. With these sub-groups,
Another task to test the effectiveness of our ap- they provided empirical evidence to support the
proach is authorship attribution (AA). To ensure importance of character n-grams features in the
that the model learns to discriminate authors and task of authorship attribution. Iacobacci et al.
not the genre, we selected books from fiction gen- (2015) showed that learning separate word em-
res for authors having at least five books that were beddings for polysemous words yields the state-
not used to train our Author2Vec model. We have of-the-art result in word similarity and relational
12 such authors in our corpus. similarity tasks. Separating the same tokens or n-
We again used the first 1k sentences, used grams helps to preserve their functional and mor-
an SVM classifier in a stratified 5-fold cross- phological information which is important for all
validation setup, and tuned the C hyper-parameter tasks. Learning embeddings for words (Mikolov
using grid search for each fold. Table 4 shows our et al., 2013), n-grams (Zhao et al., 2017), and doc-
results along with two baselines: word unigram uments (Le and Mikolov, 2014) and using them as
and character 3-grams with tf-idf. Using only input for various NLP tasks (Samih et al., 2016;
689
Overlap (%) Partial (%) Non-Overlap (%)
Methods µ±σ µ±σ µ±σ
Word Unigrams 83.33 ± 5.27 - -
Char 3-grams 85.00 ± 6.24 - -
AC3 81.67 ± 6.24 83.33 ± 9.13 83.33 ± 10.54
Mean 85.00 ± 6.24 86.67 ± 8.50 83.33 ± 9.13
Weighted 81.67 ± 12.25 80.00 ± 8.50 83.33 ± 10.54
Mean + Inferred Author 83.33 ± 11.79 95.00 ± 4.08* 90.00 ± 6.24
Weighted + Inferred Author 83.33 ± 11.79 93.33 ± 6.24 90.00 ± 6.24
Table 4: Mean and standard deviation of accuracy for 5 fold cross validation AA experiments (*statisti-
cally significant at p < 0.05 from t-test with Char 3-grams)
Zhang et al., 2015; Kim, 2014) has shown im- puter to plot the trajectory of emotion throughout
provement in performance. Again, Shrestha et al. the book and its correlation with success have been
(2017) showed that applying convolutional neu- discussed (Vonnegut, 1981; Reagan et al., 2016).
ral network (CNN) over character bigrams em- However, learning and using stylistically aware
beddings improves authorship attribution of short embeddings for authors in conjunction with other
texts like tweets. They visually showed that char- relevant stylistic features extracted from books for
acter bigrams were capturing important stylistic the problem has been overlooked. Our work fills
markers to distinguish between bot-like authors this gap by adding author’s general writing style
and other normal authors. Song and Lee (2017) learned using an external corpus of books.
jointly learned the embedding for users (senders
and receivers) and showed that these user embed- 8 Conclusions and Future Work
dings capture the semantic relationship between
In this paper we explored a new dimension of
users through an auto-foldering of emails task.
modeling authors for literary texts by jointly learn-
Following these research findings, we also distin-
ing annotated character n-gram embeddings and
guish between the same character n-grams by an-
author embeddings using an external corpus. We
notating them with categories defined by Sapkota
showed that a book representation using our pro-
et al. (2015) and learn separate embeddings for
posed embeddings significantly improves likabil-
each of them in addition to learning embeddings
ity prediction results. Our approach was also able
for authors.
to obtain competitive accuracy for authorship at-
tribution and genre classification, two tasks where
Prior works in likability prediction of books
style plays a prominent role. Moreover, we also
have shown that style is an important aspect (Ma-
demonstrated that annotated character n-gram em-
harjan et al., 2017; van Cranenburgh and Bod,
beddings yield higher quality word vectors. These
2017; Ashok et al., 2013). They captured the style
results in likability prediction, authorship attribu-
of successful and unsuccessful books using lexi-
tion, genre classification, and word similarity fur-
cal, syntactic, readability, and writing density fea-
ther demonstrate the usability of annotated char-
tures, and deep learning methods with only first
acter n-grams and author embeddings in varied
1K sentences. Since style is evident even with first
tasks. In the future, we will extend our method to
few fragments, they obtained competitive results
other domains where authors’ information is im-
using only first few fragments of books. Mahar-
portant, such as author profiling.
jan et al. (2017) even showed that around 200 sen-
tences are enough to perform book success predic- Acknowledgments
tion with reasonable accuracy. Louis and Nenkova
(2013) proposed features to capture different as- This work was partially funded by National Sci-
pects of great writing (surprising, visual and emo- ence Foundation under award 1462141. We would
tional content) and used them in combination with like to thank the anonymous reviewers as well
genre-specific features to predict high quality writ- as Gayle McElvain and Brandon Burroughs for
ings in science articles. Iwana et al. (2016) ex- reviewing the paper and providing helpful com-
tracted visual features from book covers for genre ments and suggestions.
classification. Also, the potential of using a com-
690
References 31st International Conference on Machine Learn-
ing. PMLR, Bejing, China, volume 32 of Proceed-
Eneko Agirre, Enrique Alfonseca, Keith Hall, Jana ings of Machine Learning Research, pages 1188–
Kravalova, Marius Pasca, and Aitor Soroa. 2009. 1196. http://proceedings.mlr.press/v32/le14.html.
A study on similarity and relatedness using dis-
tributional and wordnet-based approaches. In Annie Louis and Ani Nenkova. 2013. What makes
Proceedings of Human Language Technologies: writing great? first experiments on article quality
The 2009 Annual Conference of the North Amer- prediction in the science journalism domain. Trans-
ican Chapter of the Association for Compu- actions of the Association for Computational Lin-
tational Linguistics. Association for Computa- guistics 1:341–352.
tional Linguistics, Boulder, Colorado, pages 19–
27. http://www.aclweb.org/anthology/N/N09/N09- Thang Luong, Richard Socher, and Christopher Man-
1003. ning. 2013. Better word representations with recur-
sive neural networks for morphology. In Proceed-
Vikas Ashok, Song Feng, and Yejin Choi. 2013. Suc- ings of the Seventeenth Conference on Computa-
cess with style: Using writing style to predict the tional Natural Language Learning. Association for
success of novels. In Proceedings of the 2013 Con- Computational Linguistics, Sofia, Bulgaria, pages
ference on Empirical Methods in Natural Language 104–113. http://www.aclweb.org/anthology/W13-
Processing. Association for Computational Linguis- 3512.
tics, Seattle, Washington, USA, pages 1753–1764.
http://www.aclweb.org/anthology/D13-1181. Suraj Maharjan, John Arevalo, Manuel Montes,
Fabio A. González, and Thamar Solorio. 2017. A
Piotr Bojanowski, Edouard Grave, Armand Joulin, and multi-task approach to predict likability of books.
Tomas Mikolov. 2017. Enriching word vectors with In Proceedings of the 15th Conference of the Euro-
subword information. Transactions of the Associa- pean Chapter of the Association for Computational
tion for Computational Linguistics 5:135–146. Linguistics: Volume 1, Long Papers. Association for
Computational Linguistics, Valencia, Spain, pages
Ignacio Iacobacci, Mohammad Taher Pilehvar, and 1217–1227. http://www.aclweb.org/anthology/E17-
Roberto Navigli. 2015. Sensembed: Learning 1114.
sense embeddings for word and relational similar-
ity. In Proceedings of the 53rd Annual Meet- Suraj Maharjan, Sudipta Kar, Manuel Montes, Fabio A.
ing of the Association for Computational Lin- González, and Thamar Solorio. 2018a. Letting
guistics and the 7th International Joint Con- emotions flow: Success prediction by modeling the
ference on Natural Language Processing (Vol- flow of emotions in books. In Proceedings of
ume 1: Long Papers). Association for Computa- the 2018 Conference of the North American Chap-
tional Linguistics, Beijing, China, pages 95–105. ter of the Association for Computational Linguis-
http://www.aclweb.org/anthology/P15-1010. tics: Human Language Technologies, Volume 2
(Short Papers). Association for Computational Lin-
Brian Kenji Iwana, Syed Tahseen Raza Rizvi, Sheraz guistics, New Orleans, Louisiana, pages 259–265.
Ahmed, Andreas Dengel, and Seiichi Uchida. 2016. http://www.aclweb.org/anthology/N18-2042.
Judging a book by its cover. arXiv preprint
arXiv:1610.09204 . Suraj Maharjan, Manuel Montes, Fabio A. González,
and Thamar Solorio. 2018b. A genre-aware at-
Vlado Kešelj, Fuchun Peng, Nick Cercone, and Calvin tention model to improve the likability predic-
Thomas. 2003. N-gram-based author profiles for tion of books. In Proceedings of the 2018 Con-
authorship attribution. In In Proceedings of the ference on Empirical Methods in Natural Lan-
Pacific Association for Computational Linguistics. guage Processing. Association for Computational
pages 255–264. Linguistics, Brussels, Belgium, pages 3381–3391.
https://doi.org/10.18653/v1/D18-1375.
Yoon Kim. 2014. Convolutional neural networks for
sentence classification. In Proceedings of the 2014 Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey
Conference on Empirical Methods in Natural Lan- Dean. 2013. Efficient estimation of word represen-
guage Processing (EMNLP). Association for Com- tations in vector space. International Conference on
putational Linguistics, Doha, Qatar, pages 1746– Learning Representations (ICLR), Workshop .
1751. http://www.aclweb.org/anthology/D14-1181.
Fuchun Peng, Dale Schuurmans, Vlado Keselj, and
Moshe Koppel, Jonathan Schler, and Shlomo Arga- Shaojun Wang. 2003. Language independent au-
mon. 2009. Computational methods in authorship thorship attribution with character level n-grams. In
attribution. Journal of the American Society for in- 10th Conference of the European Chapter of the As-
formation Science and Technology 60(1):9–26. sociation for Computational Linguistics. Associa-
tion for Computational Linguistics, pages 267–274.
Quoc Le and Tomas Mikolov. 2014. Distributed rep-
resentations of sentences and documents. In Eric P. Andrew J. Reagan, Lewis Mitchell, Dilan Kiley,
Xing and Tony Jebara, editors, Proceedings of the Christopher M. Danforth, and Peter Sheridan Dodds.
691
2016. The emotional arcs of stories are domi- Zhe Zhao, Tao Liu, Shen Li, Bofang Li, and Xiaoy-
nated by six basic shapes. CoRR abs/1606.07772. ong Du. 2017. Ngram2vec: Learning improved
http://arxiv.org/abs/1606.07772. word representations from ngram co-occurrence
statistics. In Proceedings of the 2017 Confer-
Herbert Rubenstein and John B. Goode- ence on Empirical Methods in Natural Language
nough. 1965. Contextual correlates of syn- Processing. Association for Computational Lin-
onymy. Commun. ACM 8(10):627–633. guistics, Copenhagen, Denmark, pages 244–253.
https://doi.org/10.1145/365628.365657. https://www.aclweb.org/anthology/D17-1023.
Younes Samih, Suraj Maharjan, Mohammed Attia,
Laura Kallmeyer, and Thamar Solorio. 2016. Mul-
tilingual code-switching identification via lstm re-
current neural networks. In Proceedings of the
Second Workshop on Computational Approaches
to Code Switching. Association for Computa-
tional Linguistics, Austin, Texas, pages 50–59.
http://aclweb.org/anthology/W16-5806.
Upendra Sapkota, Steven Bethard, Manuel Montes,
and Thamar Solorio. 2015. Not all character n-
grams are created equal: A study in authorship
attribution. In Proceedings of the 2015 Confer-
ence of the North American Chapter of the As-
sociation for Computational Linguistics: Human
Language Technologies. Association for Computa-
tional Linguistics, Denver, Colorado, pages 93–102.
http://www.aclweb.org/anthology/N15-1010.
Prasha Shrestha, Sebastian Sierra, Fabio A. González,
Manuel Montes, Paolo Rosso, and Thamar Solorio.
2017. Convolutional neural networks for author-
ship attribution of short texts. In Proceedings of
the 15th Conference of the European Chapter of
the Association for Computational Linguistics: Vol-
ume 2, Short Papers. Association for Computa-
tional Linguistics, Valencia, Spain, pages 669–674.
http://www.aclweb.org/anthology/E17-2106.
Yan Song and Chia-Jung Lee. 2017. Learning user
embeddings from emails. In Proceedings of the
15th Conference of the European Chapter of the
Association for Computational Linguistics: Vol-
ume 2, Short Papers. Association for Computa-
tional Linguistics, Valencia, Spain, pages 733–738.
http://www.aclweb.org/anthology/E17-2116.
Efstathios Stamatatos. 2009. A survey of modern au-
thorship attribution methods. Journal of the Ameri-
can Society for Information Science and Technology
60(3):538–556. https://doi.org/10.1002/asi.21001.
Andreas van Cranenburgh and Rens Bod. 2017. A
data-oriented model of literary language. In Pro-
ceedings of the 15th Conference of the European
Chapter of the Association for Computational Lin-
guistics: Volume 1, Long Papers. Association for
Computational Linguistics, Valencia, Spain, pages
1228–1238. http://www.aclweb.org/anthology/E17-
1115.
Kurt Vonnegut. 1981. Palm sunday: An autobiograph-
ical collage.
Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015.
Character-level convolutional networks for text clas-
sification. In Advances in Neural Information Pro-
cessing Systems. pages 649–657.
692
Generating Challenge Datasets for Task-Oriented Conversational Agents
through Self-Play
693
Proceedings of Recent Advances in Natural Language Processing, pages 693–702,
Varna, Bulgaria, Sep 2–4 2019.
tigating the learning capabilities of neural con- simplest to implement for the task oriented setting
versational agents. For this reason, we will dis- when the flow of the dialogue is already known.
cuss some artificial conversational datasets, test- They tend to be highly brittle to patterns not seen
ing strategies for dialogue models, and the main during their construction or to porting to new do-
dialogue systems approaches. mains (Marietto et al., 2013). Information Re-
Artificial Datasets. Many conversational datasets trieval Methods usually imply the two main ap-
rely on crowd sourcing (Jurčı́ček et al., 2011; Kel- proaches, namely TF-IDF and Nearest Neighbor
ley, 1984), cooperating corporations (Raux et al., models (Isbell et al., 2000; Jafarpour et al., 2010;
2005), already available records (Lowe et al., Ritter et al., 2011; Sordoni et al., 2015). More re-
2015; Danescu-Niculescu-Mizil and Lee, 2011), cently, Neural Network Models have been heavily
or participation with the real world (Yu et al., employed for conversational agents. Sequence-
2015; Zue et al., 1994). Moreover, available task to-sequence models (Vinyals and Le, 2015) per-
oriented dialogue datasets are mainly focused on form well for short conversations but fail in longer
very specific domains, such as restaurant reser- ones (Hochreiter and Schmidhuber, 1997; Cho
vation (Jurčı́ček et al., 2011; Henderson et al., et al., 2014). Hierarchical Encoder Decoder Mod-
2014), flight booking (Zue et al., 1994), bus infor- els (Serban et al., 2016) are an extension of the
mation systems (Raux et al., 2005) and giving di- sequence-to-sequence models. They handle the
rections to rooms in a hotel (Yu et al., 2015). Since context in a separate RNN and use this context for
the collection of whole dialogues is usually very response generation. Finally, Latent Variable Hi-
expensive and time consuming, there have been erarchical Encoder Decoder models (Serban et al.,
efforts in building methodologies that allow fast 2017) represent a further improvement of the Hi-
and cheap data collection (Shah et al., 2018; Kel- erarchical Encoder Decoder Model. Other recent
ley, 1984). Artificial data generation is an effec- approaches have focused on Memory Networks,
tive methodology for obtaining well defined train- that use an external memory component build into
ing datasets. Bordes et al. (2017) uses a simulator the system to store long term contexts (Weston
for this purpose. Other approaches artificially out- et al., 2014). In particular, end-to-end Memory
line the conversational flow of their examples and Networks (Sukhbaatar et al., 2015), an extension
then use crowdsourcing to convert dialogue turns where every component is trained in an end-to-end
into natural language (Shah et al., 2018). fashion, showed promising results in task oriented
dialogues (Bordes et al., 2017).
Testing Neural Dialog Models. With the rise of
Neural NLP, interpretability has become a major
3 Data Generation through Self-Play
issue. Belinkov and Glass (To appear) survey var-
ious methods addressing interpretability. In par- Unlike the natural challenge sets discussed in Sec-
ticular, one line of research deals with challenge tion 2, our focus is on testing structured reason-
sets (Lehmann et al., 1996). While the majority of ing in conversational context, by using synthetic
NLP datasets reflects a natural distribution of lan- dialogue generation to grant that phenomena un-
guage phenomena, challenge sets are meant to re- der investigation are present only at test time and
strict their focus on a specific phenomenon (quan- not at training time. To our knowledge, this is the
tifiers, plurals, anaphora, etc.) at a time (Cooper first attempt to build challenge datasets in an ar-
et al., 1996). Analyzing the ability to deal with tificial way and to use them for the dialogue do-
a specific phenomenon allows evaluating the sys- main. Natural data would not be suitable for our
tems in a more principled way. Challenge datasets purpose since the challenging test phenomena can
have been applied to diverse tasks, such as Nat- also be found in the training set. In particular,
ural Language Inference (Wang et al., 2018) and our dataset is constructed using a dialogue self-
Machine Translation (King and Falkedal, 1990). play approach (Shah et al., 2018). This approach
These datasets offer insight if a model is capable suggests that we can always simulate a conversa-
of handling a specific line of reasoning from train- tion if we treat it like a game. During the simula-
ing data to specific structures at test time. tions, we instantiate two bots; the system and the
Methods for conversational scenarios. Many user. For each conversation, we pick up a user pro-
methods have been proposed to deal with conver- file, then user and system bots carry on the conver-
sational scenarios. Rule Based Systems are the sation through pseudo-language actions regarding
694
Logical Form Example Annotation
BOT: How can I help you today ? BOT: How can I help you today ?
inform intent = transfer I need to send some money.
inform intent = transfer {amount} I want to transfer {amount}
inform intent = transfer {partner} Can I send money to {partner}?
inform intent = transfer {partner} {amount} I would like to send {amount} to {partner}.
BOT: Who is the recipient? BOT: Who is the recipient?
inform {partner} It is {partner}.
BOT: What is the transfer amount? BOT: What is the transfer amount?
inform {amount} Roughly {amount}.
Table 1: Automatically generated logical forms provided to annotators and annotated template samples
for a Money Transfer intent with 2 slots. Bot request is provided to annotators to give better context.
695
signed to handle all these possible cases. In Ta- S : How can i help you today?
ble 2, we give an example dialogue to demonstrate U : I am really concerned that i lost the bank card,
how API intensive these dialogues might become. is its deactivation possible?
Note that, for each interaction we have a specific S : Please enter the card id.
user profile, i.e. the possible entities for slots are U : I am not sure about the card id
predetermined. For example, “partner list” for S : Which account does the card belong to?
money transfer is fixed, and a user cannot pro- U : It is associated with my checking account.
vide a random name. Additionally, dialogues are A: CALL : CHECK ACCOUNT, USER ACCOUNT: CHECKING
formed with respect to policy restrictions such that A: RESPONSE : ACCOUNT CHECK , RESULT: FAILED , MES -
i) the slots for each intent are asked in a strict or- SAGE : ‘list of available accounts joint , savings’
der, ii) each slot value should be validated through S : I couldn’t find the account you are looking for.
an api call, iii) there are maximum 3 consecutive Your available accounts are joint , savings. Would
invalid user utterances after which the dialogue is you like to change your answer?
terminated, iv) the final API call for intent exe- U : Yes
cution randomly returns a success or failure. The S : Which account does the card belong to?
valid/invalid user utterances are chosen randomly U : It must be savings
in each turn. We now list the domain intents and A : CALL : CHECK ACCOUNT, USER ACCOUNT: SAVING
their descriptions as follows: A : RESPONSE : ACCOUNT CHECK , RESULT: SUCCESS
S : You have the following cards that belong to
Making a Transfer. In this intent, U SER wants
your account: prepaid card, credit card. which
to transfer some money. The system is tasked to
card would you like to block?
fill three slots i.e user account, partner name and
U : Can you block the prepaid card?
amount.
A : CALL : CHECK CARD NAME , CARD NAME : PREPAID
Canceling a Transfer. This intent handles cases A : RESPONSE : CARD NAME CHECK , RESULT: SUCCESS
where U SER wants to cancel a transfer made ear- S : Do yo confirm blocking your card?
lier. The first required slot is the transaction id and U : yes
U SER may or may not recall it. If the transaction A : CALL : BLOCK CARD , CARD ID : SAVINGS - PREPAID
id cannot be filled, SYSTEM asks for partner name A : RESPONSE : BLOCK CARD , RESULT: SUCCESS
and checks for previous transactions. S : Your card has been successfully blocked.
Blocking a Card. The main slot required for Table 2: A complete Block card example. U cor-
this intent is the card id. Similar to the canceling a responds to User, S to System, and A to API calls.
transfer intent, there are two possible paths for the
conversation flow. If U SER cannot recall the card
id, SYSTEM asks another set of slots, which are the 3.3 Challenge Test Set
user account and the card type.
We generate a wide variety of test cases using the
Transaction History Query. This intent is for same dialogue self-play methodology employed
finding a transaction in U SER’s transaction history for the training data, by modifying the structure
by searching the note field. Notes are specific texts of the original dialogues in a principled manner.
giving information about the transaction reason, Each of these test cases represents a particular
such as “Salary” or “Electricity bill”. conversational situation that has never been seen
during training but that might happen in real inter-
Account Balance Query. To search for the ac-
actions. While we cannot precisely quantify the
count balance, U SER is asked to provide the user
exhaustiveness of the considered cases in real life
account name.
scenarios due to the domain restrictions, the syn-
Account Limit Query. U SER searches for the thetic dialogue generation allows us to easily in-
maximum transferable amount from his/her ac- tegrate new challenge cases when necessary. Be-
count. This intent, together with account balance, low, we discuss in detail each test case and what
serves to evaluate the disambiguation performance function it serves to test. We also present a more
of the conversational models when facing seman- conventional test set, i.e. out of templates set,
tically similar intents - in fact their dialogue flow to disclose how much we elevate the difficulty of
and required slots are identical. the task by the challenge test cases. In our ex-
696
periments, we will disregard Out Of Vocabulary utterances in training data either without any extra
(OOV) cases, i.e. cases where entities of interest slot value or odd number of slot values while in
at test time are not seen at training time, e.g. the the new API challenge set, intent utterances con-
partner name “Michael”, (Bordes et al., 2017). tain only even number of slot values. See Table 4
for an example.
Out of Templates (OOT) Out of templates test
case is constructed by sampling annotation tem- Sys : How can i help you today ?
plates into training, development and test sets. By User : I want to see if my salary was credited
doing so, we can test whether the agent can handle → Sys : API CHECK SALARY
user utterances in an unseen linguistic form. Sys : How can i help you today?
Out of Pattern (OOP) Out of pattern test cases User : I want to see if my salary was credited
challenge the agents through the conversations from Facebook
with a structure (dialogue flow) that has not been → Sys : API CHECK SALARY FACEBOOK.
seen during training. We have constructed five Table 4: Train and test excerpt of New API
out of pattern cases for the challenge test set. To-
gether with each OOP description, we provide an
example, where the arrow indicates the answer of 3. Reordering The training dialogues have a
the system bot to be predicted, and the part in particular slot filling order for each intent. In the
italic highlights the differences between training reordering challenge set, on the other hand, the
and test sets for readability purposes. order of slots in dialogues is shuffled. The pur-
pose of reordering is to evaluate whether an agent
1. Turn Compression The training dialogue is able to generate the system utterance to fill the
structure contains a confirmation step after each correct missing slot after all other slots for the cur-
attempt to fill the current slot with an invalid value. rent intent have been filled. In the example given
Following a confirming answer from U SER for in Table 5, we show the correct system utterance
changing the slot value, S YSTEM repeats the slot of a new API challenge test case for the missing
filling question to U SER. In the turn compression slot: user account.
challenge test case, we concatenate the confirma-
tion and the new slot answers as the user utterance User : I’d like to transfer some money.
for the change confirmation question of S YSTEM. Sys : From which account?
The correct system utterance is the validation API User : from Savings please.
call of the new value instead of the slot filling Sys : Who is the recipient ?
question again. Table 3 shows an example where User : It is Michael
S YSTEM asks if U SER wants to change the partner → Sys : What is the amount?
name, and during testing the user gives a confirm- User : I’d like to transfer some money.
ing answer together with the correct partner name. Sys : What is the amount ?
User : 100 euros.
Sys : Partner name is incorrect, would you like Sys : Who is the recipient ?
to change it? User : It is Michael
User : Yes → Sys : From which account?
→Sys : What is the name of the recipient?
Sys : Partner name is incorrect, would you like Table 5: Train and test excerpt of Slots Reordering
to change it?
User : Yes, change it to Megan
4. Another Slot In a natural conversation, U SER
→Sys : API PARTNER CHECK Megan
can provide an utterance that is irrelevant to the
Table 3: Train/test excerpt of Turn Compression current turn while including a slot value relevant
to another one. Example dialogue excerpts for
Making a Transfer intent with and without such
2. New API The new API challenge is designed another slot test case are given in Table 6.
to evaluate the performance of an agent in terms of
its ability of issuing the required API calls with ap- 5. Audit More In the training dialogues, after
propriate slot values. To this end, we create intent S YSTEM requests a new slot value for the value
697
Sys : What is the partner name ? operations such as issuing API calls to KBs
User : It is Michael. (Bordes et al., 2017), which are a key element
→Sys : API PARTNER CHECK Michael to our scenario. We implemented 2 variations of
Sys : What is the partner name ? memory networks in order to test the feasibility
User : Make the amount to be 10k euros . of challenge set by analyzing the performance
→Sys : API AMOUNT CHECK 10k euros differences of the networks.
698
racy, which is the number of correct system re- another slot for SMN, turn compression and audit
sponses predicted out of all dialogue turns in the more for MMN).
whole test set. Each response under test is com-
pared to the predicted response by the model when Single vs Multiple Memory Network. Con-
given a dialogue memory, i.e. context, and a new cerning the IT cases, MMN slightly outperforms
user utterance. SMN in the non-challenge test setting. In addi-
Considering the banking domain and task sce- tion, it shows a substantial accuracy increase in
narios, for which it is almost impossible to record turn compression and audit more OOP cases. On
and collect real human interactions, we cannot the other hand, SMN surpasses MMN in reorder-
perform an experiment on real life conversations. ing and another slot OOP challenges. A similar
Although using WoZ technique and challenge pat- outlook of performances is observed for the OOT
tern annotation could be applicable, it could not be non-challenge and OOP cases aside from new api
guaranteed that we can collect reasonable amount challenge, which turns out to be the most difficult
of data for the same challenge pattern and we OOP challenge and will be discussed later.
would be obliged to know every pattern in ad- We observe that on average MMN outperforms
vance. Therefore, we also employ dialogue syn- SMN both in IT and OOT cases. One possi-
thesis for test/challenge set dialogues, which al- ble explanation is based on how the memory net-
lows for a continuous challenge design by modi- work finds the correct response. The Single Mem-
fying the template combinations and dialogue pat- ory Network does so by incorporating all the in-
terns. tents in its selection of the responses. Therefore,
it searches for more general responses while the
5.1 Test Cases and Results Multiple Memory Network assigns the same task
We compare the memory networks against the per- to a specialized Memory network, which is trained
response accuracy for In Template (IT henceforth) on that very specific intent. The specialized mem-
and Out Of Template setting. In Table 8, we show ory network is better at finding the correct re-
the test results of SMN and MMN models for IT sponse since it is trained only on one particular in-
and OOT configurations (including non-OOP and tent data and its search space is implicitly reduced.
OOP test settings). As a particular OOP performance comparison,
we noticed that SMN is better at selecting the right
IT OOT response during the reordering challenge, which
Test Case SMN MMN SMN MMN evaluates the ability of the model in learning the
Non OOP 88.62 90.17 87.39 88.27 necessary slots to accomplish an intent.
Turn Comp. 27.80 55.00 27.90 54.70
In template vs Out of Template. We found out
New API 7.42 7.83 8.17 6.67
that there is not a major difference between IT and
Reordering 54.50 45.50 54.00 41.50
OOT test case performances (slightly better for
Another Slot 38.00 25.00 41.50 27.50
SMN and slightly worse for MMN). One possi-
Audit More 15.50 34.00 16.00 35.00
ble explanation is that the tests have not been con-
OOP Avg. 28.64 33.47 29.51 33.07
ducted in an OOV setting. Therefore, the SMN
Table 8: In template non-challenge/OOP test and MMN may not learn the linguistic patterns to
results, OOT non-challenge/OOP test results in find entities (templates) but they directly learn to
terms of per-response accuracy. recognize the entities and to predict the API calls
accordingly.
OOP impact. As expected, OOP cases represent Multiple Out of pattern. As a final experiment,
a compelling challenge for our MNs, see Table 8. we wanted to inspect the effect of having the chal-
When we compare the results of the non-OOP and lenge phenomenon of interest appearing more than
the OOP cases, we observe drastic performance once in a dialogue, such as the example proposed
differences, both in IT (88.62 vs. 28.64, 90.17 vs in Table 10. For this last case we could use only
33.47) and OOT (87.39 vs. 29.51, 88.27 vs. 33.07) turn compression and audit more, that have a suffi-
settings. Still, in some settings both MNs are able cient number of slots to replicate the phenomenon
to show reasonable performances and different be- of interest over different slots in the same dia-
haviors on different challenge sets (reordering and logue. What we observe is that indeed it is difficult
699
Per-Response Accuracy handling new APIs. The results suggest that it
Test Case SMN MMN is harder for the memory networks to interpret a
Turn Compression 29.61 55.16 new combination of slots and issue the related API
Turn Compr. OOT 29.07 55.13 calls. This could be partially explained by the po-
Audit More 10.50 15.02 sition of the new API cases in the dialogue. By
Audit More OOT 10.90 15.43 design, new API cases happen at the beginning
of the conversation (i.e. giving an intent together
Table 9: Multiple Out of Pattern per Dialog with unexpected slots, see example in Table 4).
Therefore, the system has no context (no interac-
Sys : The partner name you provided is incor- tion memory) to reason on while selecting the re-
rect, would you like to change ? sponse. On the contrary, for the easier turn com-
User : Yes, change it to Michael (first Occur- pression case, the memory network is already ex-
rence of Turn Compression) pecting a possible change in the slot value (e.g.
Sys : Okay, your amount also exceeds the limit, “do you want to change the amount?”) in the fol-
would you like to change ? lowing turns, regardless of receiving it in the re-
User : Yes, It is 100 euros (The second occur- spective turn or in the next few. In fact, the net-
rence of turn compression) work is already ‘primed’ on selecting an amount
Table 10: Example of Multiple OOP dialogue related API call. Consequently, the memory net-
works have a better performance on turn compres-
sion rather than new API challenge.
for both the SMN and the MMN to handle the case
of more than one OOP. 6 Conclusions and Future Work
It can be seen that there is a drop in the per-
In this paper, we explored some challenges
formances of both MNs in Table 9 as compared
connected to dataset creation for conversational
to previous challenge tests, only for audit more.
agents and interpretability of neural models. In
This drop can be attributed to the differences of
particular, we propose a methodology to create
the contexts of the conversations that are present
rich datasets for training end-to-end conversa-
in the memory while selecting the response. Since
tional agents and challenge them on unseen pat-
the context is the previous conversation until the
terns at test time. We then experimented with
chosen turn, for the first case, i.e. the first example
Memory Networks and investigated their perfor-
in 10, the context is a usual sub-dialogue pattern
mance on the custom test cases. The apparently
that is seen during the training. So the responsi-
low accuracy levels on unseen challenge cases
bility of the agent is to understand the new unseen
suggest that the synthetic data and challenge gen-
compressed user utterance and choose the correct
eration for low resource dialogue domains can act
response. However, when we test the second turn
as a reasonable approximate to real life challenges
compression in the same dialogue, i.e. the second
in such domains. In other words, the more a dia-
turn in 10 where U SER changes the amount, the re-
logue model is able to handle these diverse chal-
sponsibility of the agent is compounded by the fact
lenges, the more it will be able to handle the
that in addition to understanding the compressed
unstructured or less structured dialogues in real
turn, it has to reason about an unseen context pat-
human-machine interaction. As a future work we
tern. In other words, the context also contains a
would like to test further neural models and cre-
form of turn that the agent has never seen during
ate additional OOP challenge sets, even combin-
training, which is the first turn compression of the
ing additional configurations.
second example in Table 10.
700
Yonatan Belinkov and James Glass. To appear. Anal- Charles Lee Isbell, Michael Kearns, Dave Kormann,
ysis methods in neural language processing: A sur- Satinder Singh, and Peter Stone. 2000. Cobot in
vey. Transactions of the Association for Computa- lambdamoo: A social statistics agent. In AAAI/IAAI.
tional Linguistics (TACL) . pages 36–41.
Antoine Bordes, Y-Lan Boureau, and Jason We- Sina Jafarpour, Christopher JC Burges, and Alan Rit-
ston. 2017. Learning end-to-end goal-oriented di- ter. 2010. Filter, rank, and transfer the knowledge:
alog. In 5th International Conference on Learn- Learning to chat. Advances in Ranking 10:2329–
ing Representations, ICLR 2017, Toulon, France, 9290.
April 24-26, 2017, Conference Track Proceedings.
https://openreview.net/forum?id=S1Bb3D5gg. Filip Jurčı́ček, Simon Keizer, Milica Gašić, Fran-
cois Mairesse, Blaise Thomson, Kai Yu, and Steve
Kyunghyun Cho, Bart Van Merriënboer, Dzmitry Bah- Young. 2011. Real user evaluation of spoken
danau, and Yoshua Bengio. 2014. On the properties dialogue systems using amazon mechanical turk.
of neural machine translation: Encoder-decoder ap- In Twelfth Annual Conference of the International
proaches. arXiv preprint arXiv:1409.1259 . Speech Communication Association.
Robin Cooper, Dick Crouch, Jan Van Eijck, Chris John F Kelley. 1984. An iterative design methodol-
Fox, Johan Van Genabith, Jan Jaspars, Hans Kamp, ogy for user-friendly natural language office infor-
David Milward, Manfred Pinkal, Massimo Poesio, mation applications. ACM Transactions on Infor-
et al. 1996. Using the framework. Technical re- mation Systems (TOIS) 2(1):26–41.
port, Technical Report LRE 62-051 D-16, The Fra-
CaS Consortium. Margaret King and Kirsten Falkedal. 1990. Using test
suites in evaluation of machine translation systems.
Cristian Danescu-Niculescu-Mizil and Lillian Lee. In COLNG 1990 Volume 2: Papers presented to
2011. Chameleons in imagined conversations: A the 13th International Conference on Computational
new approach to understanding coordination of lin- Linguistics. volume 2.
guistic style in dialogs. In Proceedings of the
2nd Workshop on Cognitive Modeling and Compu- Sabine Lehmann, Stephan Oepen, Sylvie Regnier-
tational Linguistics. Association for Computational Prost, Klaus Netter, Veronika Lux, Judith Klein,
Linguistics, pages 76–87. Kirsten Falkedal, Frederik Fouvry, Dominique Es-
tival, Eva Dauphin, et al. 1996. Tsnlp: Test suites
Jesse Dodge, Andreea Gane, Xiang Zhang, Antoine for natural language processing. In Proceedings of
Bordes, Sumit Chopra, Alexander H. Miller, Arthur the 16th conference on Computational linguistics-
Szlam, and Jason Weston. 2016. Evaluating prereq- Volume 2. Association for Computational Linguis-
uisite qualities for learning end-to-end dialog sys- tics, pages 711–716.
tems. In 4th International Conference on Learn-
ing Representations, ICLR 2016, San Juan, Puerto Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle
Rico, May 2-4, 2016, Conference Track Proceed- Pineau. 2015. The ubuntu dialogue corpus: A large
ings. http://arxiv.org/abs/1511.06931. dataset for research in unstructured multi-turn dia-
logue systems. arXiv preprint arXiv:1506.08909 .
Mihail Eric and Christopher D Manning. 2017. Key-
value retrieval networks for task-oriented dialogue. Maria das Graças Bruno Marietto, Rafael Varago
arXiv preprint arXiv:1705.05414 . de Aguiar, Gislene de Oliveira Barbosa, Wag-
ner Tanaka Botelho, Edson Pimentel, Robson
Marco Guerini, Simone Magnolini, Vevake Balara- dos Santos França, and Vera Lúcia da Silva. 2013.
man, and Bernardo Magnini. 2018. Toward zero- Artificial intelligence markup language: A brief tu-
shot entity recognition in task-oriented conversa- torial. arXiv preprint arXiv:1307.3091 .
tional agents. In Proceedings of the 19th Annual
SIGdial Meeting on Discourse and Dialogue. pages Antoine Raux, Brian Langner, Dan Bohus, Alan W
317–326. Black, and Maxine Eskenazi. 2005. Let’s go pub-
lic! taking a spoken dialog system to the real world.
Matthew Henderson, Blaise Thomson, and Jason D In Ninth European conference on speech communi-
Williams. 2014. The second dialog state tracking cation and technology.
challenge. In Proceedings of the 15th Annual Meet-
ing of the Special Interest Group on Discourse and Alan Ritter, Colin Cherry, and William B Dolan. 2011.
Dialogue (SIGDIAL). pages 263–272. Data-driven response generation in social media. In
Proceedings of the conference on empirical meth-
Sepp Hochreiter and Jürgen Schmidhuber. 1997. ods in natural language processing. Association for
Long short-term memory. Neural computation Computational Linguistics, pages 583–593.
9(8):1735–1780.
Iulian Vlad Serban, Alessandro Sordoni, Yoshua Ben-
John E Hopcroft, Rajeev Motwani, and Jeffrey D gio, Aaron C Courville, and Joelle Pineau. 2016.
Ullman. 2001. Introduction to automata theory, Building end-to-end dialogue systems using gener-
languages, and computation. Acm Sigact News ative hierarchical neural network models. In AAAI.
32(1):60–65. volume 16, pages 3776–3784.
701
Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe,
Laurent Charlin, Joelle Pineau, Aaron C Courville,
and Yoshua Bengio. 2017. A hierarchical latent
variable encoder-decoder model for generating di-
alogues. In AAAI. pages 3295–3301.
Pararth Shah, Dilek Hakkani-Tür, Gokhan Tür, Ab-
hinav Rastogi, Ankur Bapna, Neha Nayak, and
Larry Heck. 2018. Building a conversational agent
overnight with dialogue self-play. arXiv preprint
arXiv:1801.04871 .
Alessandro Sordoni, Michel Galley, Michael Auli,
Chris Brockett, Yangfeng Ji, Margaret Mitchell,
Jian-Yun Nie, Jianfeng Gao, and Bill Dolan. 2015.
A neural network approach to context-sensitive gen-
eration of conversational responses. arXiv preprint
arXiv:1506.06714 .
Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al.
2015. End-to-end memory networks. In Advances
in neural information processing systems. pages
2440–2448.
Oriol Vinyals and Quoc Le. 2015. A neural conversa-
tional model. arXiv preprint arXiv:1506.05869 .
Alex Wang, Amapreet Singh, Julian Michael, Felix
Hill, Omer Levy, and Samuel R Bowman. 2018.
Glue: A multi-task benchmark and analysis platform
for natural language understanding. arXiv preprint
arXiv:1804.07461 .
Jason Weston, Sumit Chopra, and Antoine Bordes.
2014. Memory networks. CoRR abs/1410.3916.
Zhou Yu, Dan Bohus, and Eric Horvitz. 2015. Incre-
mental coordination: Attention-centric speech pro-
duction in a physically situated conversational agent.
In Proceedings of the 16th Annual Meeting of the
Special Interest Group on Discourse and Dialogue.
pages 402–406.
Tiancheng Zhao and Maxine Eskenazi. 2018. Zero-
shot dialog generation with cross-domain latent ac-
tions. arXiv preprint arXiv:1805.04803 .
Victor Zue, Stephanie Seneff, Joseph Polifroni,
Michael Phillips, Christine Pao, David Goodine,
David Goddeau, and James Glass. 1994. Pegasus: A
spoken dialogue interface for on-line air travel plan-
ning. Speech Communication 15(3-4):331–340.
702
Sentiment Polarity Detection in Azerbaijani Social News Articles
703
Proceedings of Recent Advances in Natural Language Processing, pages 703–710,
Varna, Bulgaria, Sep 2–4 2019.
ble of extracting public opinion regarding a spe-
2 Literature Review cific topic and also works excellently with high-
level features. During the research, deep learning
As sentiment analysis is one of the prominent top- model was developed using word embedding and
ics nowadays there exists vast amount of experi- linear machine learning model was implemented.
ments applied with different methods. Sentiment Sentiment analysis can also be used with the
analysis of text for Azerbaijani language had been combination of different methodologies in the im-
investigated by Aida-zade et al. (2013). Multi ma- plementation of various applications. One of the
chine learning algorithms had been applied for researches that benefited from application of sen-
news classification in Azerbaijani language in timent analysis is done by Rosa et al. (2019). In
(Suleymanov and Rustamov, 2018; Suleymanov et the research, applications collect data about a user
al., 2018; Aida-zade et al., 2018). Cambria (2016) and give recommendations based on the data pro-
distinguishes three main approaches in the field of duced by the user. The research proposes an ap-
sentiment analysis: knowledge-based, statistical proach for a recommendation system which takes
and hybrid. The first one is the method of classi- user’s current psychological state into account us-
fying text using rule-based algorithm to extract the ing sentiment analysis. Based on the mood of a
sentiment. To help the organizations to improve user system sends different messages to the user
their decision making and improve customer satis- including relaxing, peaceful, calm and etc.
faction Zaw and Tandayya (2018) applied rule- Bansal and Srivastava (2018) applied word2vec
based algorithm called Contrast Rule-Based sen- model with machine learning algorithms to classi-
timent analysis to classify customer reviews au- fy user reviews. The word2vec model was used to
tomatically. Another rule-based algorithm is pro- represent 400.000 consumer reviews data from
posed by Tan et al. (2015) for classifying finan- Amazon as vectors. Later, the vector representa-
cial news articles. According to the research, ini- tion of data was given to the classifier as an input.
tial stage was to determine the sentiment of each In order to classify the data both Continuous Bag
single sentence in the given financial news. Next of Words (CBOW) and Skip-Gram model were
stage was calculating positivity/negativity for the implemented in combination with 4 machine
whole content of an article. learning algorithms including SVM, Naive Bayes,
Sentiment analysis is widely used to measure random forest, and logistic regression.
the public opinion about a given topic. Li et al. Severyn and Moschitti (2015) worked on an
(2017) investigated relationship between Dow application that does sentiment analysis of tweets
Jones Industrial Average and public emotions. with deep learning models. They have applied un-
During experiment approximately 200 million supervised natural language model to initialize
twitter data was collected that mentions 30 com- word embeddings which were used as distant su-
panies which are part of the New York Stock Ex- pervised corpus in their deep learning model. The
change. Researchers applied a different methodol- model was initialized by using pre-trained param-
ogy called SMeDA-SA. The method initially ex- eters of network, then trained on supervised train-
tracts all uncertain sentences from the document ing data from Semeval-2015. According to official
to create the vocabulary. As a result, the research test sets’ result, their model ranked first in phrase-
indicated that stock price of companies can be level and second in message-level tasks.
predicted with the given methodology. In some researches it is observed that diction-
One of sentiment analysis task – subjectivity ary-based approaches are also effective to extract
detection applied before sentiment analysis for in- the sentiment from text data. Nigam and Yadav
creasing accuracy performance. Rustamov et al. (2018) divided collected tweets into lexemes and
applied Adaptive Neuro Fuzzy Inference System matched the words with the terms in the diction-
(2013a), Hidden Markov Models (2013b) and ary. Matched words were weighted so that nega-
Hybrid models (2018) for detection of subjectivity tive word gets -1 score and positive word gets +1
analysis. Same techniques had been applied for score. The overall sentiment of document was cal-
document level sentiment analysis (Rustamov et culated by subtracting the weights of positive
al., 2013). The method described by Araque et al. words from weights of negative words.
(2017) is an example of deep learning algorithm
usage for sentiment analysis. Recently, deep learn-
ing is widely used to classify the text as it is capa-
704
3 Methodology assessing the result of the applied algorithm. It
helped us to estimate how accurately the model
3.1 Data Collection would work in real case situations. By considering
Data collection and annotation is an essential step that we may not have satisfactory amount of data to
for training supervised machine learning algorithms. train the model, and while splitting data into test and
Not only the collected data should be relevant to the train we eliminate some of the train data and it may
research objective but also the annotation process cause underfitting problem, we have applied 10-fold
should be carefully designed in order to have no in- cross validation. After each splitting we got the ac-
consistencies among labeled data as they could curacy score and when the splitting ended, we ob-
scale up during training phase and lead to unsatis- tained final accuracy by calculating the average.
factory results. For conducting this research using
3.2 Data Preprocessing
supervised machine learning techniques, approxi-
mately 30000 news articles had been collected and Getting clean data was first step of the experi-
monitored from online websites of famous Azerbai- ment. In order to get sentiment from data, stop
jani newspapers. News articles under social catego- words were cleaned, which have no sentiment but
ry have been observed to contain more sentiment are highly frequent in the dataset and could de-
polarity and therefore found to be more suitable for crease accuracy. Especially while applying fre-
sentiment analysis. The inter-annotator agreement quency based vectorizer, noisiness of data inter-
had been established and an additional procedure rupts quality of classification process.
had been implemented as a control mechanism in Dataset contained XML and HTML tags since
order to verify that the agreement had been fol- they were taken from online resources. Especially
lowed during annotation. One of the main require- names of websites, sources of the article, dates,
ments of agreement was to label only the articles in URL links and JS tags were present in the dataset
which sentiment has been explicitly expressed. For and they affected the prediction accuracy nega-
instance, an article about a social event cannot be tively. For example, website names which end
annotated for its sentiment unless author explicitly with the domain names such as “.az”, “.com”,
shows its social benefits or downsides. This was “.org”, “.net”, “.ru”, “.edu”, “.gov” had been re-
done to eliminate any inconsistencies emerging moved. In addition, the ones that started with
from subjective assessment of annotators. The con- “http”, “https” and “www” and extra time tags
trol mechanism had been implemented as follows. were also deleted from data. Furthermore, all un-
Firstly, each news article was given a unique identi- necessary punctuations were cleaned except semi-
fier and after random shuffling, articles were divid- colon and dash, since they are used in compound
ed into small chunks each containing 500 articles. In words in Azerbaijani language.
the first run each annotator was given a chunk. After Finally, to eliminate difference between same
all annotators finished the first chunk, the second words with different cases (uppercase and lower-
run began. In the second run, each annotator was case), all tokens had been converted to lowercase.
given a chunk and additionally 50 more already It is needed to mention that after preprocessing,
labeled articles without their labels. By comparing number of words decreased, and consequently,
these 50 articles’ new labels with old ones, we could dictionary size was reduced as well which speeded
determine in which aspects annotators did not agree up the processing of data and the classification al-
and made relevant adjustment to the annotation gorithms.
agreement to minimize the amount of disagreement
in subsequent runs. In the following runs the steps 3.3 Feature Extraction
of the second run were repeated until there was no In terms of natural language processing, there is
chunk remaining. 12210 articles were labelled ac- an essential need to convert text data into a specif-
cording to above mentioned rules. Among the la- ic format which is appropriate for applying statis-
beled news articles, 4565 of them were labeled as tical machine learning algorithms. The process is
positive and 7645 were labeled as negative. In the called feature extraction and there exists different
next stages, we had applied k-folds cross validation. methodologies for feature extraction. One of the
In this method k stands for the number of repetitions commonly used methods is called bag of words
of splitting data into test and train part. Cross valida- model (BOW) that treats each single word as a
tion is a method in machine learning that is used for feature. This approach takes collection of docu-
705
ments and converts it into a list of unordered 4 Classifiers
words called vocabulary (Chen et al., 2017). Next
step is to create a vector representation of docu- 4.1 Random Forest
ments according to the size of given vocabulary Random forest is one of the supervised learning
where existence of each unique word defines the algorithms, which is implemented in both regres-
size. In a basic vectorizing models, it counts oc- sion and classification problems. This classifier is
currence of each word in text and converts it to an a collection of recursive, tree structured models.
array of real numbers. In decision tree, the prediction is done by split-
In the domain of machine learning, there exists ting root training set into subsets as nodes, and
various vectorization approaches one of which is each node contains output of the decision, label or
counting based. Also called frequency based, it condition. After sequentially choosing alternative
provides a sparse representation of corpus of doc- decisions, each node recursively is split again and
uments as a matrix. However, frequency based at the end classifier defines some rules to predict
vectorizer has several drawbacks. Firstly, not all result. Conversely, in random forest, classifier
words have sentiment value despite how frequent- randomly generate tree without defining rules.
ly it is used in the document, namely, frequency of We have implemented random forest algorithm
commonly used words could shadow other more with the different vectorizer methodologies, and
significant and sentiment containing words. n-gram models as shown in Table 1 and got differ-
Therefore, to solve this problem term-frequency ent outcomes as described below.
and inverse-document-frequency method (tf-idf)
is widely used. In addition to the number of occur- Feature n- F1- Precision Recall
rence in the document tf-idf also takes into con- extraction grams Score
sideration word’s density in the whole corpus of
documents. It consists of two parts where term Frequency Unigram 93.21 91.02 95.87
based
frequency provides count of each word and in-
vectorizer
verse document frequency reduces value of words Bigram 92.15 88.47 95.97
that are densely used in the corpus.
Another approach of bag of words model is Trigram 89.79 82.91 97.38
hashing based vectorizer. It maintains vectors rep-
resenting each term as an integer value. Different
from others, it does not create dictionary, still hav- Unigram 93.33 90.65 95.93
ing larger matrix to encode a document. tf-idf and Tf-idf
frequency based vectorizer generates high- Bigram 92.27 88.96 96.61
dimensional vector representations. Unlike them
hashing based vectorizer suggests an efficient way Trigram 89.95 83.29 97.5
to reduce the dimensionality of the vector. It of-
fers default size for feature vector and provides an
option to reduce and increase vector size. Howev- Table 1: Random Forest Classifier Result
er, probability of collision should be considered
when choosing the size. If the number of unique While considering the highest F1 score, we got
words in the vocabulary exceeds the feature size it 93.33 percent from the combination of tf-idf
could lead to collision where several unique words vectorizer and unigram model. On the other hand,
could map to the same integer value. when taking highest recall and precision score
It is not necessity to present each single unique separately, we obtain the highest recall score of
term as an input to the vectorization method. Dif- 97.38 percent. The highest recall score yields the
ferent word combinations can be used here includ- precision of 82.91 and F1 score of 89.79 percent.
ing unigram, bigrams, and trigrams. In unigram Even though the recall is the highest, we got the
each word represents one feature. Additionally, we lowest F1 score from that combination. Addition-
can also take combination of two words called bi- ally, the highest precision score which is 91.02
grams, and combination of three words at a time percent provides 93.21 F1 score and the lowest re-
called trigrams (Bhavitha et al., 2017). call score of 95.87 percent. As visible in the Table
1 the second highest F1 score of 93.21 was ob-
706
tained from the combination of frequency based est F1 score of 95.47 percent was obtained from
vectorizer and unigram model. It is noticeable that the combination of frequency based vectorizer and
the difference between the first highest F1 score of bigram model. Additionally, it is noticeable from
93.33 percent and the second highest F1 score of the table that while considering tf-idf results the
93.21 is too close to each other and therefore was highest F1 score we got 94.66 percent which was
not statistically important. The lowest F1 score of obtained from bigram model. This combination
89.79 percent was obtained from the combination yields the recall score of 98.24 percent which is
of frequency based vectorizer and trigram model. the second highest recall score and precision score
of 91.34 which is the lowest precision score.
4.2 Naïve Bayes
The second machine learning algorithm we have
applied to our data set is Naive Bayes classifier. It Feature n- F1- Precision Recall
extraction grams Score
is a probabilistic model, which is derived from
Bayes Theorem that finds the probability of hy- Frequency Unigram 94.87 95.87 93.9
pothesis activity to the given evidence activity. based
According to naive Bayes rule each feature is in- vectorizer Bigram 95.47 97.24 93.77
dependent from each other and because of the as-
sumption about independence, occurrence of one Trigram 91.2 97.87 85.4
feature has no impact to others. Depending on the
features, Bayes classifier has several forms includ- Unigram 94.1 91.54 96.97
ing Gaussian Naive Bayes classifier, Multinomial Tf-idf
Naive Bayes, and Bernoulli Naive Bayes. In this Bigram 94.66 91.34 98.24
research Multinomial Naive Bayes which is most-
ly used for document level analysis is implement-
Trigram 90.9 84.2 98.75
ed. In this classifier, feature representation has
been generated by multinomial distribution which
Table 2: Multinomial Naive Bayes Classifier
reflects frequency of words like vectorization.
Result
Probability of an event i happening with multino-
mial Naive Bayes is formulated as:
4.3 Support Vector Machine
The third classification algorithm that we have
applied is SVM. the purpose of SVM classifica-
tion algorithm is toewgwrhwhe
define optimal hyperplane in
In this paper we have applied Multinomial Na- N dimensional space to separate the data points
ive Bayes classifier using alpha=1 with frequency from each other. N dimensional space here is
based vectorizer and tf-idf vectorizer as described number of features:
in the following table. Alpha parameter prevents
model assigning null probabilities in case of 0
term frequency. Below, table 2 clearly indicates
the impact of combination of different
vectorization methodologies, n-gram models on
the F1, recall, and precision score. Equation (1) describes the calculation of cost
Table 2 presents that combination of frequency function and hypothesis for SVM. One of the im-
based vectorizer and bigram model provided the portant terms used in SVM is kernel parameter.
highest F1 score result of 95.47 percent. In con- When the data is so huge and hardly computation-
trast, the lowest F1 score result we got among the al, kernel is used to speed up and optimize SVM.
above combinations was 90.9 percent and was ob- There are different types of the kernel parameter
tained from tf-idf and trigram model. As can be such as linear kernel, polynomial kernel, rbf ker-
seen from the table the lowest F1 score yields the nel, and sigmoid kernel. In this research linear
lowest precision score of 90.0 percent. The high- kernel parameter is used. In the research, we ana-
est precision score we got was 97.87 which yields lyzed accuracy of SVM classification algorithm
the lowest recall score of 85.4 percent. The high-
707
with different vectorization and n-gram models type of articles is catchier and preferred more by
which is described in the following table. the readers.
Having data skewness can have a direct impact
Feature n- F1- Precision Recall on the results of a machine learning model and
extraction grams Score gaining further insights can contribute to obtaining
higher results. Therefore, the impact of skewness
Frequency Unigram 95.51 95.4 95.63 on our dataset had been researched. Firstly, we ex-
based amined the precision and recall score per class to
vectorizer
Bigram 95.45 94.17 96.76 see if skewness has a significant impact on the re-
sults.
Trigram 92.82 88.73 97.31 Class Recall Precision
Trigram 93.35 89.19 97.93 Table 4: Per class precision and recall scores of
SVM classifier
Table 3: Linear SVM Result Table 4 demonstrates that skewness in the da-
taset affects the performance of the classifier.
Table 3 depicts the F1, recall, and precision re- Therefore, we explored several approaches for op-
sults for SVM classifier using various feature se- timizing the results while working with skewed
lection and n-grams models. As can be seen from dataset. The considerable change came from the
the table the best F1 score we got 96.79 percent application of adjusting sample weights inversely
was from the combination of tf-idf vectorizer and to the frequencies of each class in SVM classifier.
unigram model. While considering frequency By this way, the classifier is penalized more for
based vectorizer the highest F1 score which is the mistakes made on samples from underrepre-
95.51 percent was obtained from unigram model. sented class, namely positive. This allowed to ele-
When comparing the F1 score of unigram and bi- vate the recall score to 95.14 percent for the posi-
gram models we do not observe huge difference tive class and increase the precision score to 97.07
between results. Recall values for the SVM classi- percent for the negative class, while slightly low-
fier ranged from 95.63 percent to 97.93 percent. ering the recall to 95.95 percent.
The highest recall score was 97.93 which was ob-
tained from the combination of tf-idf and trigram
model. As described in table the lowest F1 score
92.82 percent and was gained from trigram model
and frequency-based vectorization.
708
Figure 1 demonstrates the accuracy results ob- In Zadeh L., Yager R., Shahbazova S., Reformat M.,
tained from implementation of tf-idf and frequen- Kreinovich V. (eds) Recent Developments and the
cy-based vectorization methods with each classifi- New Direction in Soft-Computing Foundations and
Applications. Studies in Fuzziness and Soft Compu-
cation algorithm. As can be seen from the figure,
ting. Springer, 361(2018):63-70.
combination of SVM and tf-idf outperforms other https://doi.org/10.1007/978-3-319-75408-6_6
classifiers with the 96 percent accuracy. The min-
imum accuracy was obtained from the combina- Aliaksei Severyn and Alessandro Moschitti. 2015.
UNITN: Training Deep Convolutional Neural Net-
tion of random forest classifier and tf-idf
work for Twitter Sentiment Classification. In Pro-
vectorizer with 91.4 percent accuracy. While con- ceedings of the 9th International Workshop on Se-
sidering accuracy result it is observable that accu- mantic Evaluation (SemEval 2015). Association for
racy range changes between 91 percent and 96 Computational Linguistics, pages 464-469.
percent. Random forest classifier does not offer https://doi.org/10.18653/v1/s15-2079
huge difference between two vectorization meth- Ayda-zade K., Rustamov S., Mustafayev E. 2013.
ods as there is only 0.2 percent accuracy differ- Sentiment Analysis: Hybrid Approach. In Transac-
ence between tf-idf and frequency-based tions of Azerbaijan National Academy of Scienc-
vectorization methods. When analyzing Naive es.”Informatics and control problems”. Volume
Bayes classifier, the best accuracy we got was XXXIII №6., p. 100-108.
94.4 percent and was obtained from the frequency Barkha Bansala and Sangeet Srivastava. 2018. Senti-
based vectorizer. ment classification of online consumer reviews us-
ing word vector representations. Procedia Comput-
5 Conclusion er Science, 132:1147-1153.
https://doi.org/10.1016/j.procs.2018.05.029
In conclusion, we experimented with three ma- Bing Li, Keith C.C. Chan, Carol Ou and Sun Ruifeng.
chine learning algorithms on news articles in 2017. Discovering public sentiment in social media
Azerbaijani language. Comparing one classifier for predicting stock movement of publicly listed
with another, depending on the n-gram model and companies. Information Systems, 69(1):81-92.
vectorization method we obtained different results https://doi.org/10.1016/j.is.2016.10.001
for different combinations. According to the re- B. K. Bhavitha, Anisha P. Rodrigues and Niranjan N.
search the highest F1-score we got was 96.79 per- Chiplunkar. 2017. Comparative study of machine
cent with the implementation of SVM on tf-idf learning techniques in sentiment analysis. In 2017
vectorizer and unigram model. Research also re- International Conference on Inventive Communica-
vealed that Naive Bayes classifier give its best re- tion and Computational Technologies (ICICCT).
sult 95.47 percent with the combination of fre- IEEE, pages 216-221.
https://doi.org/10.1109/ICICCT.2017.7975191
quency based vectorizer and bigram model, while
random forest acquires highest F1-score 93.33 D Shubham, P Mithil, Meesala Shobharani and S
percent by using tf-idf based feature extraction Sumathy. 2017. Aspect level sentiment analysis us-
and unigram model. ing machine learning. IOP Conference Series: Ma-
terials Science and Engineering, 263:042009.
Additionally, for future work we plan to im-
https://doi.org/10.1088/1757-899x/263/4/042009
prove our research by enlarging our dataset, add-
ing neutral class, applying rule-based algorithms Erik Cambria. 2016. Affective Computing and Senti-
and observing their performance in our dataset. In ment Analysis. IEEE Intelligent Systems,
31(2):102-107.
addition to them, we are going to apply word em- https://doi.org/10.1109/MIS.2016.31
bedding with different classification algorithms.
Evangelos Psomakelis, Konstantinos Tserpes,
Acknowledgments Dimosthenis Anagnostopoulos and Theodora
Varvarigou. 2014. Comparing Methods for Twitter
This work has been carried out in Center for Data Sentiment Analysis. In Proceedings of the Interna-
Analytics Research at ADA University in partner- tional Conference on Knowledge Discovery and In-
ship with E-GOV Development Center. formation Retrieval - Volume 1: KDIR. SciTePress,
pages 225-232.
References https://doi.org/10.5220/0005075302250232
Aida-zade Kamil, Samir Rustamov, Mark A. Clements Freha Mezzoudj and Abdelkader Benyettou. 2018. An
and Elshan Mustafayev. 2018. Adaptive Neuro- empirical study of statistical language models: N-
Fuzzy Inference System for Classification of Texts. gram language models vs. neural network language
709
models. International Journal of Innovative Com- Samir Rustamov, Elshan Mustafayev and Mark A.
puting and Applications, 9(4):189. Clements. 2013. An Application of Hidden Markov
https://doi.org/10.1504/ijica.2018.10016827 Models in subjectivity analysis. In 2013 7th Inter-
national Conference on Application of Information
Khanvilkar Gayatri and Vor, Deepali. 2018. Sentiment
and Communication Technologies. IEEE, pages 64-
Analysis for Product Recommendation Using Ran-
67. https://doi.org/10.1109/ICAICT.2013.6722756
dom Forest. International Journal of Engineering
and Technology(UAE), 7(3):87-89. Samir Rustamov, Elshan Mustafayev and Mark A.
https://doi.org/10.14419/ijet.v7i3.3.14492 Clements. 2013. Sentiment Analysis using Neuro-
Fuzzy and Hidden Markov Models of Text. In 2013
Li I. Tan, Wai S. Phang, Kim O. Chin and Patricia An-
Proceedings of IEEE Southeastcon. IEEE, pages 1-
thony. 2015. Rule-Based Sentiment Analysis for Fi-
6. https://doi.org/10.1109/SECON.2013.6567382
nancial News. In IEEE International Conference on
Systems, Man, and Cybernetics. IEEE, pages 1601- Samir Rustamov and Mark Clements. 2013. Sentence-
1606. https://doi.org/10.1109/SMC.2015.283 Level Subjectivity Detection Using Neuro-Fuzzy
Model. In Proceedings of the 4th Workshop on
Mirsa Karim and Smija Das. 2018. Sentiment Analysis
Computational Approaches to Subjectivity, Senti-
on Textual Reviews. IOP Conference Series: Mate-
ment and Social Media Analysis. Association for
rials Science and Engineering, 396:012020.
Computational Linguistics, pages 108–114, Atlan-
https://doi.org/10.1088/1757-899x/396/1/012020
ta, USA. https://www.aclweb.org/anthology/W13-
Myint Zaw and Pichaya Tandayya. 2018. Multi-level 1615
Sentiment Information Extraction Using the CRbSA
Sara Rosenthal, Noura Farra and Preslav Nakov.
Algorithm. In 15th International Joint Conference
2017. Sentiment analysis in Twitter. In Proceedings
on Computer Science and Software Engineering
of the 11th International Workshop on Semantic
(JCSSE). IEEE, pages 1-6.
Evaluation (SemEval-2017). Association for Com-
https://doi.org/10.1109/jcsse.2018.8457328
putational Linguistics, pages 502-518.
Nitika Nigam and Divakar Yadav. 2018. Lexicon- https://doi.org/10.18653/v1/S17-2088
Based Approach to Sentiment Analysis of Tweets
Umid Suleymanov and Samir Rustamov. 2018. Auto-
Using R Language. Singh M., Gupta P., Tyagi V.,
mated News Categorization using Machine Learning
Flusser J., Ören T. (eds) Advances in Computing
methods. IOP Conference Series: Materials Sci-
and Data Sciences. ICACDS 2018. Communica-
ence and Engineering, 459:012006.
tions in Computer and Information Science,
https://doi.org/10.1088/1757-899x/459/1/012006
905:154-164. https://doi.org/10.1007/978-981-13-
1810-8_16 Umid Suleymanov, Samir Rustamov, Murad
Zulfugarov, Orkhan Orujov, Nadir Musayev and
Oscar Araque, Ignacio Corcuera-Platas, J. Fernando
Azar Alizade. 2018. Empirical Study of Online
Sánchez-Rada and Carlos A. Iglesias. 2017. En-
News Classification Using Machine Learning Ap-
hancing deep learning sentiment analysis with en-
proaches. In 2018 IEEE 12th International Confer-
semble techniques in social applications. Expert
ence on Application of Information and Communi-
Systems with Applications, 77(1):236-246.
cation Technologies (AICT). IEEE, pages 1-6.
https://doi.org/10.1016/j.eswa.2017.02.002
https://doi.org/10.1109/ICAICT.2018.8747012
Po-Hao Chen, Hanna Zafar, Maya Galperin-
Yashaswini Hegde and S.K. Padma. 2017. Sentiment
Aizenberg and Tessa Cook. 2017. Integrating Natu-
Analysis Using Random Forest Ensemble for Mo-
ral Language Processing and Machine Learning Al-
bile Product Reviews in Kannada. In 2017 IEEE 7th
gorithms to Categorize Oncologic Response in Ra-
International Advance Computing Conference
diology Reports. Journal of Digital Imaging,
(IACC). IEEE, pages 777-782.
31(2):178-184. https://doi.org/10.1007/s10278-
https://doi.org/10.1109/IACC.2017.0160
017-0027-x
Renata Lopes Rosa, Gisele Maria Schwartz, Wilson
Vicente Ruggiero, Demóstenes Zegarra Rodríguez.
2019. A Knowledge-Based Recommendation Sys-
tem That Includes Sentiment Analysis and Deep
Learning. IEEE Transactions on Industrial Infor-
matics,15(4):2124-2135.
https://doi.org/10.1109/tii.2018.2867174
Samir Rustamov. 2018. A Hybrid System for Subjec-
tivity Analysis. Advances in Fuzzy Systems. Volume
2018: 9. https://doi.org/10.1155/2018/2371621
710
Inforex — a Collaborative System
for Text Corpora Annotation and Analysis Goes Open
Michał Marcińczuk and Marcin Oleksy
G4.19 Research Group
Department of Computational Intelligence
Faculty of Computer Science and Management
Wrocław University of Science and Technology, Wrocław, Poland
{michal.marcinczuk,marcin.oleksy}@pwr.edu.pl
711
Proceedings of Recent Advances in Natural Language Processing, pages 711–719,
Varna, Bulgaria, Sep 2–4 2019.
ical tagging, annotation attributes, metadata and tagger based on the document language set in the
many others. A more comprehensive list of func- metadata.
tions can be found in (Marcinczuk et al., 2017).
3.5 Extended Annotation Attribute Editor
3 Recent Changes and Improvements We have extended the annotation attribute editor
to handle dictionary-based attributes with a large
3.1 Open Source Project
number of possible values (see Figure 2). The im-
After 10 years of development the project has been provements include the following:
finally released as an open source project. The
source codes are available under the LGPL license • Filtering of the list of values,
and can be obtained from https://github.
• Feature to add a new element to the dictio-
com/CLARIN-PL/Inforex.
nary directly from the value picker level.
3.2 Easy Installation
• Suggestions based on values assigned to
The installation process was simplified by convert- other annotations. We have implemented two
ing the system and all required components to run heuristic of generating the candidates with
withing a set of Docker containers3 defined in a different levels of certainty, i.e.:
Compose file. The Compose4 file defines four
containers: (1) www — web server running the – values for other annotations matched by
Inforex application with background services (see the text form with the Soundex algo-
Section 3.3), (2) db — MySQL database server, rithm6 . The list of candidates is sorted
(3) liquibase — Liquibase database schema con- by their frequency,
trol and (4) phpmyadmin — web-based access – attribute values matched by the annota-
to the database (for development and maintenance tion text (full or partial matching).
purposes).
A new installation of Inforex boils down to run-
ning the following two lines of code:
sudo apt-get install composer \
docker docker-compose
./docker-dev-up.sh
712
Figure 1: Document tokenization perspective
consuming and error-prone because it was easy to same text form and category. For attribute value
miss some annotations. To overcame these prob- we collect annotations with the same text form or
lems we have created a page for batch annotation lemma and category. In case of ambiguity, i.e.
attribute modification. The page consists of two there are more than one possible value of lemma
main components, i.e. a document content with or attribute, the value remain empty and the user
annotation preview and a table with annotations has to fill it manually. The lemma auto fill feature
with their attribute (see Figure 3). is available in the Annotation lemmas perspective
and the attribute auto fill feature is available in the
3.7 Document Auto Annotation Annotation attributes perspective.
This feature was designed to reduce user effort in
3.9 Tokenization of XML documents
annotating repeatable phrases in and across doc-
uments. Auto annotation works by annotating in Inforex allows to store documents in one of the
given documents all phrases that were already an- two formats: plain text or XML. The XML format
notated in other documents. This feature works for is used to encode document structure, like in the
both untokenized and tokenized documents, how- KPWr (Broda et al., 2012) and PCSN (Marcińczuk
ever we advise to use it on tokenized documents et al., 2011) corpora. During tagging the XML
as the phrases are aligned with token boundaries tags should be ignored and only the content should
and we avoid matching of incomplete words. Af- be processed. Thus, we made the tokenization pro-
ter running auto annotation the new annotations cess to be aware of the document format (see Fig-
are presented to the user for verification. User can ure 1). For XML format the document content is
decide whether given annotation is correct, incor- cleaned from XML tags, than the content is pro-
rect or the annotation type needs a change (see cessed by the tagging service and at the end the
Figure 4). The discarded annotation are stored in tokenization is aligned with the original XML doc-
the system for future run of auto annotation. Dur- ument.
ing the next use of auto annotation the new an-
3.10 Annotation Attribute Browser
notations which were previously discarded are ig-
nored. The attribute value browser (see Figure 5) allows
to browse corpus annotations by given attribute
3.8 Lemma and Attribute Auto Fill value. The page consists of three elements:
These features were designed to reduce user effort • View configuration — provides a set of fil-
in setting annotation lemmas and attribute values. ters, including: shared attribute, document
They both works in a similar way — for each an- language and subcorpus,
notation in the document the lemma or attribute
value is set based on other annotations in the cor- • Attribute values assigned to annotations —
pus. For lemma we collect annotations with the list of values their frequency,
713
Figure 3: Batch annotation attribute editor
• Annotations with the selected value. • Final — export only the final tags. If there
are tokens without final tags than the missing
3.11 Export of Morphological Tagging and tags are reported as errors.
Annotations Agreements
• User (agreement) — export tags created by
For tokenized document Inforex can store up to
selected user. For tokens which does not con-
three layers of morphological tags:
tain user agreement tags the tagger tags are
• Tagger — tags produced by a tool, taken.
• Agreement — tags entered by a user in the • Tagger — export tagger tags.
agreement mode,
• Final — tags approved by the super user.
3.12 Improved Support for Word Sense
During export it is possible to define which Annotation
layer of tags should be exported. It is possible to
The existing mechanism for word sense annota-
choose one of the following options (see Figure 6):
tion was limited to a single set of words and their
• Final or tagger (if final not present) — export senses (Marcińczuk et al., 2012). We have re-
the final tags. For tokens which does not have moved the limitation and allow to define and use
the final tag a tagger tag is taken. any number of sets of word senses in the WSD
714
Figure 5: Annotation attribute browser
4 Case Studies
In this section we present two uses cases in which
various features of Inforex were used in real-life
projects.
715
Figure 7: The extended perspective for word sense annotation
716
Figure 8: Summary of morphological disambiguation agreement
annotated. All tags verified by the team leader ob- and primarily used for creation of The Adventure
tained the status of final annotations. They were of the Speckled Band corpus, it was successfully
added to the version published within CLARIN- applied to prepare Corpus of the colloquial Pol-
PL infrastructure (Błaszczak et al., 2019). ish language (Oleksy, 2019) in another project.
This corpus has been designed to address the prob-
lem of morphological tagging of user-generated
content (UGC) as part of the project ”SentiCog-
nitiveServices — next generation service for au-
tomating voice of customer and social media sup-
port based on artificial intelligence methods”8 .
The whole corpus (approximately 400000 tokens)
is manually annotated with morphological infor-
mation and furthermore the sample of 100 docu-
ments was prepared as a result of 2+1 annotation.
5 Summary
The last two years have been productive in the
development of the Inforex system. Many new
features and extensions were implemented during
that time and the most important were presented in
this paper. Majority of the features and improve-
ments were dedicated by users. The most impor-
tant news is the that Inforex has been finally re-
leased as an open source project.
Figure 10: Morphological information provided
Work financed as part of the investment in the
for human annotators
CLARIN-PL research infrastructure funded by the
717
Figure 9: The perspective for morphological disambiguation agreement
Polish Ministry of Science and Higher Education. and Tools for Digital Humanities (LT4DH) at COL-
ING 2016. pages 76–84. http://tubiblio.ulb.tu-
darmstadt.de/97939/.
References
Michał Marcinczuk, Marcin Oleksy, and Jan Kocon.
Marta Błaszczak, Kacper Paszke, Ewa Rudnicka, 2017. Inforex - a collaborative system for text cor-
Marcin Oleksy, Jan Wieczorek, Wioleta Kobylińska, pora annotation and analysis. In Ruslan Mitkov
Dominika Fikus, and Dagmara Kałkus. 2019. and Galia Angelova, editors, Proceedings of the
The adventure of the speckled band 1.0 (man- International Conference Recent Advances in Nat-
ually tagged). CLARIN-PL digital repository. ural Language Processing, RANLP 2017, Varna,
http://hdl.handle.net/11321/667. Bulgaria, September 2 - 8, 2017. INCOMA Ltd.,
pages 473–482. https://doi.org/10.26615/978-954-
Bartosz Broda, Michał Marcińczuk, Marek Maziarz, 452-049-6 063.
Adam Radziszewski, and Adam Wardyński. 2012.
KPWr: Towards a Free Corpus of Polish. In Nico-
letta Calzolari, Khalid Choukri, Thierry Declerck, Michał Marcińczuk, Monika Zaśko-Zielińska, and Ma-
Mehmet Uğur Doğan, Bente Maegaard, Joseph Mar- ciej Piasecki. 2011. Structure annotation in the pol-
iani, Jan Odijk, and Stelios Piperidis, editors, Pro- ish corpus of suicide notes. In Ivan Habernal and
ceedings of LREC’12. ELRA, Istanbul, Turkey. Václav Matoušek, editors, Text, Speech and Dia-
logue, Springer Berlin Heidelberg, volume 6836 of
Wei-Te Chen and Will Styler. 2013. Anafora: Lecture Notes in Computer Science, pages 419–426.
A web-based general purpose annotation
tool. In Proceedings of the 2013 NAACL Michał Marcińczuk, Jan Kocoń, and Bartosz Broda.
HLT Demonstration Session. Association 2012. Inforex – a web-based tool for text corpus
for Computational Linguistics, pages 14–19. management and semantic annotation. In Nico-
http://www.aclweb.org/anthology/N13-3004. letta Calzolari (Conference Chair), Khalid Choukri,
Thierry Declerck, Mehmet Uğur Doğan, Bente
Richard Eckart de Castilho, Éva Mújdricza-Maydt, Maegaard, Joseph Mariani, Asuncion Moreno, Jan
Seid Muhie Yimam, Silvana Hartmann, Iryna Odijk, and Stelios Piperidis, editors, Proceedings
Gurevych, Anette Frank, and Chris Biemann. 2016. of the Eight International Conference on Language
A web-based tool for the integrated annotation of Resources and Evaluation (LREC’12). European
semantic and syntactic structures. In Proceedings Language Resources Association (ELRA), Istanbul,
of the workshop on Language Technology Resources Turkey.
718
Marcin Oleksy. 2019. Corpus of the colloquial
polish language. CLARIN-PL digital repository.
http://hdl.handle.net/11321/637.
Maciej Piasecki, Stan Szpakowicz, Marek Maziarz,
and Ewa Rudnicka. 2016. PlWordNet 3.0 – Al-
most There. In Verginica Barbu Mititelu, Corina
Forăscu, Christiane Fellbaum, and Piek Vossen, edi-
tors, Proceedings of the 8th Global Wordnet Confer-
ence, Bucharest, 27-30 January 2016. Global Word-
net Association, pages 290–299.
Adam Radziszewski. 2013. A tiered crf tagger for pol-
ish. In Intelligent Tools for Building a Scientific In-
formation Platform.
Pontus Stenetorp, Sampo Pyysalo, Goran Topić,
Tomoko Ohta, Sophia Ananiadou, and Jun’ichi Tsu-
jii. 2012. brat: a web-based tool for NLP-assisted
text annotation. In Proceedings of the Demonstra-
tions Session at EACL 2012. Association for Com-
putational Linguistics, Avignon, France.
Liling Tan and Francis Bond. 2011. Building
and annotating the linguistically diverse NTU-
MC (NTU-multilingual corpus). In Proceed-
ings of the 25th Pacific Asia Conference on
Language, Information and Computation. Insti-
tute of Digital Enhancement of Cognitive Process-
ing, Waseda University, Singapore, pages 362–371.
https://www.aclweb.org/anthology/Y11-1038.
719
Semantic Language Model for Tunisian Dialect
Abstract 1 Introduction
720
Proceedings of Recent Advances in Natural Language Processing, pages 720–729,
Varna, Bulgaria, Sep 2–4 2019.
the Tunisian Railway Transport Net- language model (LM) and phonetic dictio-
work. This corpus is based on speech nary. As our works aim at suggesting a
transcriptions. In order to obtain method in order to construct a LM, we pro-
a standardized and normalized cor- vided further explanations of this compo-
pus, we employed our Tunisian Di- nent by concentrating on its different types
alect CODA (Conventional Orthogra- that have been proposed in the literature.
phy for Dialectal Arabic) (30). As we mentioned earlier, the LM seeks
• We present our proposed method of to represent the language behavior in or-
creating the n-class LM. To achieve der to confirm or refute the propositions
this, we show the different stages of made by the acoustic module. In litera-
our method. ture, several statistical LM types are rec-
ognized as being the most efficient models
• We evaluate the performance of the
in ASR. Among these models, we can cite
discriminative algorithm based on the
n-gram, sequences of n-words and facto-
CRF model in order to perform the
rial LM. We will present some examples of
semantic labeling task for sponta-
the aforementioned approaches in the fol-
neous speech in the Tunisian Dialect
lowing sub-sections.
in the n-class LM construction con-
text.
2.1 N-gram model
• Testing the impact of using CRF
model in the semantic labelling field. Thanks to their simplicity and efficiency,
It is also important to assess its effect n-gram models are the most widely used
on creating semantic classes on the LM in the speech recognition field. They
one hand, on developing an n-class are based on the assumption that the ap-
LM and on its perplexity level on the pearance of a single word depends only
other hand. on its history. In practice, estimating this
• We eventually disclosed the elabo- probability is very difficult. In fact, no
rated experiments we went through learning corpus can make it possible to ob-
and the obtained results. serve all the sequences of possible words.
As a result, the basic idea of n-gram mod-
The remainder of this paper is organized
els consists in considering only the se-
as follows: Section 2 tries to show the
quences of words of length n, i. e the cal-
main role of ASR system and its compo-
culation is approached by a limited history
nents. We also shed light on the main
consisting of the n-1 preceding words. The
types LM. Section 3 discusses the related
major drawback of this modeling type is
work in this field summarizing the main
that it can lead to assigning a zero prob-
aspects of every work. Section 4 de-
ability to any n-gram that has never been
scribes the TARIC corpus in Tunisian Di-
encountered in the training corpus. This
alect used for experiments. In section 5,
problem is serious especially when this n-
we present the CRF discriminative model
gram could be perfectly valid in linguis-
used to perform semantic labeling then,
tics.
we explain our method to construct n-class
LM for Tunisian Dialect. Then Conclu-
sion is drawing in the last section. 2.2 N-class model
721
to the appearance of n-class LM. The chief of word classes. The figure 1 shows some
idea of this model is to group vocabulary words with the same semantic behavior.
words into lexical classes and to estimate
the probability of the word sequence, such
as the probability of a sequence of lexical
classes (5). One of the clearest concep-
tion of motivations for n-class models is
that a word of a given class, not necessar-
ily found in the learning corpus, inherits
Figure 1: some words with the same semantic be-
the probability of all the other representa-
havior.
tives of its class. In addition, it is possible
to add words to classes without having to
re-estimate the model probabilities. How-
ever, the problems faced by n-class models 3 Related Work
are numerous. The first major difficulty is
that this type of model requires the need In this Section, we are going to review the
for a prelabeled learning corpus (24). Nev- existing works related to classify vocab-
ertheless, manual labeling is particularly ulary words for the construction of an n-
heavy despite its exact results. type class LM. In the context of training
classes of words, (10) proposes a simple
2.3 Factorial LM word classification algorithm for statisti-
cal LM in speech recognition. The clas-
The factored LM is based on the princi- sification criterion used in this approach is
ple that a word is no longer seen as a sim- the similarity of words. Indeed, the princi-
ple graphic chain but rather as a vector of ple is based on the criterion of substitution
characteristics (3). These characteristics or replacement. According to this algo-
can include the lemma and the grammat- rithm, two words are similar since they can
ical class of a word, morphs, its kinds, its be substituted in the learning corpus (10).
numbers, or Booleans indicating the word According to this automatic word classi-
belonging to given semantic classes. On fication approach, the word accuracy rate
the theoretical side, factorial models have was increased by 8.6% with a reduction in
already shown good results for some tasks, perplexity of about 6.9% (10). The method
such as machine translation (14). At the proposed by (8) is essentially based on the
practical level, few works have relied on principle of combining different sources of
this model, especially in the speech recog- information at the class formation level. In
nition task. his work, (8) uses two types of informa-
tion: contextual information and prior in-
To conclude, due to the limitation of our formation. The former is the most com-
field of work ”Tunisian Railway Trans- monly used, corresponds to n-gram depen-
port Network”, several words with simi- dencies. This information can be collected
lar behaviors exist, (semantic for exam- not only at the words level, but also at the
ple) but they do not have the same prob- level of previously constructed classes of
ability of appearance; their class group- words (8). It is fundamental to take into
ings will therefore be possible. Moreover, account the contextual information in or-
the amount of learning data is reduced. In der to better distribute the words into the
this context, the use of the n-class model classes. Thus, the use of contextual infor-
is beneficial at several levels. For these mation is of interest in the context of im-
reasons, we propose to build an n-class proving the predictability of the model. It
LM that is based mainly on the integra- makes it possible to offer a better distribu-
tion of purely semantic data. Indeed, our tion of words into classes and thus, a more
method will be used to create a LM based balanced distribution of distributions (8).
on semantic information for the creation The second type, either semantic or syn-
722
tactic information, is formalized by cate- consists of 21,102 statements and 66,082
gories or grammars. In the approach pro- words.
posed by (8), the used information a pri-
ori is extracted from a learning corpus la- The Tunisian dialect the Tunisians’ mother
beled in grammatical categories. The ap- tongue, which is used in their daily oral
proach proposed by (29) is based on con- communication. It is becoming more
textual information (left context and right and more useful not only in interviews,
context), so words that appear frequently talk shows and public services but also
in similar contexts should be assigned to in blogs, chat rooms, SMS, e-mails, etc.
the same class. According to (29), differ- However until now the Tunisian dialect
ent vocabulary words are classified using has no standardized spelling. Indeed, this
the k-means algorithm. The particularity dialect differs from MSA and it does not
of this approach is based on the fact that have a standard spelling because there are
the number of words in a class is set to no academies of Arabic dialect. Thus, to
k and if there is a class whose number of obtain coherent learning data and to have
words is less than k then that class will be a robust and powerful language model, it
merged with another. The main advantage is necessary to utilize a standard spelling.
of this algorithm is its simplicity to find Indeed, there are words with many forms.
centroids and suddenly, the cost of merg- For example, the word àñJ
¯P PP /reserva-
ing words or classes becomes less expen- tion/ can be written in four different ways:
sive. The approach developed by (2) pro-
¯P P@ P, àñJ
àñJ P@ P and àñJ
¯P@
¯P QK
P.
poses to integrate semantic information for
As a result, each word has a unique
the formation of word classes in the statis-
form. Spelling transcription guidelines
tical LM of ASR system. This approach
CODA (Tunisian Dialect writing conven-
is based on a pivot language (called IF for
tion), (30), were adopted.
Interchange Format), which represents the
meaning of the sentence regardless of the
language (2). Thus, the criterion of choice The normalization step is essential be-
of classes is guided by the definition of the cause it presents a key point for the other
pivot language and the most used concepts steps of our method. Among the normal-
in the IF. isation Tunisian Dialect words we have:
(i)Number ”sixteen” is written as A ¢ J.
4 Tunisian dialect corpus (ii)To define the future, we must follow the
collection following form: A K. + verb, for example:
We create our own corpus of real spo-
ú
æÖß A K. . (iii) To define the negation, we
must follow the following form: AÓ + verb.
ken dialogues corresponding to the infor-
mation request task in railway stations in
collaboration with the Tunisian National The Tunisian Dialect is characterized by
Railway Company (SNCFT). This cor- the presence of foreign words, such as for
pus is called TARIC, for Tunisian Ara- instance: French, English, Spanish, Ital-
bic Railway Interaction Corpus [16]. The ian, etc.To transcribe these words, Arabic
main task of the TARIC corpus is in- alphabets have been used. These foreign
formation request in the Tunisian dialect words have a unique form, for example:
about the railway services in a railway sta- PñKP [Back], à@ Q K [train]... At the end of
tion. These requests are about consulta- this step, we obtain a standardized corpus,
tion, train type, train schedule, train des- the figure 2 represents a corpus extract be-
tination, train path, ticket price and ticket fore the normalization step.
booking. The creation of the corpus was
done based on two steps: the production
of audio recordings and the manual tran-
scription of these recordings.This corpus
723
work. As a result, we have decided to cre-
ate semantic blocks that consist of group-
ing one or more words into a single word.
Semantic block is defined as a group of
two or more words. Indeed, this pre-
treatment consists of adding a (-) between
two or more words to build a single word.
Among the words that can be grouped to-
gether to form a semantic block, we find
úæ AÓ followed by another word to indi-
cate the time for example é«A ú
æ AÓ [1
PM]. Cities whose names are composed
Figure 2: Corpus extract before the normalisation . ø YJ
[name of Tunisian
such as YK
PñK
step
5 N-class LM construction
724
task of words or semantic blocks extracted
from a sentence and their corresponding 1 X
concepts in the field of railway request in- p(x̄|ȳ; w) = exp wj Fj (x̄, ȳ)
z(x̄|w) j
formation. In this task, each word or se- (1)
mantic block is labeled with a concept in-
dicating its appropriate semantic nature. According to these equations, x̄ repre-
Thus, semantic labeling is not done word sents the sequence of words or seman-
by word because we can find words that tic blocks (observation), ȳ represents the
can have several meanings depending on sequence of labels, and w stands for the
the context in which they are used. Sub- weights. Fj corresponds to a feature func-
sequently, the labeling of a word or a se- tion. The latter depends on the sequence
mantic block is done while taking into ac- of words, the current label, the previous
count its left and right neighborhood in a label and the current position in an utter-
sentence. The figure 5 shows examples of ance. We utilized the CRF++ toolkit (16)
semantic labeling. in our experiments. It is a customizable
and open source, which implement the
CRF for segmenting and labeling sequen-
tial data. It is written in C++, employs
fast training based on gradient descent and
generates n-best candidates. In order to
measure the performance of the labelling
task, three evaluation metrics were gen-
erally adopted. The latter allow express-
ing Recall and Precision. These measures
could be calculated as follows:
725
classes. Each class represents an abstrac- tic classes, we use it in combination with
tion of similar labels. In fact, a seman- the data of LM to build our new ”seman-
tic class may correspond to a label or a tic” model. In the LM learning corpus all
group of labels, whereas a label cannot words (or semantic blocks) are replaced by
belong to only one class. Thus, if we class names. The result is a ”prepared”
consider
the two statements: Sentence 1: corpus that contains both words ( or se-
X ñK [Tunis second class].
C¿ ÐAK
Pð mantic blocks) and Semantic classes. Fi-
nally, we use the SRILM 1 toolbox to learn
Sentence 2: C¿ PAJ
ÓðQK éñ [Sousa
LM including semantic classes. SRILM is
first class]. They become similar from a toolkit for building and applying statisti-
this point of view because the words ñK cal LM, primarily for use in speech recog-
[name of city] and éñ
[name of city] nition, statistical tagging and segmenta-
belong to the same semantic class group- tion, and machine translation.
X and PAJ
ÓðQK
ing ”city” and the words ÐAK
Pð
belong to the same semantic class group- 5.5 Evaluation of a LM
ing ”Class”. Indeed, the number of classes
Several measures are used to evaluate the
must be limited while the number of oc-
quality of LM. We present perplexity as
currences of words belonging to same
the most used method. Perplexity (PPL) is
class must be large enough to come up
a quick method to evaluate LM. It is com-
with correct learned probabilities. We
monly used for several years to judge the
therefore choose to limit ourselves to the
quality of LM (15). This evaluation met-
selection of classes most frequently en-
ric is used to measure the prediction abil-
countered in our corpus, corresponding
ity of LM on a test corpus not seen during
to our Field of study ”Tunisian Railway
learning. The principle of perplexity is to
Transport Network”. In this step, we iden-
check how much LM can predict the word
tify the most common label and group
sequences of the language it is supposed to
them into classes. A class then corre-
model. Perplexity is defined by:
sponds to a set of words leading to the
same semantic labelling representation. n
1 X
P P L = 2− n logP (wt |h) (4)
Table 1: Examples of Semantic Classes t=1
Semantic classes Variants semantic tags Where P (wt |h) represents the probability
City Destination-Station.. proposed by the LM for the word knowing
Schedule Trip-time, Arrival-hour .. the history h. Indeed, the perplexity of LM
Response Confirmation, Negation.. is between 1 and V, V is the size of vocabu-
lary, that is to say the number of words that
compose it. A reduced value of perplexity
In our case, this classification provides leads to better LM prediction capability.
about 20 semantic classes such as [City],
[Response], [Request-Concept], [Compar- As we have already mentioned, a low
ison], [Schedule]. After obtaining the list value of perplexity reflects a strong pre-
of semantic classes, as shown in Table 1, dictive power of LM. Thus, to better judge
we can then directly associate each word our choice of creating an n-class word
of our corpus with the class to which it be- model, we compared the created model
longs. with the 3-gram type model on the same
test corpus of evaluation. Additionally,
5.4 N-class calculation to assess the impact of using the CRF
model to perform the semantic labelling
We have already done the first experience
task in order to construct semantic classes,
for n-class LM calculation for Tunisian
Dialect. After obtaining the list of seman- 1
http://www.speech.sri.com/projects/srilm/
726
we compared the n-class created model ilar but they do not have the same proba-
with using the CRF in the semantic la- bility of appearance, so their class group-
belling task and the n-class model without ings will be possible because of the lim-
using the CRF in the semantic labelling itation of our field of work ”the Tunisian
task. The table below shows a compari- Railway Transport Network”. The main
son between the n-class model based on idea of this model is to group vocabulary
the CRF model in the Semantic labeling words into semantic classes and to con-
task and the n-class model without apply- sider mainly the calculation of the prob-
ing the CRF model in the Semantic label- ability of a sequence of words such as the
ing task together with the n-gram model in probability of a sequence of these seman-
terms of perplexity. tic classes. To do this, we have followed
these steps:We firstly construct semantic
Table 2: Value of perplexity calculated on the blocks that consists in grouping one or
same test corpus more words into a single word that we
call ”semantic blocks”. Secondly, we do
Type of model Perplexity the semantic tagging. So in order to ob-
3-gram 74.46 tain a labeled corpus, the semantic label-
n-class without CRF 4.17 ing step consists in giving a label for each
n-class with CRF 3.87 single word or for each semantic block.
To improve the sequence labeling task, we
proposed to use a discriminative algorithm
based on the Conditional Random Field
Table 2 shows the very significant rela- (CRF) model. Thirdly, we form seman-
tive reduction in perplexity by applying tic classes. In fact, a semantic class may
the CRF model in the semantic labelling correspond to a label or a group of labels,
compared to the other models. These re- whereas a label cannot belong to only one
sults are consistent with what could be ex- class.LM calculation based on the SRILM
pected: (1) it is the classification based on tool. Evaluating the constructed model by
semantic data that has minimized the per- calculating its perplexity. In order to test
plexity of the obtained LM. The value of the model generated by our statistical LM
the n-class model perplexity remains well system, we compared the created model
below that of the 3-gram model on the with the 3-gram type model on the same
test corpus. Interestingly, the same mod- test corpus. Secondly, to better judge the
els as for the learning corpus have the low- impact of using the CRF model to perform
est perplexity value on the test corpus. (2) the semantic labelling task in order to con-
the application of CRF model to perform struct semantic classes, we compared the
the semantic labelling task affects the im- n-class created model based on the CRF
provement in creating semantic classes, by in the semantic labelling task and the n-
taking into account the n-class LM, and class model without using the CRF in the
calculating the perplexity rates. semantic labelling task. According to this
evaluation, we can say that the classifica-
6 Conclusion tion based on semantic data has minimized
the perplexity of the LM obtained as com-
In this work, we have interested in con- pared to the rapport 3-gram LM. More-
structing a statistical LM as one of the over, the use of the CRF model to perform
components of ASR system. The main the semantic labelling task has an impact
role of this model is to give the probability on the improvement in creating semantic
of distribution on sets of word sequences. classes, by taking into account the n-class
In particular, we are interested in n-class LM and calculating the perplexity rates.
LM by using semantic information for the
creation of word classes. Our choice is jus-
tified by the fact that some words are sim-
727
References [10] Farhat A., Isabelle J.F., O’Shaughnessy
D.: Clustering Words for Statistical Lan-
[1] Bahl L.R., Brown P.F., Souza P.V., guage Mod-els based on Contextual Word
Mercer R.L.: A tree-based Statistical Similarity, Dans Proc. IEEE International
Language Model for Natural Language Conference on Acoustics, Speech and Sig-
Speech Recognition, IEEE Transactions nal Processing, Atlanta, USA (1996).
on Acoustics, Speech and Signal Process-
ing (1989). [11] Jelinek F.: Continuous Speech Recogni-
tion by Statistical Methods, Proc. Of the
[2] Bigi B : Modle de langage smantique pour IEEE (1976).
la reconnaissance automatique de parole
dans un contexte de traduction, Labora- [12] Ji G., Bilmes J.: Dialog Act Tagging Us-
toire Parole et Langage - Aix-en-Provence ing Graphical Models, Proc. Of the IEEE
(2012). Intl Conf. on Acoustics, Speech, and Sig-
nal ProcessingICASSP (2005).
[3] Bilmes J. A., Kirchhoff K.: Factored Lan-
guage Models and Generalized Parallel [13] Habash, N., Diab, M. and Rambow, O.:
Backo, Proc. of Human Language Tech- Conventional Orthography for dialectal
nologies, North American (2003). Ara-bic,Proceedings of the Eighth Interna-
tional Conference on Language Resources
[4] Bouagres F. : Attelage de systmes de tran- and Evalua-tion, LREC-2012, (2012).
scription automatique de la parole, thse
[14] Kirchhoff K., Yang M.: Improved Lan-
de doctorat, universit du Maine-Le Mans-
guage Modeling for Statistical Machine
France (2012).
Translation, Proc. of the ACL Work-
[5] Brown P. F., DellaPietra V. J., Souza P. V., shop on Building and Using Parallel Texts
Lai J. C., Mercer R. L.: Class-Based N- (ParaTex), Morristown, NJ, USA. As-
Gram Models of Natural Language, Com- sociation for Computational Linguistics
putational Linguistics (1992). (2005).
[6] Chelba C., Engle D., Jelinek F., Jimenez [15] Kneser R., Ney H.: Improved Cluster-
V., Khudanpur S., Mangu L., Printz H., ing Techniques for Class-Based Statisti-
Ristad E., Rosenfeld R., Stolcke A., Wu cal Language Modeling. Proc. Euro-
D.: Structure and Performance of a De- pean Conference on Speech Communica-
pendency Language Model, Dans Proc. tion and Technology, Berlin, Allemagne
of the European Conf. on Speech Com- (1993).
munication and Technology (Eu-rospeech) [16] Kudo, T.: crf++,
(1997). http://chasen.org/ taku/software/CRF++/
[7] Chelba C., Jelinek F.: Structured Lan- [17] Lecorv G. : Adaptation thmatique non su-
guage Modeling , Computer Speech and pervise d’un systme de reconnaissance au-
Language (2000). tomatique de la parole, thse de doctorat,
Universit europenne de Bretagne (2010).
[8] Damnati G. : Modles de langage et clas-
sification automatique pour la reconnais- [18] Masmoudi A., Ellouze M., Estve Y.,
sance de la parole continue dans un con- Hadrich Belguith L., Habash N.: A Cor-
texte de dialogue oral homme-machine, pus and Pho-netic Dictionary for Tunisian
thse de doctorat, uni-versit dAvignon et Arabic Speech Recognition, LREC’2014,
des pays du vaucluse (2000). Reykjavik, Iceland (2014).
[9] Graja M., Jaoua M. and Belguith L.: [19] Masmoudi A Ellouze M., Estve Y.,
Discriminative Framework for Spo- Bougares F HadrichBelguith L: Phonetic
ken Tunisian Dialect Understanding. tool for the Tunisian Arabic, SLTU’2014,
SLSP(2013) Russia, (2014).
728
[20] Masmoudi, A., Habash, N., Khmekhem, Computational Linguistics (ACL), Mor-
M., Esteve, Y. and Belguith, L.: Ara- ristown, NJ, USA. Association for Com-
bic Transliteration of Romanized Tunisian putational Linguistics, (2002).
Dialect Text: A Preliminary Investiga-
tion, Computational Linguistics and Intel- [29] Zitouni I.: Linearly Interpolated Hier-
ligent Text Processing - 16th International archical N-gram Language Models for
Conference, CICLing 2015, Cairo, Egypt, Speech Recogni-tion Engines, IBM T.J.
p.608-619, (2015). Watson Research Center, NY, Bell Labs
Alcatel-Lucent, NJ,USA, (2008).
[21] Masmoudi, A., Bougares,F., Ellouze,
[30] Zribi I., Boujelben R., Masmoudi A., El-
M., Estve, Y., Belguith, L.: Automatic
louze M., Belguith L., Habash N.: A
speech recognition system for Tunisian di-
conventionnal Orthography for Tunisian
alect. Language Resources and Evaluation
Arabic, LREC’2014, Reykjavik, Iceland,
52(1): 249-267 (2018).
(2014).
[22] Abir Masmoudi, Rim Laatar, Mariem
Ellouze, Lamia Hadrich Belguith: N-
Class Language Model for Tunisian Di-
alect Automatic Speech Recognition Sys-
tem. LPKM (2018)
729
Automatic diacritization of Tunisian dialect text using Recurrent Neural
Network
1 Introduction
In this respect, we performed the task of
Modern Standard Arabic (MSA) as well as Arabic restorating diacritical marks without taking into
dialects are usually written without diacritics(24). account any previous morphological or contextual
It is easy for native readers to infer correct pro- analysis. Moreover, we diagnosed different as-
nunciation from undiacritized words not only from pects of the proposed model with various training
the context but also from their grammatical and options. The latter include the choice of transcrip-
lexical knowledge. However, this is not the case tion network (long short-term memory (LSTM)
for children, new learners and non-native speak- networks, bidirectional LSTM (B-LSTM)) and the
ers as they dont have a good mastery of rich lan- impact of RNN sizes. The size of the neural net-
guage derivations. Moreover, the absence of dia- work is a function of the number of hidden lay-
critical marks leads to ambiguity that affects the ers. Our goal is to choose the most pertinent layers
730
Proceedings of Recent Advances in Natural Language Processing, pages 730–739,
Varna, Bulgaria, Sep 2–4 2019.
in the Tunisian dialect based on the final findings 2.2 Challenges in the absence of
provided by various experiments. diacritization in Tunisian dialect
This model will be compared to our previous The absence of diacritics signs in the Tunisian di-
CRF and SMT models (24) by utilizing the same alect texts often increases the morphological, syn-
training and testing corpus. tactic and semantic ambiguity in the Tunisian di-
alect. Some of them are presented as follows:
The remaining of this paper is structured as fol-
lows: In Section 2 we describe the linguistic back- • Morphological ambiguity: The absence
ground of the Tunisian dialect, we try to show the of the diacritical marks poses an important
complexity of diacritization tasks based on exam- problem at the association of grammatical in-
ples and we present the main level of diacritiza- formation of the undiacritized word (24). For
tion. Section 3 introduces our proposed model and example, the word I . ªË /lEb/ admits several
experiments. Section 4 provides an exhaustive ex- possible words that correspond to different
perimental evaluation that illustrates the efficiency solutions at the grammatical labeling level.
and accuracy of our proposed method. Section 5 We can find the plural noun ”toys” and the
summarizes the key findings of the present work verb ”play” in 3rd person masculine, singular
and highlights the major directions for future re- of passive accomplishment.
search.
• Syntactic ambiguity: It should be noted that
2 Linguistic background the ambiguities in the association of gram-
2.1 Tunisian dialect matical information, related to the undiacritic
words, pose difficulties in terms of syntactic
The language situation in Tunisia is ”poly-
analysis (24). For example,
the undiacritic
glossic”, where distinct language varieties, such as
the normative language (MSA) and the usual lan-
Q
expression ð
J.ËA« YJ
ËñË@ I
. J» can admit two
guage (the Tunisian dialect) coexist (24). different diacritization forms that are syntac-
tically accepted.
As an official language, MSA is extensively
present in multiple contexts, namely education, – We find the diacritization
form
business, arts and literature, and social and legal ð Q
J.ËA« YJ
ËñË@ I.J» [The boy
wrote
written documents. However, the Tunisian dialect on the desk] whose syntactic structure
is the current mother tongue and the spoken lan- corresponds to a verbal sentence.
guage of all Tunisians from different origins and – In addition, we find the diacritiza-
distinct social belongings. For this purpose, this tion form whose syntactic structure
dialect occupies a prominent linguistic importance corresponds
in Tunisia. Q
to a nominal sentence
ð
J.ËA« YJ
ËñË@ I.J» [The boy’s books are
on the desk].
Another salient feature of the Tunisian dialect
is that it is strongly influenced by other foreign
• Semantic ambiguity:Tthe different diacriti-
languages. In fact, it is the outcome of the in-
zation of a word can have different meanings,
teraction between Berber, Arabic and many other
even if they belong to the same grammatical
languages such as French, Italian, Turkish and
category. For example, among the possible
Spanish. The manifestation of this interaction be-
Q¯ /qryt/ we find
dicaritization of the word IK
tween these languages is obvious in introducing
borrowed words. As a result, the lexical register the following dicaritization:
of the Tunisian dialect seems to be more open and
–
Q¯ /qryt/ [I read]
IK
contains a very rich vocabulary.
The Tunisian dialect has other specific aspects. –
Q¯ /qaryt/ [I taught].
IK
Indeed, since this dialects spoken rather than writ-
ten or taught at school, there is neither grammati- These two diacritic words have the same
cal, nor any orthographical or syntactic rules to be grammatical category: verb but they have two
followed. different meanings.
731
2.3 Diacritization level output sequence y1T by performing the following
The use of diacritic symbols in several instances operations for t = 1 to T (13):
is quite crucial in order to disambiguate homo-
ht = H(Wxh xt + Whh ht−1 + bh ) (1)
graphic words. Indeed, the level of diacritization
refers to the number of diacritical marks presented
on a word to avoid text ambiguity for human read- yt = Why ht + by (2)
ers. There are four levels of possible diacritization.
where H is the hidden layer activation function,
• Full Diacritization: this is the case where Wxh is the weight matrix between input and the
each consonant is followed by a diacritic. hidden layer, Whh is the recurrent weight matrix
This type of diacritization is used mainly between the hidden layer and itself, Why is the
in classical Arabic, especially in religion- weight matrix between the hidden and output lay-
related books and educational writings. ers, bh and by are the bias vectors of the hidden and
output layers, respectively. In a standard RNN, H
• Half Diacritization: the objective of this cat- is usually an element-wise application of sigmoid
egory is to add diacritics of a word except the function. Such a network is usually trained using
letters that depend on the syntactic analysis the back-propagation through time (BPTT) train-
of the word. Often, it is the before last letter ing (27).
that depends on syntactic analysis of a word
but it is not always the case due to the use of • Long short-term memory: LSTM
suffixes.
An alternative RNN called Long Short-Term
• Partial Diacritization: is the case of adding Memory (LSTM) is introduced where the conven-
only lexical vowels. The latter can be de- tional neuron is replaced with a so-called memory
fined as the vowels with which the mean- cell controlled by input, output and forget gates
ing of words changes. The goal of marking in order to overcome the vanishing gradient prob-
these vowels is to remove ambiguity from the lem of traditional RNNs (12). In this case, H can
meaning of words. be described by the following composite function
(13):
• No Diacritization: This level is completely
underspecified. The script is subject to ambi-
it = σ(Wxi xt + Whi ht−1 + Wci ct−1 + bi ) (3)
guity, especially with homographs (4).
732
A BLSTM processes the input sequence in both and an output layer. Each layer fulfils a particu-
directions with two sub-layers in order to account lar purpose. In what follows, we will explain the
for the full input context. These two sub-layers advantages of each layer.
compute forward and backward hidden sequences
→
− ← − • Input layer: This level consists in mapping
h , h respectively, which are then combined to
compute the output sequence y as follows (13): the letter sequence w to a vector sequence x.
We have checked and prepared data of our
corpus. In combining the gemination mark
→
− →
−
ht = H(W →
− xt + W−
→→− h →)
+ b− (8) with another diacritic, each character in the
xh h h t−1 h
corpus has a label corresponding to 0,1 or 2
diacritics. Character embedding input, which
←
−
ht = H(W ←
←
−
(9) is initialized randomly and updated during
− xt + W←
−←−h −)
+ b←
xh h h t−1 h training, means that each character in the in-
put sentence is represented by a vector of d
→
− ←
− real numbers. It is worth pointing out that
→ h t + W←
yt = W− − h + by (10)
hy hy t
adding a linear projection after the input layer
3.2 Model architecture affects the learning of a new representation
In our diacritization task, the basic idea is to at- for the latter embedding.
tribute a corresponding diacritical label to each • Hidden layer: This layer consists in map-
character. Hence, we apply RNN to model our se- ping the vector sequence x to a hidden se-
quence data, where a sequence of characters con- quence h. Several types of hidden layers have
stitutes the input and the probability distribution been used to choose the best performance and
over diacritics forms the output. Schematically, the best result in the automatic diacritization
our RNN structure is employed in this work as the of the Tunisian dialect. Hence, these experi-
following figure: ments were based on LSTM different layers
ranging from one layer to multiple B-LSTM
layers.
exp(yt [l])
P (l|wt ) = P 0
(11)
l0 exp(yt [l ])
733
The network is trained using Gradient Descent pus is a representation of spontaneous dis-
optimizer with learning rate 0.0003 and a mini- courses in Tunisian dialect. This dialect cor-
batch size of 200. For dropout, a rate of 0.2 pus of transcribed discourses deals with mul-
is used both on embedded inputs and after each tiple themes, such as social affairs, health, re-
type of hidden layers; either LSTM or B-LSTM. ligion, etc.
Weights are randomly-initialized with normal dis-
tribution of zero mean and 0.1 standard deviation • We utilized another type of corpus that is the
and weight updates after every batch. The loss result of a conversion tool from Latin written
function is the cross-entropy loss summed over all texts (also called Arabizi) into Arabic scripts
outputs. The GPU used is Nvidia GTX 580 that following the CODA conversion. The Ara-
has 16 streaming multiprocessors and 1.5 GB of bizi corpus is collected from social media like
memory Facebook, Twitter and SMS messaging (22).
4 Results and discussion • In order to solve the problem of the lack of re-
4.1 Evaluation Metric sources for the Tunisian dialect, we have cho-
sen to gather corpora from blog sites written
In order to measure our model performance, an
in this dialect using Arabic alphabets (24).
evaluation metric, known as Diacritic Error Rate
(For more details see (24))
(DER) is generally used. DER indicates how
many letters have been incorrectly restored with
As mentioned above, the Tunisian dialect dif-
their diacritics. The DER can be calculated as fol-
fers from MSA and it does not have a standard
lows (24):
spelling because there are no academies of Arabic
(1 − |T S|) dialect. Thus, to obtain coherent learning data, it
DER = ∗ 100 (12) is necessary to utilize a standard spelling. Indeed,
|T G|
there are words with many forms. For example,
Where |T S| is the number of letters assigned
¯P PP /reservation/ can be written
the word àñJ
correctly by the system, and |T G| is the number
of diacritized letters in gold standard texts. in four different ways:
¯P P@ P, àñJ
àñJ P@ P and
¯P@
¯P QK
P.
àñJ
4.2 Datasets
This section shows a breakdown of different sizes
of our data sets, which were gathered from various In this work, spelling transcription guidelines
sources. So far, we have used four existent types CODA (Tunisian Dialect writing convention),
of corpora for our teamwork. (36), were adopted. CODA is a conventional-
ized orthography for Dialectal Arabic. In CODA,
• We made use of our TARIC corpus (Tunisian every word has a single orthographic representa-
Arabic Railway Interaction Corpus) (24). tion. It uses MSA-consistent and MSA-inspired
The latter collected information about the orthographic decisions (rules, exceptions and ad
Tunisian dialect used in a railway station. hoc choices). CODA preserves, also, dialectal
This corpus was recorded in the ticket offices morphology and dialectal syntax. CODA is eas-
of the Tunis railway station. We recorded ily learnable and readable. CODA has been de-
conversations in which there was a request signed for the Egyptian Dialect (11) as well as the
for information about the train schedules, Tunisian Dialect (36) and the Palestinian Levan-
fares, bookings, etc. This corpus consists of tine Dialect (20). For a full presentation of CODA
several dialogues; each dialogue is a com- and an explanation of its choices, see ((11), (36)).
plete interaction between a clerk and a client.
All the words are written using the Arabic al- The normalization step is essential because it
phabet with diacritics. The diacritics indicate presents a key point for the other steps of our
how the word is pronounced. The same word method. Among the normalisation Tunisian Di-
can have more than one pronunciation. alect words we have:
• The second corpus is called STAC (Spo-
ken Tunisian Arabic Corpus)(35). This cor- • Number ”sixteen” is written as A ¢ J.
734
• To define the future, we must follow the
following form: A K. + verb, for example:
ú
æÖß A K. .
• To define the negation, we must follow the
following form: AÓ + verb.
735
the 3-layer B-LSTM configuration was adopted.
Table 2: Tunisian dialect diacritization corpus
statistics.
Train Dev Test To conclude, a 3-layer BLSTM models
# Statements 23,255 1,550 6,202 achieved the best results.
# words 129,649 8,643 34,574
# Letters 64,8247 43,216 172,867 4.4 Error Analysis
In order to reveal the weaknesses of our automatic
diacritic restoration RNN models, we analyzed all
errors that are mainly due to the following reasons:
4.3 Result
In this section, we present the evaluation outcome • We noticed that these errors are due to the
of our established diacritization models. We use presence of foreign words in our corpus.
DER as an evaluation metric. The adopted RNN
has from 1 to 4 hidden layers, each with 250 neu- • Some error words with prefixes, or suffixes
rons. This number is chosen after different exper- or both can be significantly perceived. It is
iments. We come up with the conclusion that a hard to diacritize these complex words in a
smaller number of neurons (less than 250) have correct way, as the in flection diacritical mark
an impact on accuracy rate and a greater number is related to the last letter of the stem rather
do not improve it in a significant way. Table 3 than to the last letter of the suffix.
gives an overview of the RNN models outcomes
in terms of diacritic error rate (DER). • Errors due to form/spelling diacritization er-
rors. Errors caused by ”Shadda” (consonant
doubling), or Tanween (nunation). We per-
Table 3: Diacritization Error Rate Summary for ceived that restoring ”shadda” is harder than
the Tunisian dialect RNN model restoring the other diacritics.
Model DER
LSTM 13.86% • Errors due to missing and incorrect short
B-LSTM 12.31% vowels (i.e. lexical diacritics).
2-layer B-LSTM 11.53%
3-layer B-LSTM 10.72%
4-layer B-LSTM 10.83% We have manually checked 150 error samples of
our input RNN model. The following figure shows
an example of 4 sample sequences that have errors.
Table 3 shows the effect of using LSTM and The words that have errors are underlined and in
B-LSTM models, and the number of hidden lay- red.
ers on the DER. According to this table, results
show a DER of 13.56% for LSTM and 12.31%
for B-LSTM. Based on the results of our RNN,
we detected an enhancement of 1.55% in DER of
the B-LSTM model as compared to LSTM model.
This means that B-LSTM is more performant than
LSTM.
Figure 4: Sample sequences with errors
Moreover, we noticed that increasing the num-
ber of B-LSTM layers from one hidden layer
In about 21% of the samples, we have remarked
to three layers improves accuracy. But, we ap-
that the absence of ”shadda” in some words can
plied the 3-layer BLSTM because its accuracy is
lead to a semantic ambiguity of the verb. For in-
not only closer but also fasther than the 4-layer
Q¯ [I
stance, sample 1 shows that target verb IK
BLSTM. Indeed, the training time rises from 3:52
to 6:78 hours when the number of layers pro-
Q¯ [I read]. These two dia-
taught] is output as IK
gresses monotonically from 3 to 4 and the testing critic words have the same grammatical category:
time icreases from 3.65 to 5.41 minutes. Hence, verb but they have two different meanings.
736
Diacritization errors in test samples can cause ous published researches in the Tunisian dialect
about 4% of errors. For example, sample 2 dis- field.
plays that the ”Fatha” in the word B was mistak-
enly entered after the last letter rather than the first Table 4: Diacritization results of related work
letter. (CRF and SMT models) and our RNN model
Model DER
Some error words with prefixes, suffixes or both 3-layer B-LSTM 10.72%
can be significantly perceived, in about 41% of the CRF 20.25%
samples. In another illustration, the error word in SMT 33.15%
sample 3 has both the prefix ”la” È and the pro-
noun suffix ”haA” Aë . As depicted in Table 4, our RNN model (3-
layer B-LSTM) provides the best results(DER of
We also noticed a significant fraction of error 10.72%) compared to both SMT (DER of 33.15%)
words (34%) due to the presence of foreign words and CRF (DER of 20.25%) models.
in our corpus as in sample 4..
5 Conclusion
4.5 Comparison with State-of-art Systems
The absence of short vowels gives rise to a great
In this section, we compare our proposed RNN
ambiguity which influences the results of such
model with two other models, namely SMT and a
NLP applications. An outcome of this study
discriminative model as a sequence classification
was the development of RNN diacritic restoration
task based on CRFs (24). These two models were
model for Tunisian dialect. To the best of our
realized in our previous works in order to carry on
knowledge, this is the first work that deals with
the dialect restoration for the Tunisian dialect. To
the problem of Tunisian dialect diacritizers using
achieve this comparison, we employed the same
RNN.
dialectical corpus and evaluation metrics.
Concerning the first model, we regarded the di- In order to choose the best configuration of the
acritization problem as a simplified phrase-based RNN network, we did several preliminary experi-
SMT task. The source language is the undiacritic ments with different training options. These op-
text while the target language is the diacritic text. tions concern the hidden layer where we tested
The basic idea of SMT is to analyze automatically the impact of the change of the neural network
existing human sentences called parallel corpus in size and the topology on its performance. Sev-
order to build translation model. The alignment eral types of hidden layers are tested, ranging
from words without diacritics to words with dia- from one layer LSTM to multiple B-LSTM layers.
critics is a monotonic mapping. To do this, we The best accuracy is obtained when using the 3-
employed moses (21) a SMT tool.Word alignment layer B-LSTM model (DER of 10.72%). We com-
was done with GIZA++ (25). We implemented a pared our RNN diacritization model with two ma-
5-gram language model using the SRILM toolkit jor models, namely a SMT and CRF models (24).
(31). We decoded using Moses (21). These two models were realized in our previous
works in order to carry on the dialect restoration
In the second model, we decided to get the for the Tunisian dialect. During this comparison,
diacritical marks restoration by focusing on dia- we employed the same dialectical corpus and eval-
critization based on grammatical information. We uation metrics. About 9.53% DER improvement
intended to build dependency relations between of RNN model was achieved over the best reported
words and ”POS” tags and to perceive their effects CRF model.
on word diacritizations. In fact, we proposed to
scrutinize the integration of grammatical informa- We have two future plans for the diacritization
tion ”POS” for the diacritization with the aid of problems of Tunisian dialect. The first plan con-
Conditional Random Fields (CRF)(24). sists in expanding a rule-based diacritizer system
and integrating it into our RNN model in order
The following table reviews the accuracy results to ameliorate the outcomes. The second plan fo-
to restore diacritics automatically from our previ- cuses on providing morphological analysis of such
737
words to the RNN in order to achieve higher accu- Diab, M., Habash, N., Owen, R.: Conventional
racy. The presence of significant fraction of er- Orthography for Dialectal Arabic. In Proceedings
rors in complex words that contain prefixes, suf- of the Language Resources and Evaluation Con-
fixes, or both open up new perspectives for future ference , Istanbul,2012
research. Gers,F. : Long short-term memory in recurrent
Abandah, G., Graves, A., Al-Shagoor, B., Ara- neural networks, Ph.D. dissertation, Department
biyat, A., Jamour, F. and Al-Taee. M. 2015. Auto- of Computer Science, Swiss Federal Institute of
matic diacritization of arabic text using recurrent Technology, Lausanne, EPFL, Switzerland, 2001.
neural networks. International Journal on Docu- Graves, A., Mohamed, A., Hinton, G: Speech
ment Analysis and Recognition (IJDAR). recognition with deep recurrent neural networks,
Alotaibi, Y. A., Meftah, A. H. and Selouani. S.A. IEEE International Conference on Acoustics,
2013. Diacritization, Automatic Segmentation Speech, and Signal Processing, Canada,2013.
and Labeling for Levantine Arabic Speech. In Fashwan, A., Alansary, S. 2017. SHAKKIL: an
Digital Signal Processing and Signal Processing automatic diacritization system for modern stan-
Education Meeting (DSP/SPE). dard Arabic texts. The Third Arabic Natural Lan-
Alqudah,S., Abandah,G., Arabiyat, A., 2017.In- guage Processing Workshop (WANLP).
vestigating Hybrid Approaches for Arabic Text Habash, N., Shahrour, A., and Al-Khalil, M. 2016,
Diacritization with Recurrent Neural Networks. Exploiting Arabic Diacritization for High Qual-
2017 IEEE Jordan Conference on Applied Elec- ity Automatic Annotation, the Tenth International
trical Engineering and Computing Technologies. Conference on Language Resources and Evalua-
tion, LREC 2016.
Al-Badrashiny, M., Hawwari, A. and Diab, M.
2017. A Layered Language Model based Hybrid Habash, N. and Rambow, O. 2007. Arabic diacriti-
Approach to Automatic Full Diacritization of Ara- zation through full morphological tagging. The
bic. Third Arabic Natural Language Processing Conference of the North American Chapter of the
Workshop. Association for Computational Linguistics.
Hamed, O and Zesch, T. 2017. A Survey and
Ameur, M. Moulahoum, Y. and Guessoum, A.
Comparative Study of Arabic Diacritization Tools.
2015. Restoration of Arabic Diacritics Using a
Journal of Language Technology and Computa-
Multilevel Statistical Model. In IFIP International
tional Linguistics, volume 32, number 1.
Federation for Information Processing.
Harrat, S., Abbas, M., Meftouh, K., Smaili, K.,
Ayman, A. Z., Elmahdy, M., Husni, H. and Al
Bouzareah, E.N.S. and Loria, C. 2013. Diacritics
Jaam, J. 2016. Automatic diacritics restoration for
restoration for Arabic dialect texts. 14th Annual
Arabic text. International Journal of Computing &
Conference of the International Speech Commu-
Information Science..
nication.
Azmi, A. and Almajed, R. 2015. A survey of au- Hifny, Y. 2012. Higher order n-gram language
tomatic Arabic diacritization techniques. Natural models for Arabic diacritics restoration. In Pro-
Language Engineering, 21, pages:477495. ceedings of the 12th Conference on Language En-
Belinkov, Y. and Glass. J. 2015. Arabic diacritiza- gineering.
tion with recurrent neural networks. Conference Jarrar, M., Habash, N., Akra, D. and N. Zalmout,
on Empirical Methods in Natural Language Pro- N.: Building a Corpus for Palestinian Arabic: a
cessing. Preliminary Study :In Proceedings of the Arabic
Bouamor, H., Zaghouani, W., Diab, M., Obeid, O., Natural Language Processing Workshop, EMNLP,
Kemal, O., Ghoneim, M. and Hawwari, A. 2015. Doha,2014
A pilot study on Arabic multi-genre corpus dia- Koehn, P. Hoang; H. Birch, A. Callison-Burch, C.
critization annotation. The Second Workshop on Federico, M. and Bertoldi, N. 2007. Moses: Open
Arabic Natural Language Processing. Source Toolkit for Statistical Machine Translation.
Diab, M., Ghoneim, M. and Habash. N. 2007. ACL 2007, demonstration session.
Arabic Diacritization in the Context of Statistical Masmoudi, A., Habash, N., Khmekhem, M., Es-
Machine Translation. In Proceedings of MTSum- teve, Y. and Belguith, L.: Arabic Transliteration of
mit, Denmark. Romanized Tunisian Dialect Text: A Preliminary
738
Investigation, Computational Linguistics and In- Zitouni, I. and Sarikaya, R. 2009. Arabic Dia-
telligent Text Processing - 16th International Con- critic Restoration Approach Based on Maximum
ference, CICLing 2015, Cairo, Egypt, p.608-619, Entropy Models. In Journal of Computer Speech
(2015). and Language.
Masmoudi, A., Bougares,F., Ellouze, M., Esteve, Zribi, I., Ellouze, M., Belguith, L.H. and Blache,
Y., Belguith, L.: Automatic speech recognition P. 2015. Spoken Tunisian Arabic Corpus ”STAC”:
system for Tunisian dialect. Language Resources Transcription and Annotation. Res. Comput. Sci.
and Evaluation 52(1): 249-267 (2018). 90.
Masmoudi, A., Mdhaffer,S., Sellami, R. Belguith, Zribi I., Boujelben R., Masmoudi A., Ellouze M.,
L.: Automatic Diacritics Restoration for Tunisian Belguith L., Habash N.: A conventionnal Orthog-
Dialect. TALLIP2018: Transactions on Asian and raphy for Tunisian Arabic, LREC’2014, Reyk-
Low-Resource Language Information Processing. javik, Iceland, (2014).
Och, F. and Ney, H. 2003. A Systematic Com-
parison of Various Statistical Alignment Models.
Computational Linguistics, 29(1):19–52.
Rashwan, M., Al Sallab, A., Raafat, H. and Rafea,
A. Deep Learning Framework with Confused Sub
Set Resolution Architecture for Automatic Arabic
Diacritization. IEEE/ACM Transactions on Au-
dio, Speech, and Language Processing, 2015.
Rumelhart, D., Hinton,G., and Williams, R.
Learning representations by back-propagating er-
rors, Nature, no. 323, pp. 533 536, 1986.
Said, A., El-Sharqwi, M., Chalabi, A , and Ka-
mal, E. A hybrid approach for Arabic diacritiza-
tion, Application of Natural Language to Informa-
tion Systems, pp. 53-64, Jun 2013.
Shaalan, K., Abo Bakr, M. and Ziedan. I. 2009.
A hybrid approach for building Arabic diacritizer.
In Proceedings of the EACL 2009 Workshop on
Computational Approaches to Semitic Languages.
Shaalan, K., Abo Bakr, H. and Ziedan. I. 2008. A
statistical method for adding case ending diacrit-
ics for Arabic text. In Proceedings of Language
Engineering Conference.
Stolcke A. 2002. SRILM an Extensible Language
Modeling Toolkit. Proceedings of ICSLP.
Zaghouani, W., Bouamor, H., Hawwari, A., Diab,
M., Obeid, O., Ghoneim, M., Alqahtani, S and
Oflazer, K. 2016. Guidelines and framework for
a large-scale Arabic diacritized corpus. Proceed-
ings of the Tenth International Conference on Lan-
guage Resources and Evaluation (LREC 2016).
Zitouni, I., Sorensen, J. and Sarikaya, R. 2006.
Maximum entropy based restoration of Arabic di-
acritics. In Proceedings of the 21st International
Conference on Computational Linguistics and the
44th annual meeting of the Association for Com-
putational Linguistics.
739
Comparing MT Approaches for Text Normalization
740
Proceedings of Recent Advances in Natural Language Processing, pages 740–749,
Varna, Bulgaria, Sep 2–4 2019.
statistical machine translation (SMT) and neu- nel model. In this model, the goal is to find the
ral machine translation (NMT), on English and intended word w given a word x where the letters
Dutch parallel corpora with data coming from have been changed in some way. Correct words in
three genres (text messages, message board posts the text remain untouched. This model is prob-
and tweets). For SMT we explore the added value ably the most popular and successful approach
of varying background corpora for training the to spelling correction (Dutta et al., 2015; Goot,
language model. For NMT we have a look at 2015). Although spelling correction is mostly per-
data augmentation since the parallel datasets we formed on languages which are morphologically
are working with are limited in size. Our results simple and with a fairly strict word order, like En-
reveal that when relying on SMT to perform the glish, there has been some progress for normal-
normalization it is beneficial to use a background ization applied to other languages as well, such as
corpus that is close to the genre to be normalized. Russian (Sorokin, 2017) and French (Beaufort and
Regarding NMT, we find that the translations - or Roekhaut, 2010).
normalizations - coming out of this model are far
Text found in social media shares features with
from perfect and that for a low-resource language,
spoken language and the automatic speech recog-
like Dutch, adding additional training data works
nition metaphor exploits this similarity. This ap-
better than artificially augmenting the data.
proach starts by converting the input message into
In the following section, we discuss related a phone lattice, which is converted to a word lat-
work on text normalization. In section 3, we give tice using a phoneme-grapheme dictionary. Fi-
more information about our parallel data and de- nally, the word lattice is decoded by applying a
scribe the methodology we used to perform the language model to it and using a best-path algo-
SMT and NMT experiments. Section 4 gives an rithm to recover the most likely original word se-
overview of the results, whereas section 5 con- quence. This metaphor has mainly been merged
cludes this work and offers some prospects for fu- with the machine translation (infra) and spell
ture work. checking (supra) metaphors to improve the qual-
ity of the normalization. Kobus et al. (2008), for
2 Related Works example, incorporated ideas from speech recogni-
Previous research on UGC text normalization has tion to text message normalization and combined
been performed on diverse languages using dif- it with a machine translation system. Beaufort
ferent techniques ranging from hand-crafted rules and Roekhaut (2010); Xue et al. (2011) and Han
(Chua et al., 2018) to deep learning approaches and Baldwin (2011) also combined the automatic
(Ikeda et al., 2016; Sproat and Jaitly, 2016; Lusetti speech recognition approach with spell checking
et al., 2018). Kobus et al. (2008) introduced and machine translation techniques.
three metaphors to refer to these normalization ap- The machine translation metaphor treats so-
proaches: the spell checking, automatic speech cial media text as the source language and its
recognition and translation metaphors. normalized form as the target language. Sev-
In the spell checking metaphor, corrections eral works have tackled the problem of text nor-
from noisy to standard words occur at the word malization using this approach. Statistical Ma-
level. As in conventional spelling correction one chine Translation (SMT) models, especially those
has to deal with both non-word and real-word er- trained at the character-level, have proven highly
rors (Clark and Araki, 2011). The disadvantage effective for the task because they capture well
of this approach is that all non-standard words intra-word transformations. One of the first works
(NSWs) have to be represented in the dictionary following this approach was presented by Aw et al.
in order to obtain the corresponding normaliza- (2006). They adapted phrase-based SMT to the
tion. Therefore, the success of this kind of sys- task of normalizing English SMS producing mes-
tems highly depends on the dictionary coverage. sages that collated well with manually normalized
However, as UGC is a very generative language ones. Besides, they studied the impact of the nor-
and new variants of canonical words and phrases malization on the task of SMS translation, show-
appear constantly, it is very difficult and expensive ing that SMS normalization, as a preprocessing
to maintain a high coverage lexicon. Other works step of MT, can boost the translation performance.
have approached the problem using a noisy chan- Kaufmann (2010) used a two-step approach for
741
Source sentence Target sentence Translation
iz da muzieksgool vnavnd ? is dat muziekschool vanavond ? is that music school tonight? I
kwt da niemr . ik weet dat niet meer . don’t know that anymore.
wa is je msn k en e nieuwe msn wat is je msn ik heb een nieuwe what is your msn i have a new
omda k er nie meer op graal . msn omdat ik er niet meer op msn because i can’t get it any-
xxx geraak . xx more. xx
@renskedemaessc dm me je <user> doormail me je gsm- <user> mail me your cell-
gsmnummer eens ;-) nummer eens <emoji> phone number once <emoji>
Table 1: Source and target pairs as parallel data for a machine translation approach.
the normalization of English tweets: he first pre- lation from this representation. Some works on
processed the tweets to remove as much noise as text normalization have followed the same ap-
possible and then used a machine translation ap- proach. Ikeda et al. (2016) performed text nor-
proach to convert them into standard English. MT malization at the character level for Japanese text
approaches when used at the character level, also and proposed a method for data augmentation
have the advantage of being effective when small using hand-crafted rules. They proved that the
training data is provided, thanks to their small vo- use of the synthesised corpus improved the per-
cabulary size. De Clercq et al. (2013) proposed formance of Japanese text normalization. Man-
a phrase-based method to normalize Dutch UGC dal and Nanmaran (2018) presented an architec-
comprising various genres. They performed ex- ture for automatic normalization of phonetically
periments at several levels of granularity: charac- transliterated words to their standard forms in a
ter and word level. In a preprocessing step they code-mixed scenario improving the accuracy of a
handled emoticons, hyperlinks, hashtags and so pre-existing sentiment analysis system by 1.5%.
forth. Then they worked in two steps: first at the Lusetti et al. (2018) performed text normalization
word level and then at the character level. This ap- over Swiss German WhatsApp messages and com-
proach revealed good results across various genres pared it to a state-of-the-art SMT system. They
of UGC; however a high number of phonetic al- showed that integrating language models into an
ternations still remained unresolved. Schulz et al. encoder-decoder framework can reach and even
(2016) made a modification to the previous work improve the performance of character-level SMT
by combining the three metaphors (machine trans- methods for that language.
lation, spell checking and speech recognition) in In this work, we also consider the normalization
a multi-modular system and by using a novel ap- task as a MT problem and test both statistical and
proach for decoding. This led to an improve- neural machine translation. For SMT, we explore
ment in the selection of the best normalization op- the added value of varying background corpora for
tion. Furthermore, they showed a performance im- training the language model. For NMT, we inves-
provement of state-of-the-art NLP tools on UGC tigate whether we can overcome the limited data
data when normalization is used as a previous step. set size by using data augmentation.
Recently, neural networks have proven to out-
perform many state-of-the-art systems in sev-
eral NLP tasks (Young et al., 2018). The
3 Methodology
encoder-decoder model for recurrent neural net-
works (RNN) was developed in order to address Our objective is to go from noisy to standard text
the sequence-to-sequence nature of machine trans- and we tackle this normalization problem using a
lation and it obtains good results for this task Machine Translation (MT) approach. As in gen-
(Sutskever et al., 2014; Cho et al., 2014; Bah- eral MT, a translation model is trained on paral-
danau et al., 2014; Luong et al., 2015). The model lel data consisting of pairs (x, y) of source sen-
consists of two neural networks: an encoder and tences/words (= noisy text) and their correspond-
a decoder. The encoder extracts a fixed-length ing target equivalents (= standard). Table 1 lists
representation from a variable-length input sen- some examples of the noisy data we are dealing
tence, and the decoder generates a correct trans- with.
742
3.1 Parallel Corpora y1 , y2 , ..., yn . The language model P (y) is respon-
We relied on existing Dutch (Schulz et al., 2016) sible for the fluency of the sentence in the target
and English (De Clercq et al., 2014) corpora that language. W is the set of all target sentences.
were manually normalized1 (Table 2). Three gen- To achieve better context-sensitive source-
res were included for both languages: target mappings, traditional SMT systems rely on
Tweets (TWE) which were randomly selected phrase-level translation models. These models al-
for both languages from the social network. low to build a phrase table to store aligned phrase
Message board posts (SNS) which were in pairs in the source and target language. This is a
both languages sampled from the social network difficult task since one word in one language may
Netlog, which was a Belgian social networking correspond to several words in the other language.
website targeted at youngsters. However when translating from noisy to standard
Text messages (SMS) which were sampled text we can assume that most of the words have a
from the Flemish part of the SoNaR corpus one-to-one mapping. Figure 1 illustrates the archi-
(Treurniet et al., 2012) for the Dutch language and tecture of an SMT system.
from the NUS SMS corpus (Chen and Kan, 2013)
for English.
743
taken for Dutch, for which we used an in-house Neural systems, however, require huge amounts
subtitles dataset, Europarl, and the combination of of data in order to perform properly. The train-
both. ing data we have available for text normalization
amounts to only a few hundred sentences, as can
Corpus Sentences
English
be derived from Table 2. Moreover, manually an-
OPUS 22,512,649 notating more training is highly time-consuming.
Europarl 2,005,395 Under these conditions, we decided to experimen-
Combined OPUS+Europarl
Dutch
tally verify whether it is more beneficial to use a
Subtitles 8,056,693 data augmentation technique (step A) which possi-
Europarl 2,000,113 bly resolves the data scarcity problem (Saito et al.,
Combined OPUS+Europarl
2017) or to annotate more data (step B). We tested
Table 3: Size (expressed in sentences) of the this on the Dutch corpus and one particular genre,
monolingual corpora used for training our LMs. namely text messages (SMS). For step A, we aug-
mented the parallel data by duplicating monolin-
gual subtitles data on both the source and target
3.3 NMT Approach side. For step B, we sampled and manually anno-
tated ten thousand extra tokens from the Flemish
Neural Machine Translation incorporates the ad- part of the SoNaR corpus (Treurniet et al., 2012)2 .
vantages of newly developed deep learning ap- We relied on OpenNMT3 to train our encoder-
proaches into the task. Sequence-to-Sequence decoder model. OpenNMT is an open source
(seq2seq) models have been used for a variety of (MIT) initiative for neural machine translation and
NLP tasks including machine translation obtaining neural sequence modeling (Klein et al., 2017).
state-of-the-art results (Luong et al., 2015; Young The main system is implemented in the Lua/Torch
et al., 2018). In this approach both input and out- mathematical framework, and can easily be ex-
put sentences are going in and out of the model. tended using Torch’s internal standard neural net-
As described in the literature overview, the model work components. We used the version of the sys-
consists of two neural networks: an encoder and tem with the basic architecture which consists of
decoder (See Figure 2). The encoder extracts a an encoder using a simple LSTM recurrent neu-
fixed-length representation from a variable-length ral network. The decoder applies attention over
input sentence (A B C D), and the decoder gener- the source sequence and implements input feeding
ates a correct translation from this representation (Luong et al., 2015).
(X Y Z). In the figure <eos> marks the end of a
sentence. The encoder-decoder model is trained 3.4 Evaluation
on a parallel corpus consisting of aligned source For evaluating the results of the normalization we
sentences and their normalized forms (see Table calculated Word Error Rate (WER), a commonly
1). used machine translation evaluation metric. WER
is derived from the Levenshtein distance (Leven-
shtein, 1966), working at the word level instead
of the character level. It takes into account the
number of insertions (INS), deletions (DEL) and
substitutions (SUBS) that are needed to transform
the suggested string into the manually normalized
string. The metric is computed as follows:
IN S + DEL + SU BS
W ER =
N
where N in the number of words in the reference.
Figure 2: Encoder-decoder architecture. The Table 4 reports WER computed between the
light-color nodes represent the encoder and the original and target parallel sentence pairs that were
dark-color ones the decoder. Image taken from 2
Following the same annotation guidelines as Schulz et al.
Luong et al. (2015). (2016)
3
http://opennmt.net
744
Figure 3: Normalization results at the token level. The left chart presents the results on the English
datasets and the right one the results on the Dutch dataset.
used for training our models. The parallel data (Table 2) used for training
the translation model were divided using 80% for
Word Error Rate (%)
Genre English Dutch
training the model and 10% for development and
TWE 12.160 10.592 testing, respectively. The target sentences from the
SNS 15.400 21.390 training set were also added to the monolingual
SMS 17.190 25.130
corpus for training the language model.
Table 4: WER values (in percentage) at the sen- Despite several works reporting better results
tence level when using a character-level approach (De Clercq
et al., 2014; Lusetti et al., 2018) our experiments
WER values were calculated per sentence and revealed the best performance with SMT at the to-
averaged within the document. The higher the ken level. The bar charts in Figure 3 present the
value, the more operations are needed to obtain the results of SMT at the token level with the different
target sentence. Looking at the values, we again LMs.
notice that genres requiring the most and least nor-
Regarding the monolingual background cor-
malization are the text messages (SMS) and tweets
pora, we notice that Europarl leads to the best re-
(TWE), respectively.
sults for the tweets (TWE) genre which was ac-
4 Experiments tually the genre with the least noise (see Section
3.1). Our experiment shows WERs of 4% and
4.1 Varying Background Corpora for SMT 6.3% for English and Dutch, respectively. This
With the first round of experiments we want to result was to be expected as the word usage in
research the influence of varying the monolin- Europarl is mostly standard and therefore close to
gual data that are used to construct the language the word usage in the tweets. The same is true
models. We trained LMs at the character level for the genre comprising the most noise, i.e text
using unigrams and bigrams and at the token messages. The word usage in the Subtitles/OPUS
level. All LMs were built using the SRILM dataset is less standard and closer to spoken lan-
toolkit (Stolcke, 2002) with Witten-Bell discount- guage and, indeed, also in this case we obtained
ing which has proven to work well on small data a WER of 9.5% for English using OPUS, and a
sets (Tiedemann, 2012a; Mahmudul Hasan et al., WER of 12% for Dutch using a combination of
2012; De Clercq et al., 2013). To evaluate the per- the Subtitles dataset and Europarl.
formance of each LM, Word Error Rate (WER) In addition, we also computed the number of in-
was calculated. sertions (INS), deletions (DEL) and substitutions
745
(SUBS) in the original sentence pairs (Ori) and af- 4.3 NMT Approach to UGC Normalization
ter normalization (Norm) (Table 5). As we explained before, neural approaches have
obtained state-of-the-art results for the task of Ma-
English
INS DEL SUBS
chine Translation. Neural approaches however,
Genre Ori Norm Ori Norm Ori Norm are well-known to require big amounts of paral-
TWE 91 36 15 30 34 22 lel data in order to perform properly. Especially
SNS 354 64 66 43 109 62
SMS 377 63 44 57 135 54
for Dutch, which can be considered a low-resource
Dutch language, it is difficult to find freely available par-
TWE 22 19 0 5 39 45 allel data and annotating new data is both money
SNS 386 159 18 36 76 47
SMS 0 0 2 3 94 85
and time-consuming.
Under these conditions we decided to experi-
Table 5: Number of operations required before mentally verify whether it is more beneficial to
(Ori) and after (Norm) normalization. use a data augmentation technique (step A) which
possibly resolves the data scarcity problem (Saito
Ideally, the number of operations after nor- et al., 2017) or annotate more data (step B). We
malization should be reduced to zero. As can tested this on the Dutch corpus and one particular
be derived from the table many operations were genre, the most noisy one, namely text messages
strongly reduced; however, many cases still need (SMS).
to be solved. We will have a closer look at some For Step A, our idea was to make use of the
of these cases in the next section. monolingual subtitles corpus in both sides of the
parallel data in order to augment the number of
sentences available for training our model. Doing
4.2 Error Analysis of the SMT Results
this, we obtain a bigger dataset consisting of the
The number of remaining insertions is mostly Dutch SMS parallel corpus and the Dutch Subti-
linked to the problem of abbreviation expansions. tles dataset which is duplicated in the source and
Very common abbreviations like lol or omg are al- target data. That is a total of 8,057,334 parallel
ways corrected, whereas others like r.e. and p.e. sentences for training.
for religious and physical education or cum on den
src sent. wa gaat je doen ? xxx
for come on then are not corrected since they never norm sent. wat gaat je doen ? xxx
appear in the training data. src sent. oeiiii misterieus <emoji> xxx
norm sent. oeiiii misterieus <emoji> xxx
When we consider the deletions, we can ob- src sent. dne dvd vn is ni goe ze . ge kunt nx zien .
serve that flooding or repetition of characters is of- mt betale . x
ten not solved with SMT. For example, tokens like norm sent. het dvd hem is niet niet het maar wat wat
in ik . niet betale . x x x x x
okkk, awwwww or sentences like immaaa dooiin
fiiine ! remained unchanged. A straightforward Table 6: Original (src) and predicted (norm) sen-
way to overcome this problem could be to reduce tences using the NMT approach.
the number of repetitions to two or three in some
cases as a pre-normalization step. The second fac- Unfortunately, as can be derived from the ta-
tor that affects the number of deletions is the hy- ble above, results following this approach are very
pernormalization of some words. This leads to an poor. It is common in the output to find repeti-
increase in the number of operations since some- tion of words like in the third sentence (niet, wat
times we will have to perform more deletion oper- and x). Besides, some sentences that needed nor-
ations on the predicted sentences than on the orig- malization like the second sentence in the table,
inal ones (these instances are indicated in bold in were not normalized at all. For example, the words
Table 5). This is for example the case with the wat and niet in the first and last sentence respec-
name al which was incorrectly normalized to all tively, were correctly normalized. These results
or the normalization of i can ’t really think... into may have been determined to a great extent by the
i can not really not think. These problems also unbalance in the parallel sentences. However, we
affect the number of substitutions. In general, could see a slight improvement compared to the
we also notice that the normalization of Dutch results using only the small parallel data in this
presents a higher number of errors. architecture. For that case, the system output con-
746
sisted of sentences of the type <emoji> , de , , , 5 Conclusions and Future Work
. . . <emoji>. These are random repetitions of
In this article, we have presented two different
the most represented tokens like ik (I in English),
approaches to text normalization of social media
punctuation marks or <emoji> labels.
text: statistical and neural machine translation. We
In order to corroborate our other hypothesis, we applied text normalization to English and Dutch
collected and manually normalized more data for text from different genres: text messages, message
step B. In order to check how this system works at board posts and tweets. Best results were achieved
different levels of granularity, we also performed at the token level for all genres and for both SMT
experiments using bigram and unigrams of char- and NMT.
acters. Regarding the results, also for NMT the For the SMT experiments, regarding the differ-
results are better at the word level than at the char- ent corpora that were used to construct the LM,
acter level, with WERs of 15% instead of 29% and we found that Europarl gave the best results for
26% for bigram and unigram, respectively. the least noisy genre (tweets). The same is true
for the noisiest genre (text messages). Considering
4.4 Error Analysis of the NMT Results our results, it seems to be important to make vari-
Using the new data configuration the system is ca- ations in the background data for building the LM,
pable to correctly translate sentences like the first depending on the amount of noise and vocabulary
one in Table 7. that is present in the genre. With respect to the re-
maining errors we believe that following a modu-
src sent. aahn , ok ma cva dan kzal dan wel wa lar approach instead of only using SMT could lead
zoeken xp merci eh x to a much better performance.
norm sent. ah , oké maar ça va dan ik zal dan wel wat
zoeken <emoji> merci h x Our NMT approach performs poorly due to the
tgt sent. ah , oké maar ça va dan ik zal dan wel wat scarcity of the data, although we did find that for
zoeken <emoji> merci h x a low-resource language like Dutch adding addi-
src sent. zal dan eentje v mezelf sturen . zorgen we
morgenavond dan voor verrassing v tional training data works better than artificially
kareltje ? augmenting the data. The data augmentation tech-
norm sent. zal dan eentje van mag sturen . zorgen we nique used, however, was very basic and we be-
morgenavond dan voor cocktail van
droomt ? lieve that other techniques could lead to better
tgt sent. zal dan eentje van mezelf sturen . zorgen results, such as hand-crafted rules for the pro-
we morgenavond dan voor verrassing voor duction of abbreviations or the use of previously
kareltje ?
trained embedding in order to build similar sen-
Table 7: Original (src), predicted (norm) and tences helping to generalize better.
target (tgt) sentences using the NMT approach Exploring those other data augmentation tech-
trained on the extended dataset. niques is a first avenue for future work. Besides
we also want to test the benefits of the integra-
However, the system still produces a large num- tion of a neural LM in the encoder-decoder model
ber of odd normalizations. In the second sen- to help with the translation of out-of-vocabulary
tence in Table 7, for example, only the bold words words.
should have been normalized. However, only one
of those two words was correctly normalized, the
References
other one was normalized but not into its correct
form. On the other hand, the system also produces AiTi Aw, Min Zhang, Juan Xiao, and Jian Su. 2006.
A phrase-based statistical model for SMS text nor-
odd translations of words that already were in their malization. Proceedings of the COLING/ACL on
standard form. For example the word mezelf is Main conference poster sessions - (July):33–40.
changed to mag and we got cocktail van droomt https://doi.org/10.1109/TCST.2005.854339.
instead of the desired normalization, ie. verrass-
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben-
ing voor kareltje. gio. 2014. Neural Machine Translation by Jointly
In general, while the results using this approach Learning to Align and Translate. arXiv preprint
are very poor, the experiments revealed that hav- pages 1–15. http://arxiv.org/abs/1409.0473.
ing a bigger parallel training corpus could improve Richard Beaufort and Sophie Roekhaut. 2010. A
the performance of this system. hybrid rule/model-based finite-state framework
747
for normalizing SMS messages. Proceedings ation for Computational Linguistics: Human Lan-
of the 48th Annual Meeting of the Associa- guage Technologies, Proceedings of the Main Con-
tion for Computational Linguistics. Association ference (June):359–369.
for Computational Linguistics 1(July):770–779.
https://doi.org/10.1097/ICO.0b013e318245c02a. Rob Van Der Goot. 2015. Normalizing Social Me-
dia Texts by Combining Word Embeddings and
Su Lin Blodgett, Green Lisa, and OĆonnor Brendan. Edit Distances in a Random Forest Regressor. In
2016. Demographic Dialectal Variation in Social Normalisation and Analysis of Social Media Texts
Media: A Case Study of African-American English. (NormSoMe). 1.
arXiv preprint .
Bo Han and Timothy Baldwin. 2011. Lexical Nor-
Tao Chen and Min Yen Kan. 2013. Creating a live, malisation of Short Text Messages : Makn Sens a
public short message service corpus: The NUS SMS #twitter. Proceedings of the 49th Annual Meeting of
corpus, volume 47. https://doi.org/10.1007/s10579- the Association for Computational Linguistics pages
012-9197-9. 368–378.
Kyunghyun Cho, Bart van Merrienboer, Caglar Gul- Taishi Ikeda, Hiroyuki Shindo, and Yuji Matsumoto.
cehre, Dzmitry Bahdanau, Fethi Bougares, Hol- 2016. Japanese Text Normalization with Encoder-
ger Schwenk, and Yoshua Bengio. 2014. Learn- Decoder Model. Proceedings of the 2nd Workshop
ing Phrase Representations using RNN Encoder- on Noisy User-generated Text (WNUT) pages 129–
Decoder for Statistical Machine Translation. Inter- 137.
national Conference on Learning Representations
ICLR http://arxiv.org/abs/1406.1078. Max Kaufmann. 2010. Syntactic Normalization of
Twitter Messages. International conference on nat-
Mason Chua, Daan Van Esch, Noah Coccaro, Eunjoon ural language processing 2:1–7.
Cho, Sujeet Bhandari, and Libin Jia. 2018. Text
Normalization Infrastructure that Scales to Hun- Jinyun Ke, Tao Gong, and William S-y Wang. 2008.
dreds of Language Varieties. In Proceedings of the Language Change and Social Networks. Communi-
Eleventh International Conference on Language cations in Computational Physics 3(4):935–949.
Resources and Evaluation (LREC). European
Language Resource Association, Miyazaki, Guillaume Klein, Yoon Kim, Yuntian Deng, Josep
Japan, pages 1353–1356. http://www.lrec- Crego, Jean Senellart, and Alexander M. Rush.
conf.org/proceedings/lrec2018/pdf/8883.pdf. 2017. OpenNMT: Open-source Toolkit for
Neural Machine Translation. arXiv preprint
Eleanor Clark and Kenji Araki. 2011. Text nor- http://arxiv.org/abs/1709.03815.
malization in social media: Progress, prob-
lems and applications for a pre-processing Catherine Kobus, Francois Yvon, and Geéraldine
system of casual English. Procedia - Social Damnati. 2008. Normalizing SMS: are two
and Behavioral Sciences 27(Pacling):2–11. metaphors better than one? Proceedings of
https://doi.org/10.1016/j.sbspro.2011.10.577. the 22nd International Conference on Com-
putational Linguistics 1(August):441–448.
Orphée De Clercq, Sarah Schulz, Bart Desmet, and https://doi.org/10.14288/1.0064682.
Véronique Hoste. 2014. Towards Shared Datasets
for Normalization Research. Proceedings of the Philipp Koehn, Wade Shen, Marcello Federico,
Ninth International Conference on Language Re- Nicola Bertoldi, Chris Callison-Burch, Brooke
sources and Evaluation (LREC’14) pages 1218– Cowan, Chris Dyer, Hieu Hoang, Ondrej Bo-
1223. jar, Richard Zens, Alexandra Constantin, Evan
Herbst, and Christine Moran. 2006. Moses:
Orphée De Clercq, Sarah Schulz, Bart Desmet, Els Open Source Toolkit for Statistical Machine Trans-
Lefever, and Véronique Hoste. 2013. Normaliza- lation. Proceedings of ACL (June):177–180.
tion of Dutch User-Generated Content. Proceed- https://doi.org/10.3115/1557769.1557821.
ings of the 9th International Conference on Recent
Advances in Natural Language Processing (RANLP Vladimir I. Levenshtein. 1966. Binary codes capa-
2013) pages 179–188. ble of correcting deletions. Soviet physics doklady
10(8):707–710.
Sukanya Dutta, Tista Saha, Somnath Banerjee, and
Sudip Kumar Naskar. 2015. Text Normalization in Minh-Thang Luong, Hieu Pham, and Christopher D.
Social Media. In 2015 IEEE 2nd International Con- Manning. 2015. Effective Approaches to Attention-
ference on Recent Trends in Information Systems based Neural Machine Translation. arXiv preprint
(ReTIS). IEEE, Kolkata, India, c, pages 378–382. http://arxiv.org/abs/1508.04025.
https://doi.org/0.1109/ReTIS.2015.7232908.
Massimo Lusetti, Tatyana Ruzsics, Anne Göhring,
Jacob Eisenstein. 2013. What to do about bad language Tanja Samardi Samardžic, and Elisabeth Stark.
on the internet. NAACL HLT 2013 - 2013 Confer- 2018. Encoder-Decoder Methods for Text
ence of the North American Chapter of the Associ- Normalization. In Proceedings of the Fifth
748
Workshop on NLP for Similar Languages, Va- Ilya Sutskever, Oriol Vinyals, and Quoc V. Le.
rieties and Dialects (VarDial). pages 18–28. 2014. Sequence to Sequence Learning with
http://www.aclweb.org/anthology/W18-3902. Neural Networks. In Advances in neural in-
formation processing systems, pages 3104–3112.
A. S. M. Mahmudul Hasan, Saria Islam, and M. Mah- http://arxiv.org/abs/1409.3215.
mudul Rahman. 2012. A Comparative Study of Wit-
ten Bell and Kneser-Ney Smoothing Methods for Jörg Tiedemann. 2012a. Character-Based Pivot Trans-
Statistical Machine Translation. Journal of Infor- lation for Under-Resourced Languages and Do-
mation Technology (JIT) 1(June):1–6. mains. Proceedings of the 13th Conference of the
European Chapter of the Association for Computa-
Soumil Mandal and Karthick Nanmaran. 2018. tional Linguistics pages 141–151.
Normalization of Transliterated Words in Code-
Mixed Data Using Seq2Seq Model & Leven- Jörg Tiedemann. 2012b. Parallel Data, Tools and In-
shtein Distance. In Proceedings of the 2018 terfaces in OPUS. Proceedings of the Eight Interna-
EMNLP Workshop W-NUT: The 4th Workshop tional Conference on Language Resources and Eval-
on Noisy User-generated Text. Association for uation (LREC’12) pages 2214–2218.
Computational Linguistics, Brussels, Belgium,
pages 49–53. https://en.wikipedia.org/wiki/ISO Maaske Treurniet, Henk van den Heuvel, Nelleke Oost-
http://arxiv.org/abs/1805.08701. dijk, and Orphée De Clercq. 2012. Collection of
a corpus of Dutch SMS. Proceedings of the Eight
Itsumi Saito, Jun Suzuki, Kyosuke Nishida, and Conference of International Language Resources
Kugatsu Sadamitsu. 2017. Improving Neural and Evaluation. pages 2268–2273.
Text Normalization with Data Augmentation at
Character- and Morphological Levels. Proceed- Cynthia Van Hee, Marjan Van De Kauter, Orphe
ings of the The 8th International Joint Conference De Clercq, Els Lefever, Bart Desmet, and Vronique
on Natural Language Processing pages 257–262. Hoste. 2017. Noise or music? Investigating the use-
http://aclweb.org/anthology/I17-2044. fulness of normalisation for robust sentiment analy-
sis on social media data. Revue Traitement Automa-
Sarah Schulz, Guy De Pauw, Orphée De Clercq, Bart tique des Langues 58(1):63–87.
Desmet, Véronique Hoste, Walter Daelemans, and
Lieve Macken. 2016. Multimodular Text Normal- Reinhild Vandekerckhove and Judith Nobels. 2010.
ization of Dutch User-Generated Content. ACM Code eclecticism : Linguistic variation and code al-
Transactions on Intelligent Systems and Technology ternation in the chat language of Flemish teenagers.
7(4):1–22. https://doi.org/10.1145/2850422. Journal of Sociolinguistics 14(5):657–677.
Andrew H. Schwartz, Johannes C. Eichstaedt, Mar- Zhenzhen Xue, Dawei Yin, and Bd Davison. 2011.
garet L. Kern, Lukasz Dziurzynski, Stephanie M. Normalizing Microtext. Analyzing Microtext pages
Ramones, Megha Agrawal, Achal Shah, Michal 74–79.
Kosinski, David Stillwell, Martin E. P. Seligman,
and Lyle H. Ungar. 2013. Personality, Gender, Tom Young, Devamanyu Hazarika, Soujanya Poria,
and Age in the Language of Social Media : The and Erik Cambria. 2018. Recent trends in deep
Open-Vocabulary Approach. PLoS ONE 8(9). learning based natural language processing. ieee
https://doi.org/10.1371/journal.pone.0073791. Computational intelligenCe magazine 13(3):55–75.
https://ieeexplore.ieee.org/abstract/document/8416973/.
Alexey Sorokin. 2017. Spelling Correction for Mor-
phologically Rich Language: a Case Study of Rus-
sian. Proceedings of the 6th Workshop on Balto-
Slavic Natural Language Processing (April):45–53.
https://doi.org/10.18653/v1/w17-1408.
Richard Sproat, Alan W. Black, Stanley Chen, Shankar
Kumar, Mari Ostendorf, and Christopher Richards.
2001. Normalization of non-standard words.
Computer Speech and Language 15(3):287–333.
https://doi.org/10.1006/csla.2001.0169.
Richard Sproat and Navdeep Jaitly. 2016. RNN Ap-
proaches to Text Normalization: A Challenge. Com-
puting Research Repository (CoRR) abs/1611.0.
http://arxiv.org/abs/1611.00068.
Andreas Stolcke. 2002. SRILM - an extensible
language modeling toolkit. In 7th Interna-
tional Conference on Spoken Language Pro-
cessing.. Denver, Colorado, pages 901–904.
https://doi.org/10.1.1.157.2429.
749
Sentiment and Emotion Based Text
Representation for Fake Reviews Detection
750
Proceedings of Recent Advances in Natural Language Processing, pages 750–757,
Varna, Bulgaria, Sep 2–4 2019.
ing gamification dynamics. This makes it impor- They trained three machine learning models i.e.
tant for data-driven FR solutions to rely on more SVM, Logistic Regression, and Multi-layer Per-
generic or higher-level data representations rather ceptron and found that LDA+Logistic Regression
than simple lexical ones based on words, phrases and LDA+Multi-layer Perceptron performed bet-
and sentences. This is because FR filters using ter with 81.3% of accuracy.
higher-level generic features may naturally be ex- With representations being only a means to en-
pected to be more robust and resistant to sim- able better FR identification, it is useful to briefly
ple workarounds by FR authors such as word and outline the classification techniques that have been
phrase replacements. Further, higher-level fea- employed for FR detection. Ott et al., (2011) used
tures may have limited volatility across domains; word n-gram features in combination with a SVM
thus, FR detection methods based on them may be classifier. Banerjee and Chua (2014) employed a
more transferable across domains. Logistic Regression classifier over POS tags and
In this paper, we evaluate the effectiveness of writing style features (e.g., tense of words) for
emotion and sentiment based representations for FR detection. Algur et al., (2010) explored a
the task of building machine learning models for similarity-oriented method for FR detection over
FR detection. In particular, we illustrate that im- domain-specific product features.
proved data representations can be achieved by As mentioned earlier, our representations are
leveraging a plurality of emotion and sentiment centred on emotion and sentiment based features.
extraction methods, as well as by estimating emo- There has been very little prior work on using such
tions and sentiments on a part-by-part basis by features for FR detection. An early work in senti-
segmenting the reviews. We illustrate the im- ment analysis for FR detection was conducted by
proved effectiveness of multiple emotion and sen- Peng and Zhong (2014), whereas (K et al., 2019)
timent features as well as review-segmented fea- explore utility of emotions in health fake news de-
tures by evaluating over real-world datasets. tection. Peng and Zhong (2014) chose SentiWord-
Net and MPQA lexicons and analysed sentiment
2 Related Work on review and product features. In our experiment,
we used IBM, Afinn, SenticNet, and Biu Liu lexi-
Representation learning focuses on developing a cons. To our knowledge, this is the first study de-
more instructive feature set for training a classi- tecting FR by means of combination emotion and
fication model that helps to boost the FR detec- sentiment analysis. Taking cue from the previous
tion process (Li et al., 2017; Yilmaz and Durahim, work of FR detection, we use Random Forest clas-
2018). Within past research, diverse features se- sifier, in our experiments.
lection methods have been employed to detect FR.
These may be divided into two classes: review- 3 Methodology
centric and reviewer-centric features. Reviewer- In this section we describe our proposed approach
centric features are related to the reviewer’s be- to online FR detection using emotion and senti-
haviour (Fontanarava et al., 2017) rather than ment based text representation.
the review itself. Those features include tex-
tual features, rating features, and temporal fea- 3.1 Emotion and Sentiment Analysis
tures. Review-centric features are derived from For the purpose of sentiment and emotion anal-
the content of a review. Commonly used review- ysis, we apply three different sentiment lexicons
centric features include Bag-of-words, TF-IDF and one emotion analysis API.
(Term-frequency inverse-document- frequency),
POS (part of speech) tags, word n-grams (Ahmed • IBM Watson Natural Language Understand-
et al., 2018), and word embedding vectors (e.g. ing. Natural Language Understanding (NLU)2
Word2vec, Doc2vec) (Krishnamurthy et al., 2018; is a collection of APIs that offer text analy-
Yilmaz and Durahim, 2018). A recent study by Jia sis through natural language processing. One
et al. (2018) explored the application of linguis- of the feature of IBM Watson NLU is emo-
tic features to distinguish between fake and non- tion analysis. The API takes a text as
fake reviews. They used Yelp filter dataset in their an input and returns the category which the
study and applied Term Frequency, Word2vec, and 2
https://www.ibm.com/services/natural-language-
Latent Topic Distribution for data representation. understanding/
751
text belongs to, stored in a list variable: < 3.2.1 Sentiment Based Representation
KeyV alueP air < String, Double >> e.g. The process of constructing sentiment based rep-
”emotion” : {”sadness”:0.336228}. Each item resentation of a review is presented in Algorithm
in the list contains the category (emotion) name 1. We first split a review into P segments,
and the categorization score. IBM Watson NLU each one containing the same number of sen-
can detect five emotions: anger, disgust, fear, tences. For example, if P=4, then we split a
joy, and sadness. For example, for an an input ’I review into 4 segments. For each segment we
love apples! I don’t like oranges’, the NLU API identify all positive and all negative words us-
returns (sadness: 0.32665, joy: 0.563273, fear: ing the lexicons. In the next step, all positive
0.033387, disgust: 0.022637, anger: 0.041796). sentiment values and all negative sentiment val-
• SenticNet lexicon. SenticNet3 performs tasks ues within the segment are accumulated together.
such as polarity detection and emotion recog- In the case of AFINN and SenticNet, all posi-
nition. Instead of merely relying on word co- tive and negative values are summed in each seg-
occurrence frequencies, it leverages semantics ment. For Biu Liu lexicon, all positive and all
and linguistics. This lexicon contains a list of negative words are counted. Following this, the
words with their polarity and intensity values. segment is represented by a two dimensional vec-
The intensity is a float number between -1 and tor [pos(si ), neg(si )], where pos(si ) and neg(si )
+1. For example, according to the SenticNet represent the accumulated/counted positive and
lexicon ’abandoned’ is a negative word with in- negative sentiment values. Finally, all P vec-
tensity of -0.85. Each word in the lexicon is tors (one generated for each segment) are concate-
assigned with only one polarity and intensity nated. The concatenated vector is returned as the
value. sentiment representation of the entire review. The
• AFINN lexicon. AFINN4 lexicon is a list of process looks the same for all sentiment lexicons.
English terms rated with valence on a scale -5
(negative) and +5 (positive). This lexicon has Algorithm 1 Sentiment Based Representation
been manually labelled by Finn Årup Nielsen Input: Review R, number of segments P , senti-
(2011). AFINN provides two versions of lexi- ment lexicon L
con: the newest version AFINN-111 with 2477 Output: Sentiment representation of R
words and phrases and AFINN-96 with 1468 1: Split R into P equal segments s1 , . . . , sP
unique words and phrases on 1480 lines. Our 2: for all s1 , . . . , sP do
experiment use AFINN-111 as it is the most up- 3: Tokenise si into set of words W
to-date version. 4: Retrieve sentiment values for all words in
• Biu Liu lexicon. Biu Liu5 lexicon consists of W using L
6789 words including 2006 positive and 4783 5: Accumulate all positive sentiment values in
negative words (Hu and Liu, 2004). This lexi- W as pos(si )
con does not provide any sentiment scores and 6: Accumulate all negative sentiment values in
only provides positive/negative labels. W as neg(si )
7: vi = [pos(si ), neg(si )]
8: end for
3.2 Representation Learning
9: v(R) := [v1 , . . . , vP ]
In this work we explore whether sentiment and 10: return v(R)
emotions extracted from a review can be used to
train machine learning models for distinguishing
3.2.2 Emotion Based Representation
between fake and non-fake reviews. We perform
the sentiment/emotion analysis with different lev- The process of generating emotion based repre-
els of granularity on a part-by-part basis by seg- sentation is presented in Algorithm 2. As in the
menting the reviews. case of the sentiment based representation, a re-
view is first divided in P segments. All sentences
3
in each segment is then passed to the IBM Watson
https://sentic.net/
4 API. As the output we obtain vector with the five
https://pypi.org/project/afinn/
5
http://www.cs.uic.edu/l̃iub/FBS/opinion-lexicon- emotions’ scores. Finally, the emotion vectors ob-
English.rar tained for all the segments are concatenated. The
752
output vector is returned as the emotion represen- Algorithm 4 Combined sentiment and Emotion
tation of the entire review. Based Representation
Input: Review R, number of segments P , senti-
Algorithm 2 Emotion based Representation ment lexicon Ls , emotion lexicon Le
Input: Review R, number of segments P , emo- Output: Vector representation of R
tion lexicon L 1: Split R into P equal segments s1 , . . . , sP
Output: Emotion representation of R 2: for all s1 , . . . , sP do
1: Split R into P equal segments s1 , . . . , sP 3: Get sentiment representation Vs (R) apply-
2: for all s1 , . . . , sP do ing Algorithm 1 with R, P and Ls
3: Get vector vi with emotions scores from L 4: Get emotion representation Ve (R) applying
4: end for Algorithm 2 with R, P and Le
5: v(R) := [v1 , . . . , vP ] 5: vi = [Vs (R), Ve (R)]
6: return v(R) 6: end for
7: v(R) := [v1 , . . . , vP ]
8: return v(R)
753
since sparse text does not allow to identify emo- better the prediction performance. It can also be
tions and sentiments well. Table 2 demonstrates noted that the multi-segment based representation
the statistics of the datasets after filtering. tends to perform better than when a single segmen-
tation is applied. The only exception is the Biu Liu
Dataset Non-fake Fake lexicon, which for Yelp, Zip, and Ott obtained the
YELP ZIP 528019 80439 best results for P = 1.
YELP NYC 322097 36860
Ott 800 800 Lexicon P=1 P=2 P=3 P=4 P1-4
IBM 0.570 0.584 0.589 0.584 0.597
SenticNet 0.506 0.510 0.522 0.523 0.524
Table 1: Statistics of the datasets Biu Liu 0.574 0.540 0.547 0.558 0.557
AFINN 0.550 0.542 0.549 0.555 0.563
Dataset Non-fake Fake Table 3: RF’s F-measure over Yelp ZIP dataset.
YELP ZIP 170261 15108
YELP NYC 105080 6185
Ott 340 270 Lexicon P=1 P=2 P=3 P=4 P1-4
IBM 0.554 0.569 0.578 0.569 0.584
SenticNet 0.511 0.520 0.523 0.525 0.526
Table 2: Statistics of the datasets after filtering Biu Liu 0.546 0.523 0.543 0.543 0.555
AFINN 0.524 0.529 0.541 0.544 0.557
Learning. As the machine learning algorithm we
used Random Forest (RF) given that it was re- Table 4: RF’s F-measure over Yelp NYC dataset.
ported as one of the most effective in FR detection
(Chowdhary and Pandit, 2018; Saumya and Singh,
2018; Viviani and Pasi, 2017). However, any other Lexicon P=1 P=2 P=3 P=4 P1-4
IBM 0.620 0.605 0.543 0.533 0.590
learning algorithm can be applied instead. We set SenticNet 0.533 0.529 0.483 0.525 0.570
n estimator=100 and random state=42 for the RF Biu Liu 0.618 0.561 0.592 0.576 0.600
parameter. All the experiments are performed with AFINN 0.523 0.560 0.580 0.545 0.600
5-fold cross-validation and the prediction perfor-
mance is evaluated with application of F-measure. Table 5: RF’s F-measure over Ott dataset.
Given the very high class imbalance in the Yelp
NYC and Yelp Zip, we randomly select number
4.3 Sentiment vs. Emotion
of non-fake reviews equal to the number of FR in
order to balance the training data. In this section we compare the results obtained by
RF applied with the sentiment and the emotion
4.2 Sentiment and Emotion Granularity based representations of data. We can see from
In this section we investigate what level of gran- Tables 3-5 that IBM emotion lexicon obtained the
ularity in terms of sentiment and emotion is the best performance in comparison to the three senti-
most representative for FR detection. Tables 3-5 ment lexicons in the Yelp Zip and NYC datasets.
demonstrate the F-measure obtained by RF with This may be considered unsurprising since emo-
each of the datasets and sentiment and emotion tions provide more fine grained information for the
based representations for reviews. For the param- classifiers to work with. For the Ott dataset, Biu
eter P we used values from 1 to 4. For each ta- Liu and AFINN lexicon obtained better results for
ble, the first row represents results obtained by RF some of the greater values of P .
applied with the emotion based representation ob- In order to perform better comparison be-
tained with the IBM Watson API. The three bot- tween the sentiment and emotion based rep-
tom columns contain results obtained for the sen- resentations we calculated average of the re-
timent based representation generated with each of sults obtained for each of the granularity levels:
the three sentiment lexicons. Each column refers P 1, P 2, P 3, P 4, P 1 − 4. The results are demon-
to a different value of parameter P = 1 × 4. The strated in Figure 1. We can observe from the
last column presents results obtained for multi- graphs that the IBM emotion lexicon performs sig-
segment based representation. nificantly better than any of the other sentiment
We can observe from the tables that in the ma- lexicons apart from the Ott dataset where it is out-
jority of cases, the higher the granularity (P ) the performed by the Biu Liu lexicon.
754
Figure 2: Combined sentiment-emotion vs. emo-
Figure 1: Average F-measure obtained for all val- tion representation learning for Zip
ues of P
755
Information Management, ICIM 2018, IEEE,
Acknowledgments 2018, pp. 280–83,
doi:10.1109/INFOMAN.2018.8392850.
The authors thank LPDP (Indonesia Endowment
Fund for Education) for funding this research. Jindal, Nitin, and Bing Liu. “Opinion Spam and
Analysis.” Proceedings of the 2008 International
Conference on Web Search and Data Mining,
ACM, 2008, pp. 219–30.
References
Krishnamurthy, Gangeshwar, et al. “A Deep Learning
Ahmed, Hadeer, et al. “Detecting Opinion Spams and Approach for Multimodal Deception Detection.”
Fake News Using Text Classification.” Security ArXiv Preprint ArXiv:1803.00344, 2018.
and Privacy, vol. 1, no. 1, 2017, p. e9,
doi:10.1002/spy2.9. Kumar, Abhinav, et al. “Spotting Opinion Spammers
Using Behavioral Footprints.” Proceedings of
Algur, Siddu P., et al. “Conceptual Level Similarity the 19th ACM SIGKDD International
Measure Based Review Spam Detection.” Conference on Knowledge Discovery and Data
Proceedings of the 2010 International Mining, ACM, 2013,
Conference on Signal and Image Processing, doi:10.1145/2487575.2487580.
ICSIP 2010, IEEE, 2010, pp. 416–23,
doi:10.1109/ICSIP.2010.5697509. Li, Luyang, et al. “Document Representation and
Feature Combination for Deceptive Spam
Anoop, K., et al. “Emotion Cognizance Improves Fake Review Detection.” Neurocomputing, vol. 254,
News Identification.” ArXiv Preprint Elsevier B.V., 2017, pp. 1339–51,
ArXiv:1906.10365, 2019. doi:10.1016/j.neucom.2016.10.080.
Bandhakavi, Anil, et al. “Lexicon Generation for Luca, Michael, and Georgios Zervas. “Fake It till You
Emotion Detection from Text.” IEEE Intelligent Make It: Reputation.” Competition, and Yelp
Systems, vol. 32, no. 1, IEEE, 2017, pp. 102–08, Review Fraud., SSRN Electronic Journal, 2016.
doi:10.1109/MIS.2017.22.
Nielsen, Finn Årup. “A New ANEW: Evaluation of a
Banerjee, Snehasish, and Alton Y. K. Chua. Word List for Sentiment Analysis in
“Applauses in Hotel Reviews: Genuine or Microblogs.” CEUR Workshop Proceedings,
Deceptive?” Proceedings of 2014 Science and vol. 718, 2011, pp. 93–98,
Information Conference, SAI 2014, The Science doi:10.1016/j.knosys.2015.06.015.
and Information (SAI) Organization, 2014, pp.
938–42, doi:10.1109/SAI.2014.6918299. Ott, Myle, et al. “Finding Deceptive Opinion Spam by
Any Stretch of the Imagination.” Proceedings of
Bengio, Yoshua, et al. “Representation Learning: A the 49th Annual Meeting of the Association for
Review and New Perspectives.” IEEE Computational Linguistics: Human Language
Transactions on Pattern Analysis and Machine Technologies-Volume 1, Association for
Intelligence, vol. 35, no. 8, IEEE, 2013, pp. Computational Linguistics, 2011, pp. 309–19.
1798–828.
Peng, Qingxi, and Ming Zhong. “Detecting Spam
Chowdhary, Neha S., and Anala A. Pandit. “Fake Review through Sentiment AnalysisPeng, Q. and
Review Detection Using Classification.” Zhong, M. (2014) ‘Detecting Spam Review
International Journal of Computer Applications, through Sentiment Analysis’, Journal of
vol. 180, no. 50, 2018, pp. 16–21. Software. Doi: 10.4304/Jsw.9.8.2065-2072.”
Journal of Software, 2014,
Fontanarava, Julien, et al. “Feature Analysis for Fake doi:10.4304/jsw.9.8.2065-2072.
Review Detection through Supervised
Classification.” 2017 IEEE International Rayana, Shebuti. “Collective Opinion Spam
Conference on Data Science and Advanced Detection : Bridging Review Networks and
Analytics (DSAA), IEEE, 2017, pp. 658–66. Metadata.” Proceedings of the 21th ACM
SIGKDD, 2015.
Hu, Minqing, and Bing Liu. “Mining and Summarizing
Customer Reviews.” Proceedings of the Tenth Rout, Jitendra Kumar, et al. “Deceptive Review
ACM SIGKDD International Conference on Detection Using Labeled and Unlabeled Data.”
Knowledge Discovery and Data Mining, vol. 50, Multimedia Tools and Applications, vol. 76, no.
no. 08, 2004, pp. 50-4466-50–4466, 3, Multimedia Tools and Applications, 2017, pp.
doi:10.5860/choice.50-4466. 3187–211, doi:10.1007/s11042-016-3819-y.
Jia, Shaohua, et al. “Fake Reviews Detection Based on
LDA.” 2018 4th International Conference on
756
Saumya, Sunil, and Jyoti Prakash Singh. “Detection of
Spam Reviews: A Sentiment Analysis
Approach.” CSI Transactions on ICT, vol. 6, no.
2, Springer, 2018, pp. 137–48.
Viviani, Marco, and Gabriella Pasi. “Quantifier Guided
Aggregation for the Veracity Assessment of
Online Reviews.” International Journal of
Intelligent Systems, vol. 32, no. 5, Wiley Online
Library, 2017, pp. 481–501.
Yilmaz, Cennet Merve, and Ahmet Onur Durahim.
“SPR2EP: A Semi-Supervised Spam Review
Detection Framework.” Proceedings of the 2018
IEEE/ACM International Conference on
Advances in Social Networks Analysis and
Mining, ASONAM 2018, 2018, pp. 306–13,
doi:10.1109/ASONAM.2018.8508314.
757
Turning Silver into Gold: Error-Focused Corpus Reannotation with
Active Learning
758
Proceedings of Recent Advances in Natural Language Processing, pages 758–767,
Varna, Bulgaria, Sep 2–4 2019.
2.1 Corpus Reannotation deep understanding of the original protocol, often
created by another team. It may also require to
The errors present in annotated corpora can have
modify the protocol or classification values, and
many sources. They can be simple typographical
also require an additional effort of the annotators
errors created during transcription or from the be-
to assess if they are not creating more errors when
gining on uncurated information. On a higher cog-
modifying an original answer. For all these rea-
nitive level, another frequent cause of error is an-
sons, it is important to explore ways by which
notators’ discrepancies over expert knowledge, in
some of the effort might be lowered, such as the
which disagreement over labels are not solved in
application of active learning on noisy corpora.
a consistent way. Protocol inconsistencies are yet
another source of noise in annotated corpora, of-
2.2 Active Learning
ten encountered when unforeseen cases fail to be
brought to light, managed, or added to the proto- Active learning (Cohn et al., 1996) has been used
col. Moreover, there is also all the possible causes to lower the annotation effort needed to train a
from the annotator side, like misunderstanding the prediction model in natural language processing
protocol, errors caused by fatigue, solving ambi- tasks. It traditionally involves a dialogue between
guities with the most favorable case instead of re- one (or more) human annotator(s) and a machine
porting it, and many others. All these issues can learning algorithm, the former being proficient at
diminish the quality of the annotations. annotating instances that the latter is providing
The nature of an annotation task can also influ- based on relevance measures. The process can be
ence the risk of generating errors. Intuitively, fine- split into three distinct steps.
grained classifications are often considered more The first one, called seeding, is where the al-
error-prone than those with a small number of gorithm must choose instances without being able
class values. This can be attributed to fuzzy fron- to rely on any annotation from the expert. The
tiers between close-by values, overlapping values chosen instances are submitted to be annotated
(i.e. choosing between ”entertainer” and ”come- by the expert. This starts the second and longest
dian”) or failing to recall a specific classification step, the querying phase. The active learning en-
option in a very large set of values. gine iterates between training a prediction model
Errors might also occur based on the interpreta- with the gold values (training set) from the ex-
tion of different annotators on the same case. This pert, choosing new instances relevant to training
can be the case in named entity classification tasks, a better model, submitting them to the expert and
when an occurrence can be considered both an or- adding the expert’s answers back into the training
ganization and a location (i.e. “I went to register at set. The third step, stopping, is applied after each
the office of University X”). These are edge cases training of the querying phase. It checks if there
that are often overlooked in annotation protocols is enough information in the model to annotate the
and might result in inconsistencies in the final an- rest of the corpus automatically so that the anno-
notated corpus. tation can stop. A bad stopping criterion might
A silver standard corpus (Rebholz-Schuhmann overfit the model, lowering its predictive power.
et al., 2010) is usually defined as a noisy set of One challenge of active learning is to balance
ground truth annotations provided automatically between specializing the classification on known
by state-of-the-art algorithms, while gold labels cases (for example, annotation errors), thus im-
are the higher quality annotations created by ex- proving the performance for current classes, or ex-
pert annotators. The silver labels are normally pro- ploring the problem space to find unknown but rel-
duced manually by human agents, but can also be evant instances that could improve the overall per-
obtained automatically by tools or trained predic- formance of the model.
tion models. Of course, the gold labels might still The goal of the current experiment is not to pro-
contain some degree of noise, but they are consid- duce the best prediction model, but to explore if
ered of better overall quality than the silver ver- the reannotation effort on a corpus can be reduced
sion. by using the active learning process with different
Corpus reannotation can be a tedious undertak- algorithms. These algorithms should not target the
ing, as it not only requires the same expert domain most informative instances for a prediction model,
knowledge to correct the annotations, but also a but instead choose those that are more likely to be
759
errors. If successful, the model should then be able conducted by adding human resources to the task
to extract more errors as it progresses. to distribute the effort among multiple users. This
is the proposition made in (Hovy et al., 2014) by
3 Related Works applying crowd sourced reannotation. In the same
multiuser settings, (Lin et al., 2016) propose an ap-
While different flavors of reannotation have been proach to reannotate labels by integrating another
used for natural language processing tasks on text human oracle in order to improve the quality of the
corpora, the work of (Rehbein and Ruppenhofer, annotations.
2017) is the most similar to our contribution from
an error finding perspective with active learning. 4 Methods
They evaluate the query-by-committee (QBC) and
variational inference (VI) active learning methods In order to improve noise detection, we focus on
to perform error detection on part-of-speech tag- the seeding and querying steps of the active learn-
ging and named entity recognition (NER) tasks. ing process. As the goal is to facilitate annotation
These approaches were tested in four contexts: in- of incorrect annotation, and not to create a predic-
domain (same type of training and testing data) on tion model per se, we did not explore the stopping
English POS, out-domain (different training and phase. The following sections detail the baselines
testing) on English POS, new task and language as well as the new methods used for the experi-
on German NER, and real-world context with hu- ments in Section 6.
man annotators on POS task. After 1,000 iter-
ations, the VI approach generally gives a better 4.1 Seeding Methods
noise reduction than QBC. While the POS task is The seeding method’s main goal from a reannota-
similar to the one used in our experiment (although tion perspective is to provide the highest error ra-
on a different language), the NER task uses only tio for the budgeted seed size. This contrasts with
4 class values (location, organization, person and a usual active learning task which is to find the
miscellaneous) while documents in our classifica- most informational instances to annotate in order
tion task can be categorized into a high number of to improve the model’s performances.
types. They do not specify the experimental seed- To capitalize on the silver classification infor-
ing methods used to choose the first examples. mation, we used outlier detection methods sepa-
Another similar research is Skeppstedt (2013) rately on each unique silver value. Our hypothe-
which uses active learning with two sources of sis for this is that most of the annotations should
tools generated annotation in order to tag and clas- be of good quality, although noisy, meaning that
sify named entities in Swedish clinical text docu- clustering the instances for a single silver value
ments. The challenge of combining multiple pre- should produce one or many clusters of correctly
annotated sources and active learning is to provide classified instances. A large enough dataset should
the right quality of information. If the sources then provide a cluster for each valid manifestation
are too noisy, the task will be more difficult and (depending on the features used) of a silver value.
unreliable. On the other hand, providing high- As the clusters represent valid cases, the outliers
quality sources might lower the attention and inter- should represent either rare cases or, ideally, noisy
est of the annotators. The proposed method tries to labels which should be reannotated by the expert.
overcome these two points by showing the sources For each silver value of a corpus, a random in-
without specifying which one is most likely to be stance was chosen in the detected outliers, to test
correct. No performance evaluation was done for the hypothesis that they should mostly be incor-
the proposed approach. rectly annotated cases.
Other experiments make usage of preanno- Four other outlier detection methods were
tated information without applying active learning tested as baselines to assess their performances
methods, like (Chou et al., 2006) who use a se- and compare them to the above method. These are
mantic role labeling tool trained on PropBank to not normally used in the seeding phase and they
pretag a biomedical corpus called BioProp. Af- do not consider the presence of a silver value in
ter the automatic annotation step, a human annota- the feature set. The first is the one-class SVM
tor manually checks the silver values and corrects (Schölkopf et al., 2001) which trains a support
them as needed. Other reannotation efforts may be vector machine with a radial basis kernel function
760
and returns the lowest supported instances in a dis- a negative one. The algorithm then selects those
tribution. that have the highest value, following the hypoth-
The local outlier factor (Breunig et al., 2000) esis that they would be similar to known errors.
computes the local deviation of density for an in- Other methods often applied in a standard ac-
stance with its closest neighbors using a k-nearest tive learning process have also been used as a basis
approach. If the local density of the close-by in- of comparison with the proposed method, namely
stances is significantly higher, the instance is con- distance to centroid, cosine similarity, hierarchi-
sidered an outlier and is selected for human anno- cal clustering and margin query.
tation.
Distance to centroid calculates a density center
The isolation forest algorithm (Liu et al., 2008)
point (centroid) from each instance of a specific
detects outliers by selecting the most isolated in-
gold value asked from the expert annotator. Each
stance of a randomly selected feature set and split
instance is then checked against each centroid and
point in the value range. These splits are then pro-
the ones which are furthest from all points are se-
jected into a tree structure and the average path
lected.
length to an instance gives its degree of isolation,
choosing the highest degree of isolation as the best The cosine similarity works in a similar way as
potential instance to annotate. the previous method but uses the cosine between
Finally, the covariance detector (Rousseeuw the instance and the centroids to assess the prox-
and Driessen, 1999) uses a Gaussian distribution imity.
around the density cluster to assess the degree to Hierarchical clustering also uses the distance
which an instance might be part of that cluster. to clusters as a degree of uncertainty to choose
ambiguous instances, but goes one step further
4.2 Querying Methods
by splitting these instances to choose only those
The proposed double centroid approach is based which are the most different from one another.
on the hypothesis that an annotator, either human This usually helps to provide a better sample to
or tool, tend to produce similar types of errors annotate and avoid ambiguous but very similar in-
through the annotation process. The source of stances.
these types could be the inability to differentiate
Margin query chooses the instances with the
between two classes, unknown terminology, etc.
smallest difference between the most probable
The method first use density-based clustering to predicted class and the second most probable.
group together newly annotated instances from the
seeding or previous querying steps. It is applied
once on erroneous instances, where the silver and 5 Datasets
new gold values did not match, and once on non-
erroneous instances, where the silver and new gold
For this experiment, we used a total of four
values matched. This gives multiple clusters con-
datasets, two targeting a document classification
taining either noisy (Cn ) or matching (Cm ) anno-
task and two for a text sequence classification task
tations.
on part-of-speech tags.
C As there were no publicly available manually
X n X
Cm corrected datasets, and correcting an existing one
|Ci | |Cj |
rank(l) = − (1)
i=1 dist(l, Ci )2
j=1
2
dist(l, Cj ) would have been too time consuming for the scope
of this project, we simulated the noise level by ap-
As shown in equation 1, for each silver in- plying a classification tool to each manually anno-
stance l that was not yet picked for relabeling, tated corpus. They are each presented in the fol-
a weighted squared distance dist(l, C)2 is calcu- lowing sections and Table 1 shows the size and
lated separately against each cluster centroid. The error rate for each of them. The error rate is calcu-
weight used is the cluster cardinality |C|, so that lated by dividing the number of errors in the silver
nearer and larger clusters have more influence on standard (compared to the gold standard) by the
an instance. These weighted distances are then total number of elements. Sections 5.5 and 5.6 de-
summed up with opposing values, in this case Cn scribe how the noise was generated for each cor-
clusters having a positive influence and Cm having pus to calculate the error rate.
761
Corpus Size Error rate sal dependency-based part-of-speech tags from the
Reuters 9,149 docs 0.0941 open class (ADJ, ADV, INTJ, etc.), the closed class
WoS 46,985 docs 0.3916 (ADP, AUX, CCONJ, DET, etc.) and the other
GSD-French 402,120 words 0.1015 class (PUNCT, SYM, X). The feature set for each
Sequoia 70,572 words 0.1080 instance included the token’s position in the sen-
tence, surface form, lemma, length, presence of
Table 1: Corpus size (documents or words) and space after the token and case type (capitalized,
error rate. all capital letters or mixed case).
5.4 Sequoia
5.1 WOS-46895 The Sequoia corpus (Candito and Seddah, 2012)
The WOS corpus (Kowsari, Kamran et al., 2018) contains 3,099 French sentences (70,572 tokens)
contains 46,985 documents in English taken from taken from different corpora such as Europarl,
the Web of Science website. Each document con- L’Est Republicain newspaper, French Wikipedia
tains the abstract from a published scientific paper. and European Medicine Agency. It has initially
These documents were picked from the fields of been annotated with constituency trees and then
psychology, medical sciences, biochemistry, com- converted to surface syntactic dependency trees.
puter science and three specialization of engineer- In our experiment, we only classified the part-of-
ing. While each document is only classified in a speech (POS) information. The feature set was the
single topic, terminology coverage of some topics same as the French-GSD corpus.
overlaps with others, such as between biochem-
istry and medical sciences. 5.5 POS Processing and Vectorisation
Each topic is further broken down into 134 spe- The GSD and Sequoia corpora were initially
cialized areas, varying from 9 to 53 areas for each tagged manually with universal part-of-speech
topic. While this dataset was primarily created tags1 which we used as the gold standard. While
for hierarchical classification, we only used the dependency information was widely available in
134 areas (the second layer of classification) for the original corpus, we removed them as they
the purpose of this research. would not be available in a raw text corpus with-
out applying a high quality dependency analyzer.
5.2 Reuters We kept morphological features as listed in Sec-
The Reuters-21578 corpus (Lewis, 2004) is com- tion 5.3.
posed of 10,788 English documents consolidated In order to create a silver version of the dataset,
for the text classification challenge of docu- each token was automatically reannotated with
ment understanding conference (DUC). There are TreeTagger (Schmid, 1997) to provide a new set
90 categories in the corpus. However, that dataset of tags. These tags were then converted to the uni-
was originally used for multi-class classification. versal part-of-speech tagset to make them compa-
To use it for our research, we extracted only rable with the original corpus tags.
the articles having a single class and kept only These corpora were then vectorized to be pro-
the classes that appeared more than once in the cessable by the classification algorithms, project-
dataset. The final dataset is made of 9,149 in- ing the information of each instance in a feature
stances across 56 categories like acq and earn. space. The feature set was also enriched for each
The labels’ distribution is heavily skewed as two instance with the information from the last five to-
of these classes make up 75% of the dataset in- kens and the next five tokens. The sentence bound-
stances. aries were respected, so that the first token in a
sentence would only get the next five tokens, the
5.3 French-GSD second token would get the previous token and the
We used the French portion of the GSD corpus next five, and so on. As the algorithms used cannot
(McDonald et al., 2013), version 2.2 at the time of deal with nominal data, each feature was one-hot
writing. We merged the three parts (dev, train, test) encoded. Once encoded, each dataset contained
into a single corpus consisting of 402,426 words 1,156 features including the silver annotation.
(16,448 sentences). While it is fully tagged with 1
https://universaldependencies.org/u/
dependence trees, we only retained the 17 univer- pos/
762
5.6 Classification Preprocessing and for active learning, choose a fixed number of in-
Vectorisation stances from each target class value (knowledge
The text documents in the WOS and Reuters cor- that would not be accessible in a real world set-
pora were stemmed using Porter stemmer (Porter, ting) or simply fail to mention the method.
1997) and had their stop words removed. In or- The methods described in the previous section
der to simulate silver level annotations, the cor- were run 100 times to smooth out the randomness
pora were vectorized and annotated using a five- effect of some outlier selection methods. It also
parts iteration. In other words, the corpus was split helped to show if most of the outliers were in fact
into five batches containing 20% of the corpus, a valid errors to be detected. Each chosen outlier
model was trained on the gold values of 80% of was then compared to the gold value to verify if
the corpus, recreating a new model to annotate the it was an error or a valid annotation. The num-
remaining 20%, and a different batch was swapped ber of unique silver values were not scaled down
to be annotated each time. to a specific seed size so as to provide an overall
The two corpora were then vectorized with two measure of performance.
word embeddings methods, fastText (Bojanowski As some clustering methods do not scale well
et al., 2016), Word2vec (Mikolov et al., 2013), with large datasets, we applied a feature reduc-
and the tf-idf (Salton and McGill, 1986) statisti- tion algorithm on each dataset. We used a random
cal method. The feature set size of the output forest (Breiman, 2001) estimator using 100 trees,
vector from each method, either produced with building each one with a random subset of fea-
neural networks or statistical measure, was set to tures from the dataset, evaluating the relevance
1,000 features. of each feature and discarding the less meaning-
ful ones. The Sequoia and French-GSD corpora
Corpus Method F1 were respectively reduced to 261 and 223 features,
Reuters fastText 0.87 while 250 and 334 features were retained from the
Word2vec 0.88 Reuters and WOS-46895 corpora. The silver an-
tf-idf3 0.91 notation was used as the relevance indicator for
WOS-46895 fasttext 0.59 this process instead of the gold which would not
Word2vec 0.57 be available in a real setting.
tf-idf3 0.61 The results presented in Table 3 show an im-
provement across all methods compared to base-
Table 2: Performances of vectorization methods. line performances. The error rate column is aver-
aged from all the results in the experimental runs.
To see which one provided the best expressiv- The gain is the ratio between the average error rate
ity with its features, the vectors for each corpus and the baseline performance from Table 1 (error
were evaluated with ten-fold validation using lo- rate column) using a random selection or a stan-
gistic regression as the passive learning algorithm. dard seeding method.
The results in Table 2 show the performances for While all seeding methods provide an improve-
each vectorization method on each corpus. While ment compared to the baseline, the local outlier
the results are quite close to one another, the statis- factor method seems the most promising when as-
tical tf-idf method using unigrams to trigrams (Tf- sessed from the average gain global score over all
idf3 ) provides higher results on all datasets. For corpora. Performances on the WOS corpus were
this reason, we use the feature set provided by this not favored by any of the four methods and were
method for the experiments. just marginally better with the local outlier factor.
6 Experiment 6.2 Querying Experiment
6.1 Seeding Experiment The four baseline methods detailed in Section 4
We use the error rate of the corpus (as previously were applied to each vectorized corpus for the doc-
reported in Table 1) as a random baseline for seed- ument classification and POS tagging tasks. To
ing, which is the equivalent of making a random avoid boosting the performance of the error seek-
selection. As studied in (Hu et al., 2010), most of ing query method, we did not use the previous
the papers either use this type of seeding method approaches for seeding. Instead we randomly se-
763
Method Measure Reuters WoS Sequoia French-GSD Average gain
SVM EDP 0.3576 0.4346 0.1429 0.2938
Gain 3.80 1.11 1.32 2.89 2.28
Covariance EDP 0.3606 0.4331 0.1786 0.2971
Gain 3.83 1.11 1.65 2.93 2.38
Isolation EDP 0.3424 0.4336 0.1929 0.2267
forest Gain 3.64 1.11 1.79 2.23 2.19
Local EDP 0.4030 0.4369 0.2000 0.3729
outlier Gain 4.28 1.12 1.85 3.67 2.73
764
Method Measure Reuters WoS Sequoia French-GSD Average gain
Distance to EDP 0.1194 0.3954 0.1049 0.1196
centroid Gain 1.27 1.01 1.03 1.11 1.10
Cosine EDP 0.0974 0.3796 0.1198 0.1003
similarity Gain 1.04 0.97 1.18 0.93 1.03
Hierarchical EDP 0.1110 0.3943 0.1196 0.1079
clustering Gain 1.18 1.01 1.18 1.00 1.09
Margin EDP 0.1032 0.3937 0.1074 0.1057
query Gain 1.10 1.01 1.06 0.98 1.03
Double EDP 0.1982 0.4283 0.2499 0.1565
centroid Gain 2.11 1.09 2.46 1.45 1.78
Table 4: Query phase error detection precision and gain after 200 instances.
765
References F. T. Liu, K. M. Ting, and Z. Zhou. 2008. Iso-
lation forest. In 2008 Eighth IEEE Interna-
Piotr Bojanowski, Edouard Grave, Armand Joulin, tional Conference on Data Mining. pages 413–422.
and Tomas Mikolov. 2016. Enriching word vec- https://doi.org/10.1109/ICDM.2008.17.
tors with subword information. arXiv preprint
arXiv:1607.04606 . Ryan McDonald, Joakim Nivre, Yvonne Quirmbach-
Brundage, Yoav Goldberg, Dipanjan Das, Kuz-
Leo Breiman. 2001. Random forests. In Machine man Ganchev, Keith Hall, Slav Petrov, Hao
Learning. pages 5–32. Zhang, Oscar Täckström, Claudia Bedini, Núria
Bertomeu Castelló, and Jungmee Lee. 2013. Uni-
Markus M Breunig, Hans-Peter Kriegel, Raymond T versal dependency annotation for multilingual pars-
Ng, and Jörg Sander. 2000. LOF: Identifying ing. In Proceedings of the 51st Annual Meet-
Density-Based Local Outliers. In ACM SIGMOD ing of the Association for Computational Lin-
2000 Int. Conf. On Management of Data. page 12. guistics (Volume 2: Short Papers). Associa-
tion for Computational Linguistics, pages 92–97.
Marie Candito and Djamé Seddah. 2012. Le cor- http://aclweb.org/anthology/P13-2017.
pus sequoia : annotation syntaxique et exploita-
tion pour l’adaptation d’analyseur par pont lex- Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S
ical (the sequoia corpus : Syntactic annotation Corrado, and Jeff Dean. 2013. Distributed
and use for a parser lexical domain adaptation representations of words and phrases and their
method) [in french]. In Proceedings of the compositionality. In C. J. C. Burges, L. Bottou,
Joint Conference JEP-TALN-RECITAL 2012, vol- M. Welling, Z. Ghahramani, and K. Q. Wein-
ume 2: TALN. ATALA/AFCP, pages 321–334. berger, editors, Advances in Neural Information
http://aclweb.org/anthology/F12-2024. Processing Systems 26, Curran Associates, Inc.,
pages 3111–3119. http://papers.nips.cc/paper/5021-
Wen-Chi Chou, Richard Tzong-Han Tsai, Ying-Shan distributed-representations-of-words-and-phrases-
Su, Wei Ku, Ting-Yi Sung, and Wen-Lian Hsu. and-their-compositionality.pdf.
2006. A semi-automatic method for annotating a
biomedical proposition bank. In Proceedings of M. F. Porter. 1997. Readings in information re-
the Workshop on Frontiers in Linguistically An- trieval. Morgan Kaufmann Publishers Inc.,
notated Corpora 2006. Association for Computa- San Francisco, CA, USA, chapter An Al-
tional Linguistics, Sydney, Australia, pages 5–12. gorithm for Suffix Stripping, pages 313–316.
https://www.aclweb.org/anthology/W06-0602. http://dl.acm.org/citation.cfm?id=275537.275705.
David A. Cohn, Zoubin Ghahramani, and Michael I. Dietrich Rebholz-Schuhmann, Antonio José Jimeno-
Jordan. 1996. Active learning with statistical mod- Yepes, Erik M. van Mulligen, Ning Kang, Jan
els. J. Artif. Int. Res. 4(1):129–145. Kors, David Milward, Peter Corbett, Ekaterina
Buyko, Katrin Tomanek, Elena Beisswanger, and
Dirk Hovy, Barbara Plank, and Anders Søgaard. 2014. Udo Hahn. 2010. The CALBC silver stan-
Experiments with crowdsourced re-annotation of a dard corpus for biomedical named entities —
POS tagging data set. In Proceedings of the 52nd a study in harmonizing the contributions from
Annual Meeting of the Association for Computa- four independent named entity taggers. In
tional Linguistics, ACL 2014, June 22-27, 2014, Proceedings of the Seventh conference on In-
Baltimore, MD, USA, Volume 2: Short Papers. pages ternational Language Resources and Evaluation
377–382. http://aclweb.org/anthology/P/P14/P14- (LREC’10). European Languages Resources Asso-
2062.pdf. ciation (ELRA), Valletta, Malta. http://www.lrec-
conf.org/proceedings/lrec2010/pdf/888 Paper.pdf.
Rong Hu, Brian Mac Namee, and Sarah Jane Delany.
2010. Off to a good start: Using clustering to select Ines Rehbein and Josef Ruppenhofer. 2017. Detect-
the initial training set in active learning. In FLAIRS ing annotation noise in automatically labelled data.
Conference. In Proceedings of the 55th Annual Meeting of the
Association for Computational Linguistics (Volume
Kowsari, Kamran, Brown, Donald, Heidarysafa, Mo- 1: Long Papers). Association for Computational
jtaba, Jafari Meimandi, Kiana, Gerber, Matthew, Linguistics, Vancouver, Canada, pages 1160–1170.
and Barnes, Laura. 2018. Web of Science Dataset. https://doi.org/10.18653/v1/P17-1107.
https://doi.org/10.17632/9rw3vkcfy4.6. Peter J. Rousseeuw and Katrien Van Driessen. 1999.
A fast algorithm for the minimum covariance de-
David Lewis. 2004. Reuters 21578 data set version terminant estimator. Technometrics 41(3):212–223.
1.0 http://www.daviddlewis.com/resources/ testcol- https://doi.org/10.1080/00401706.1999.10485670.
lections/reuters21578/readme.txt.
Gerard Salton and Michael J McGill. 1986. Introduc-
Christopher H Lin, M Mausam, and Daniel S. Weld. tion to modern information retrieval .
2016. Re-Active Learning: Active Learning with
Relabeling. In Thirtieth AAAI Conference on Artifi- Helmut Schmid. 1997. Probabilistic part-of-speech
cial Intelligence. page 8. tagging using decision trees. In Daniel Jones and
766
Harold Somers, editors, New Methods in Language
Processing, UCL Press, London, GB, Studies in
Computational Linguistics, pages 154–164.
Bernhard Schölkopf, John C. Platt, John C. Shawe-
Taylor, Alex J. Smola, and Robert C. Williamson.
2001. Estimating the support of a high-dimensional
distribution. Neural Comput. 13(7):1443–1471.
https://doi.org/10.1162/089976601750264965.
Maria Skeppstedt. 2013. Annotating named entities in
clinical text by combining pre-annotation and active
learning. In Proceedings of the ACL Student Re-
search Workshop. page 74–80.
767
NLP Community Perspectives on Replicability
Margot Mieskes Karën Fort
University of Applied Sciences, Darmstadt Sorbonne Université, EA STIH
Germany Paris, France
margot.mieskes@h-da.de karen.fort@sorbonne-universite.fr
768
Proceedings of Recent Advances in Natural Language Processing, pages 768–775,
Varna, Bulgaria, Sep 2–4 2019.
cation more visible and how the three factors de- which compared the situation in sharing research
scribed above influence this. Our results indicate artefacts between 2011 and 2016 indicates that,
that the participants in general regard replication while the situation has improved and the availabil-
as an important issue and that the NLP commu- ity of data is high, access to code is less so and re-
nity could do more to support replication, which questing code is unsuccessful in most of the cases.
would strengthen the field as a whole.6 Based on results of their actual replication exper-
iments, at most 60% of the studies are replicable,
2 Related Work but only if the need for exact replication was re-
laxed.
One of the earliest reports of a replication ef-
Fares et al. (2017) present a repository and in-
fort addresses manual word-sense disambiguation
frastructure containing texts, tools and embed-
based on four words, representing different de-
dings for English and Norwegian. Their aim
grees of difficulty (Kilgarriff, 1999). The author
is to facilitate replicability and testing of previ-
reports that humans agree in this task on average
ous results. Dror et al. (2017) propose a “repli-
in 95% of the cases. Following Fokkens et al.
cability analysis framework” and demonstrate it-
(2013), others look into parameters that influence
s use on various tasks such as part-of-speech
the replicability of results. Dakota and Kübler
tagging or cross-domain sentiment classification.
(2017); Marrese-Taylor and Matsuo (2017) and
They specifically target cases where algorithms
Horsmann and Zesch (2017) report various param-
are compared across multiple data sets. The re-
eters and problems with replication experiments
sults indicate that testing on a range of data sets is
for morphology and syntax.
only beneficial if the data sets are heterogeneous.
In the field of biomedical NLP, Olorisade et al.
(2017) assess the reproducibility of findings pub- 3 Survey Design
lished in 33 papers. They notice that data sets were
missing, making it impossible to reproduce results Our survey has 18 questions, of which many were
for 80% of the papers. These figures are in line conditional and show only if they apply to the re-
with results reported by Mieskes (2017). They spondent. Thus, not all questions have been an-
consider that a permanent link to the resources swered by all participants, while most multiple-
(data set, software, etc.) must exist along with choice questions allow for several answers, result-
published papers. As part of a NLP challenge, ing in more answers than participants for these
Névéol et al. (2016) report results on replicating questions. Questions are grouped into three cat-
experiments from three systems submitted to the egories: (i) replication work in general, (ii) repli-
CLEF eHealth track. They show that replication is cating one’s own work and (iii) replicating others’
feasible although “ease of replicating results var- work.
ied”. They suggest the allowance of extra pages General questions quiz participants on their per-
for papers, where information required to replicate ception of replication work. We also inquire about
an experiment could be reported. their current position to investigate potential cor-
Moore and Rayson (2018) illustrate how to pub- relation with other aspects of the survey. Ques-
lish relevant details to reduce efforts in repeatabil- tions addressing participants’ replication experi-
ity and generalisability. Suggestions include using ence specifically enquired about research artefact-
only open data, open source code and providing s availability (data, code, parameters, etc.) and
extensive documentation in the code. about the timeline of the replication experience in
Wieling et al. (2018) describe one example order to assess attrition.
where exact replication was possible and the au- The survey was advertised on professional mail-
thors list the parameters that allowed them to do ing lists (BioNLP, Corpora, LN and GLCL7 ) and
so: a virtual image, containing all code and all da- social network (LinkedIn). The appendix gives
ta or providing CodaLab worksheets. Their study, details on the progression of responses. With re-
6
spect to sensitive data, only e-mail addresses were
The complete results of the survey are avail-
able at https://github.com/replicateNLP/
provided, on a voluntary basis, and we follow the
Survey-RANLP2019. Please note, that due to privacy ACM Code of Ethics and Professional Conduct8 ,
regulations, we had to remove some free text answers that
7
contain personal information, such as E-Mail addresses, Biomedical NLP, French and German NLP.
8
which were given on a voluntary basis. https://www.acm.org/code-of-ethics
769
specifically sections 1.6 and 1.7. With respect to how often replication of the
Figure 1 illustrates the flow of the questionnaire. same experiment was tried, 15 participants gave
If a person did neither replicate their own or some- more than one answer, giving us 172 answers (see
one elses work, they only had to answer the blue Figure 5 for the detailed figures). Of these, 28.5%
marked questions. If a participant tried to replicate indicate that replication was tried only once, while
his/her own work, but not someone elses work, 38.4% tried 3 times or more.
they only had to answer the blue and the orange When looking at the last attempt (total answers
questions. Only persons who have experience in 234), nearly half (47.0%) report that they reached
replicating their own and someone elses work had the same general conclusions, while 23.1% s-
to go through the whole questionnaire, including tate that they reached the same figures. A lit-
the green questions.9 This flow in addition to the tle less than 10% report that they managed to re-
possibility to give more than one answer in some implement the system, but got significantly differ-
questions results in different numbers of answers ent results. Another 14.9% could not find either
for each question. the code or the data or the parameters used for the
experiment (see Figure fig:resultsOwn for the ab-
4 Results solute numbers). 52 participants gave more than
We received 225 responses and the two biggest one answer of which 25 report that they reached
groups of participants in our study identify them- the same general conclusions and the same fig-
selves as graduate students and postdocs. ures. Overall, the results indicate that even in the
With respect to when work on replication has case when researchers try to replicate their own
been done, 36 participants (16%) gave more than work, they fully succeed in only 23.1% of the cas-
one anwer, indicating that they did work on repli- es.
cation at various stages of their career. Most an-
4.3 Replicating Others’ Work
swers (50.3%) state that replication was done on
MSc or PhD level, less on PostDoc level (20.7%) About 60% (total 130) of the participants report
and slightly more as Faculty members (24.3%). that they tried to replicate someone else’s work
However, we did not find strong correlations be- (see Figure 7 for detailed, absolute numbers).
tween the respondents’ position and opinion on 51 respondents gave more than one answer with
the importance (or lack thereof) of reproducibili- respect to the results achieved when replicating
ty. Figure 2 shows the absolute numbers for this somebody else’s work, resulting in 211 answers.
question, while Figure 3 shows the numbers for Approximately 40% of the answers state that they
the participants current position. reached the same general conclusions or figures,
while another 33.6% of the answers state that they
4.1 General Stance towards Replicability managed to re-implement or re-run the system, but
The answers show that in 56.4% of the cases, work with significantly different results. Nearly 23.1%
on replication is considered “Important” and an- of the responses state that re-implementation or
other 7.5% state that it is “Somewhat Importan- re-runnning of experiments was not achieved (see
t”. Only 2 answers indicate that this work is “U- Figure 8 for the absolute responses for this ques-
nimportant”. 20% of the answers regard work on tion). This means that over half of the replication
replication as publishable, while 11.8% deem it experiments failed, either early on or at the level
unpublishable. 87 participants gave more than one of results achieved.
answer. The majority (49) consider work on repli-
4.4 Accessibililty of Research Artefacts
cation as important and publishable, while 26 of
them consider it important but not publishable (see For finding research artefacts such as code, data
Figure 4 for the absolute numbers). and parameters, respondents gave several answer-
s, resulting in 250 answers for where the code can
4.2 Replicating one’s Own Work be found, 260 answers for finding data and re-
Roughly 70% (156) of the participants declare sources and 233 answers for finding the experi-
they have tried to replicate their own work while mental parameters. GitHub is by far the most pop-
about 30% have not tried (total 225). ular (36.4% of the answers) for accessing code,
9
Please note that only the most important questions are but more than 23.6% of the answers state that
illustrated here and some questions have been left out. code is found on the authors’ personal webpage,
770
Figure 1: Illustration of the questionnaire flow.
Figure 3: Position of the participants at the time of Figure 5: Participants who tried to replicate their
the survey. own work.
771
which does not guarantee availability beyond that
person maintaining his/her webpage. More than
14% of the answers report that the code could not
be found. Data is also primarily published vi-
a GitHub or personal webpages (25% and 25.7%
of the answers respectively). 11.1% report that
the material used for the experiments could not be
found. Parameters for experiments are primarily
found in the respective publications (40.3% of the
answers), while 21.9% of the answers state they
could be found on GitHub as well. 13.7% report
that they could not find parameters at all. Fig-
ure 9 illustrates the absolute numbers concerning
Figure 6: Results achieved when trying to repli- the availability of code, while Figure 10 and Fig-
cate own work. ure 11 illustrate them for other resources and pa-
rameters respectively.
772
Figure 11: Sources for Accessing Parameters for Figure 13: Quality of Answers by original authors.
Replication.
773
cited enough times should be replicated”. build on top of it and with what results. NAA-
CL recently initiated a “test of time” award. This
6 Conclusions and Future Work could be extended to consider experimental work
Based on a survey we gave insight into the NLP that has been cited often and gained influence on a
community’s view on replicability. We targeted shorter time-scale. This work could be verified for
three different facets of this topic: Authors pub- a follow-up conference. Replicability could also
lishing their work, Researchers building on top of become a factor in the best paper awards.
other researchers’ work and the Community, sup-
porting such efforts. Our results show that on the
References
authors’ side more information has to be shared
openly, rather than via personal communication. David Blanco, Jamie J Kirkham, Douglas G Altman,
David Moher, Isabelle Boutron, and Erik Cobo.
The use of reporting guidelines formalized into a
2017. Interventions to improve adherence to report-
protocol has been suggested recently for clinical ing guidelines in health research: a scoping review
NLP (Velupillai et al., 2018). Earlier studies from protocol. BMJ Open, 7(11).
the clinical domain suggest that adherence to such
António Branco, Nicoletta Calzolari, and Khalid
guidelines is suboptimal (Samaan et al., 2013) and Choukri, editors. 2016. Proceedings of the Work-
methods to improve adherence are being investi- shop on Research Results Reproducibility and Re-
gated (Blanco et al., 2017). The task of creating sources Citation in Science and Technology of Lan-
guidelines falls to the community and the adher- guage.
ence of such guidelines could become part of the António Branco, Nicoletta Calzolary, and Khalid
reviewing process. Choukri, editors. 2018. Proceedings of the 4REAL
Experiments in replication fail more often than 2018 - Workshop on Replicability and Reproducibil-
ity of Research Results in Science and Technology of
not. If we document and store all relevant infor- Language. European Language Resources Associa-
mation so that results could be reproduced by our- tion, Paris.
selves (e.g., before the final paper submission), the
K. Bretonnel Cohen, Aurélie Névéol, Jingbo Xia, Ne-
package could be published completely. Results
gacy Hailu, Larry Hunter, and Pierre Zweigenbaum.
by Wieling et al. (2018) indicate that images con- 2017. Reproducibility in Biomedical Natural Lan-
taining all the material or technical lab books pub- guage Processing. In AMIA annual symposium pro-
lished on CodaLab might be a way to proceed. Ad- ceedings, page 1994. American Medical Informatics
ditionally, failure to replicate previous work (i.e. Association.
not achieving the same results and/or not being K. Bretonnel Cohen, Jingbo Xia, Pierre Zweigen-
able to draw the same conclusions as previous- baum, Tiffany Callahan, Orin Hargraves, Foster
ly reported), should be publishable in a way that Goss, Nancy Ide, Aurélie Névéol, Cyril Grouin, and
Lawrence E. Hunter. 2018. Three Dimensions of
gives us scientific merit and could be encouraged Reproducibility in Natural Language Processing. In
more. Proceedings of the Eleventh International Confer-
Based on the comments, each of us, in all of our ence on Language Resources and Evaluation (LREC
individual roles can improve the situation: As au- 2018) Miyazaki, Japan, May 7–12, 2018. European
Language Resources Association (ELRA).
thors, we can be more diligent when reporting our
experiments and experimental setup—even testing Daniel Dakota and Sandra Kübler. 2017. Towards
the replicability of our experiments ourselves. As Replicability in Parsing. In Proceedings of the In-
ternational Conference Recent Advances in Natural
reviewers, we can be more careful to check the
Language Processing, RANLP 2017 Warna, Bulgar-
supplementary material for relevant information, ia, 2–8 September 2017, pages 185–194. INCOMA
pointing out missing elements. As a community, Ltd.
we can appreciate replication more and develop
Rotem Dror, Gili Baumer, Marina Bogomolov, and Roi
guidelines both for authors and for reviewers. Reichart. 2017. Replicability Analysis for Natu-
ral Language Processing: Testing Significance with
Future Work The next steps include, but are not Multiple Datasets. Transactions of the Association
limited to, analyzing whether the supplementary for Computational Linguistics, 5:471–486.
material and appendices actually do improve repli-
Murhaf Fares, Andrey Kutuzov, Stephan Oepen, and
cability, as stated by Névéol et al. (2016). Further- Erik Velldal. 2017. Word vectors, reuse, and repli-
more, evaluating the repository offered by Fares cability: Towards a community repository of large-
et al. (2017), whether other researchers actually text resources. In Proceedings of the 21st Nordic
774
Conference on Computational Linguistics Gothen- Ted Pedersen. 2008. Empiricism Is Not a Matter of
burg, Sweden, 22–24 May, 2017, pages 271–276. Faith. Computational Linguistics, 34(3):465–470.
Association for Computational Linguistics.
Zainab Samaan, Lawrence Mbuagbaw, Daisy Kosa,
Antske Fokkens, Marieke van Erp, Marten Postma, Ted Victoria Borg Debono, Rejane Dillenburg, Shiyuan
Pedersen, Piek Vossen, and Nuno Freire. 2013. Off- Zhang, Vincent Fruci, Brittany Dennis, Monica Ba-
spring from Reproduction Problems: What Repli- wor, and Lehana Thabane. 2013. A systematic scop-
cation Failure Teaches Us. In Proceedings of the ing review of adherence to reporting guidelines in
51st Conference of the Association for Computa- health care literature. Journal of Multidisciplinary
tional Linguistics Sofia, Bulgaria 4–9 August 2013, Healthcare, 6:169–88.
pages 1691–1701.
Sumithra Velupillai, Hanna Suominen, Maria Liaka-
Tobias Horsmann and Torsten Zesch. 2017. Do LSTMs ta, Angus Roberts, Anoop D. Shah, Katherine Mor-
really work so well for PoS tagging? – A replication ley, David Osborn, Joseph Hayes, Robert Stewart,
study. In Proceedings of the 2017 Conference on Johnny Downs, Wendy Chapman, and Rina Dutta.
Empirical Methods in Natural Language Process- 2018. Using clinical Natural Language Processing
ing Copenhagen, Denmark, September 9–11, 2017, for health outcomes research: Overview and action-
pages 727–736. Association for Computational Lin- able suggestions for future advances. Journal of
guistics. Biomedical Informatics, 88:11 – 19.
Adam Kilgarriff. 1999. 95% Replicability for Manual Martijn Wieling, Josine Rawee, and Gertjan van No-
Word Sense Tagging. In Proceedings of the Ninth ord. 2018. Squib: Reproducibility in Computational
Conference of the European Chapter of the Asso- Linguistics: Are We Willing to Share? Computa-
ciation for Computational Linguistics (EACL-1999) tional Linguistics, 44(4):641–649.
Bergen, Norway, 8–12 June, 1999. Association for
Computational Linguistics.
775
Unsupervised Data Augmentation
for Less-Resourced Languages with no Standardized Spelling
Alice Millour and Karën Fort
Sorbonne Université / STIH - EA 4509
28, rue Serpente, 75006 Paris, France
alice.millour@etu.sorbonne-universite.fr,
karen.fort@sorbonne-universite.fr
776
Proceedings of Recent Advances in Natural Language Processing, pages 776–784,
Varna, Bulgaria, Sep 2–4 2019.
STEP 1 STEP 2 STEP 3 and 4
crowdsourcing rules extraction rules application & lookup
CORPUS
Southern variant
Yes
+ New VARIANT PAIR
OOV words widdersch
wittersch
Northern variant VOCABULARY
- wittersch
- etc.
etc. Modify corpus
guages presents additional difficulties such as ac- These steps are detailed in sections 2 and 3 and
cessibility to the speakers, or representativity of illustrated with their application on Alsatian.
contents (Millour and Fort, 2018b). In the context of OOV words reduction in a
Yet, when a community of speakers can be given corpus, step 4 is followed by a transposition
found on-line, it seems necessary to empower of those for which a variant has been identified in
them to produce raw corpora and to document RLookup . Especially, in the context of supervised
variability. In fact, the speakers appear to be, col- machine learning, one cannot expect to find all ex-
lectively, the only experts of the mechanisms at isting variants in a training corpus. By replacing
stake. an OOV word by one of its already known spelling
To meet this goal, we developed a crowd- variants, we make the most of the annotations we
sourcing platform that collects two types of re- have at our disposal (see Section 4).
sources: (1) raw texts and (2) spelling variants on
these texts. These resources are used to seed the 1.3 The Case of Alsatian
unsupervised augmentation of the multi-variant Alsatian is a French regional language count-
lexicon following a process that we detail in Fig- ing 550,000 speakers in 2004 (Barre and Vander-
ure 1. schelden, 2004). This continuum of Alemannic di-
alects is an example of language in which the di-
1.2 Process Overview
alectal variants are not erased in the written form
Given an existing linguistic resource (corpus, lex- by any spelling system.
icon, or both) RLookup and a set of out-of- Initiatives such as the Orthal guide-
vocabulary words V ocOOV , the process consists lines (Crévenat-Werner and Zeidler, 2008)
of four steps: have been developed to unify the Alsatian spelling
while being respectful of its variations. Yet, these
1. crowdsourcing spelling variant pairs, keep the variability (Kı̀risch and Kı̀ch are the
Orthal version of Kerisch and Kı̂ch, Northern
2. automatic rules extraction,
and Southern possible versions for the word
3. application of the rules on elements of “church”), and are still unknown by a majority
V ocOOV , of Alsatian Internet users as shown by a recent
survey (Millour, 2019).
4. lookup of the resulting transformed spelling For this reason, the dialectal variations (6 to
in RLookup . 8 variants emerge from the continuum) combine
777
with the variety of spelling habits, which might de- (with two to six variants per word), e.g. {bı̀tsi,
pend, for instance, on the linguistic backgrounds bessel, béssel}, “a bit of”.
of the speakers. The only information we possess about these
Also, since there exist an active community of participants is the languages they speak and their
on-line speakers, Alsatian is a good candidate for place of origin, when they fill it in in their pro-
crowdsourcing experiments. file. Based on the information provided by 8 of
them, we can assume that 3 to 4 dialectal areas are
2 Crowdsourcing Spelling Variants covered by the towns of origin of the participants.
We developed a slightly gamified crowdsourcing No assumption can be made regarding their profi-
platform, Recettes de Grammaire1 which ciency in Alsatian.
allows us to collect (i) raw corpora in the shape The size of this resource does not allow us to
of cooking recipes, (ii) part-of-speech annotations perform direct lookup for any out-of-vocabulary
on the recipes, and (iii) alternative spellings. The word me might encounter. However, we can use
platform is language independent and its source the aligned variants to identify substitution pat-
code is freely available on GitHub2 under the Ce- terns, and extract sets of rules we apply to any
CILL v2.1 license.3 OOV word as described in the following section.
We do not differentiate variants due to a varia-
tion in dialects, in spelling or in an accumulation 3 Unsupervised Data Augmentation
of these two factors during collection. 3.1 Rules Extraction
The addition of a new spelling variant can be
performed: (i) by adding a variant to any word that In the manner of (Prokić et al., 2009), we used AL-
is present on the platform by clicking on a word PHAMALIG4 , a multi sequence alignment tool, to
cloud on the main page (see Figure 2), (ii) by dy- perform the alignment of our variants necessary to
namically editing the written contents on the web- the extraction of substitution patterns. The tool re-
site thanks to a feature called “Personally, I would quires to be provided with an alphabet of symbols,
have said it like that!”, illustrated on Figure 3. weighted with the match, mismatch, insertion and
deletion scores of given characters. Since we have
These features enable the participants to mod-
no a priori knowledge of these scores, the only as-
ify the content they read and further annotate in
sumption we made is that vowels are more likely
a manner that suits their writing habits. In fact,
to match vowels than consonants and vice versa.
feedback we received on previous experiments led
Insertion and deletion are given the same scores
on crowdsourcing part-of-speech annotations for
for all characters. An example of the alignment
Alsatian (Millour and Fort, 2018a) highlighted the
obtained for four crowdsourced variants is given
fact that some participants felt unrepresented by
in table 1.
the texts on the platform, and that annotating di-
alectal or spelling variants they are not familiar ˆ G A L - R Ì E W L E K Ü E C H E $ (1)
ˆ G A L E R I E B L E K Ü E C H A $ (2)
with was an obstacle hard to overcome. ˆ G A L E R - E W L E K Ù - C H E $ (3)
The interface allows the participants to provide ˆ G A L - R Ì A W L A K Ü A C H A $ (4)
an alternative spelling for either a single word or a Table 1: Alignment of four variants of the Alsatian
sequence of words. Although the latter facilitates (compound) word for “carrot cake”.
the task for the participants, it sometimes leads to
alternative spellings which number of words did
From the produced alignments we can identify
not match the original version, hence could not
substitution patterns of different degrees of rigid-
be immediately aligned. In such cases and when
ity, depending on the size of the context. We
possible, the alternative spellings were manually
extract three sets of rules which either force the
aligned with the original version.
matching of the left (L), right (R) or both contexts
So far, the collected resource contains 367 vari-
(L+R).
ants provided by 10 participants for 145 words
The ˆ and $ characters, respectively represent-
1
“Grammar’s Recipes”, see https://bisame. ing the beginning and the end of a word, are in-
paris-sorbonne.fr/recettes.
2 4
See https://github.com/alicemillour/ Source code: http://alggen.lsi.
Bisame/tree/recipes. upc.es/recerca/align/alphamalig/
3
See http://www.cecill.info/index. intro-alphamalig.html.
778
Figure 2: Spelling addition using the wordcloud. The word is shown in its context, with the proposed
part-of-speech (if available).
779
3. lookup: the sequence of rules apply until In the following experiments, we chose to train
the produced form is matched with a word MElt, which enables us to take advantage of
present in Vlookup . available lexicons existing for Alsatian. The dif-
ferential in performance is more interesting to us
Although this “brute-force” method generates a than the performance per se, which is why we
great quantity of noise, the filtering operated by chose not to focus on testing our methodology on
Vlookup leads to the matching of OOV word with other taggers.
existing variant candidates. Two POS-tagged corpora are available for Alsa-
Since part of the dialectal and spelling varia- tian. Both are made of texts produced in an uncon-
tion mechanisms may be similar to some of the trolled environment (such as Wikipedia5 ) and
language morphological rules (such as gender, contain multiple variants of the language:
number, conjugation or declension), the generated
variant pairs should be manually checked in con- • The Crowdsourced Corpus (Millour
text. and Fort, 2018b), Crowd C, annotated by
This phenomenon is illustrated by the analysis benevolent participants on a dedicated
of the pairs generated for Alsatian in Section 5. crowdsourcing platform Bisame6 with the
universal POS tagset (Petrov et al., 2012)
4 Evaluation on a Downstream Task extended with two categories: APPART
(preposition-determiner contraction), and
To illustrate the benefits of the identification of FM (foreign words). The corpus contains
variant pairs, we evaluate its impact on a down- 9,282 tokens (439 sentences), and is avail-
stream task: part-of-speech (POS) tagging. able under CC BY-NC-SA license. The
Previous experiments on Alsatian from (Millour accuracy of the annotations provided by the
and Fort, 2018b) have shown that using multiple benevolent participants has been evaluated to
variants for training can lead to a drop of accuracy 93% (Millour and Fort, 2018b).
on the sections of the evaluation corpus which do
not match the variants represented in the training • The Annotated Corpus for the
corpus (-1.4% accuracy when a corpus of Stras- Alsatian Dialects (Bernhard et al.,
bourg specific variant is added in the training of a 2018a), Trad C, annotated with the tagset
Southern variant only). described above, extended with the cate-
In this context, we use our methodology to gories EPE (epenthesis) and MOD (modal
match OOV words from the evaluation corpus verb) (Bernhard et al., 2018b). The corpus
with their potential spelling variant appearing in contains 12,570 tokens (533 sentences) and
the training corpus. is available under CC BY-SA license. It was
It is important to understand that this process annotated manually by expert linguists.
is independent from the tagger, and occurs after it We manually corrected Trad C to match the
has been trained. The extraction of pairs is per- tagset used in Crowd C.
formed at the time of annotation on a previously The corpus resulting from the concatenation of
unseen corpus. the two corpora, Concat C, was used for the fol-
lowing experiments. We performed a cross vali-
4.1 Language Resources and Tools Used for
dation on 4 subdivisions (80% used for training,
Evaluation
Concat C80, 20% for the evaluation, Concat C20).
Experiments in POS tagging Alsatian include our We also have at our disposal two lexica:
previous work (Millour and Fort, 2018b), which
uses MElt (Denis and Sagot, 2012), a freely avail- • a multi-variant lexicon MultiV ar L of 54,355
able sequence labeller achieving at best 84% ac- entries annotated with their POS, containing
curacy when the variants in the training and the grammatical words (Bernhard and Ligozat,
evaluation corpus are carefully controlled. Experi- 2013), verbs from (Steiblé and Bernhard,
ments using word embeddings have been also been 2016), and various entries from (i) the Office
carried on Alsatian by (Magistry et al., 2018), us- for Alsatian Language and Culture (OLCA)
ing a raw corpus of 200 000 tokens and reaching 5
See https://als.wikipedia.org
91% accuracy. 6
See https://bisame.paris-sorbonne.fr.
780
bilingual lexicons, (ii) the dictionary com- Before transp. After transp.
piled by the Culture and Heritage of Alsace Overall 0.859 0.864
Association (ACPA), and (iii) a multilingual OOV words 24% 22%
French-German-Alsatian dictionary (Adolf,
Table 2: Accuracy of the model trained on multi-
2006).
variant corpora, before and after the corpus trans-
• the Lexicon of Place Names in position.
the Alsatian Dialects which con-
tains 1,346 entries (Bernhard, 2018), used
during training only. The proportion of OOV words was diminished
by around 2% resulting in an improvement of the
4.2 Application of the Methodology tagging performance of 0.5 points (see table 2).
This minimal impact is expected since the per-
Since the identification of potential variant pairs
formance on “known words” is around 10 points
depends on the initial conditions of the experi-
higher than on OOV words in this setup. In fact,
ment, i.e. the corpus, and optionally, the lexica
considering the sizes of our corpora, lowering the
used to train the model beforehand, we present two
number of OOV words of 100 is expected to im-
experiments in which these parameters vary.
prove the overall results of 0.2 points.
For each experiment, we extract from the train-
ing corpus the vocabulary V T lookup and from
the external lexicon, the vocabulary V L lookup. 4.4 Experiment 2: Controlled Setup
We use the set of rules presented in Section 3.1. By “controlled”, we mean that training and eval-
We prioritize the lookup in V T lookup to fur- uation each contain a specific variant of Alsatian
ther ease the evaluation of the generated pairs re- selected in a multi-variant corpus. In the follow-
lying on the context. ing, we compare homogeneous and heterogeneous
If the length of the OOV word is less or equal to setups, in which the training and evaluation cor-
four characters (ˆ and $ excluded), only the L+R pora either contain the same or distinct variants of
rules are applied: it has been observed in prelimi- Alsatian.
nary tests that shorter words were more likely to
To highlight the effect of our methodology in
lead to erroneous matching such as das (deter-
an heterogeneous context, met when no corpus
miner) /dass (subordinating conjunction) or dien
of each possible variant is available, we manually
(auxiliary) /dene (determiner). Additionally, we
split Concat C in two sub-corpora North C (4,880
force the variant candidates to have the same letter
words) and South C (7,690 words) based on the
case as the OOV word.
frequencies of the -e and -a noun endings, which
After variant pairs have been generated, the
are specific of the Northern and Southern variants
OOV words are replaced by their variant candi-
respectively.
date, and the pre-trained model is applied on the
The results of these experiments are presented
transposed evaluation corpus. After the corpus has
in table 3.
been tagged, the transposed words are replaced by
they original form. Unsurprisingly, the best results are obtained
when training and evaluation corpora are of the
4.3 Experiment 1: Uncontrolled Setup same variant. Yet, we can observe that in this
By “uncontrolled”, we mean that training and setup, the effect of transposition to identified vari-
evaluation corpora are both extracted from a shuf- ants has a higher impact on the proportion of OOV
fled corpus that contains multiple variants. words and the tagging performances.
Our first model is trained with Concat C80 The efficiency of the methodology largely de-
(17,136 words) and evaluated on Concat C20 pends on: (i) the respective and relative sizes of
(4,374 words) before and after its transposi- the training and evaluation corpora, (ii) the varia-
tion. After the application of the three sets of tion in variants existing between them.
rules, using both the vocabularies extracted from This experiment shows that the performance of
Concat C80 and MultiV ar L for the lookup, 56 vari- a tool trained on a given corpus can be improved
ant pairs were discovered and the same number of by modifying the corpus it is applied on to match
words were transposed. the vocabulary it was trained with.
781
North C20 South C20
North C80 Before transp. After transp.
Overall 0.853 0.714 0.752
OOV words 40% 54% 52%
South C80 Before transp. After transp.
Overall 0.788 0.809 0.864
OOV words 51% 48% 29%
Table 3: Accuracy of the model trained on mono-variant corpora, before and after the corpus transposi-
tion.
782
for work described in (Theron and Cloete, 1997), Acknowledgments
in which the rules are automatically extracted but
We wish to thank the participants of Bisame and
with a known (morphological) goal.
Recettes de Grammaire for their motiva-
The closest work to ours is that described
tion and comments, as well as J-N. S. Kempf for
in (Barteld, 2017), as it focuses on detecting
his time and expertise on the Alsatian variants. We
spelling variants in Middle Low German unre-
also thank B. Sagot who inspired us with the title
lated to a standard. Yet, the described method
of this paper.
requires the training of a classifier to filter the
generated pairs. This classifier is based on a re-
source that contains 1,834 pairs of spelling vari- References
ants, a resource that is unavailable for most non-
Paul Adolf. 2006. Dictionnaire comparatif mul-
standardized languages. tilingue: français-allemand-alsacien-anglais.
Regarding Alsatian more specifically, Bernhard Midgard, Strasbourg, France.
(2014) aligns spelling variants relying on a multi-
variant bilingual French-Alsatian lexicon anno- Alistair Baron and Paul Rayson. 2008. Vard 2: A tool
for dealing with spelling variation in historical cor-
tated with part-of-speech and a phonetization of pora. In Aston University, editor, Proceedings of
Alsatian. This high dependency on existing re- the Postgraduate Conference in Corpus Linguistics,.
sources make this method challenging to adapt to Birmingham, UK.
other languages for which the only available ex-
Alistair Baron and Paul Rayson. 2009. Automatic stan-
perts are the very speakers of the language. dardisation of texts containing spelling variation:
How much training data do you need? In Univer-
7 Conclusion sity of Liverpool, editor, Proceedings of the Corpus
Linguistics Conference. Liverpool, UK.
We have presented a method to automatically gen-
erate pairs of spelling variants based on a small Corinne Barre and Mélanie Vanderschelden. 2004.
L’enquête ”étude de l’histoire familiale” de 1999 -
subset of crowdsourced pairs. Résultats détaillés. INSEE, Paris.
The method does not require manual rules def-
inition by experts and is language independent. Fabian Barteld. 2017. Detecting spelling variants in
non-standard texts. In Proceedings of Student Re-
The resources needed to perform variant pair de- search Workshop (EACL 2017). Valencia, Spain.
tection can be easily produced by the speakers,
who hold the knowledge of the of the variation Delphine Bernhard. 2014. Adding dialectal lexical-
mechanisms. The crowdsourcing of variants, un- isations to linked open data resources: the exam-
ple of alsatian. In Proceedings of the Workshop
like that of POS tags, requires no prior training. on Collaboration and Computing for Under Re-
In fact, even the expertise necessary for the vali- sourced Languages in the Linked Open Data Era
dation of the variant pairs is about to be transferred (CCURL 2014). Reykjavik, Iceland, pages 23–29.
to the participants of the crowdsourcing platform. https://hal.archives-ouvertes.fr/hal-00966820.
The originality of this methodology is that once Delphine Bernhard. 2018. Lexicon of
the rules have been extracted, the process feeds place names in the alsatian dialects.
from previously unseen texts. This is particularly https://doi.org/10.5281/zenodo.1404873.
useful in a less-resourced scenario where a raw Delphine Bernhard, Pascale Erhart, Dominique Huck,
corpus is being collected from various sources. and Lucie Steiblé. 2018a. Annotated corpus for the
The code of both the gamified crowdsourc- alsatian dialects. Guide d’annotation, LiLPa, Uni-
ing platform and the variants generation is freely versité de Strasbourg.
available on GitHub7 . The created multi-variant Delphine Bernhard and Anne-Laure Ligozat. 2013.
lexicon is also available under a CC license. Es esch fàscht wie Ditsch, oder net? étiquetage
We plan to extend this work to other non- morphosyntaxique de l’alsacien en passant par
l’allemand. In Proceedings of TALARE (Traitement
standardized languages. We have started working Automatique des Langues Régionales de France et
on adapting the platform to Mauritian, a French- d’Europe) (TALN’13). Les Sables d’Olonne, France,
based Creole, the morphology of which is very pages 209–220.
different from that of Alsatian.
Delphine Bernhard, Anne-Laure Ligozat, Fanny
7
See https://github.com/alicemillour/ Martin, Myriam Bras, Pierre Magistry, Mari-
Bisame/tree/recipes. anne Vergez-Couret, Lucie Steible, Pascale Erhart,
783
Nabil Hathout, Dominique Huck, Christophe Rey, Alice Millour and Karën Fort. 2018b. Toward
Philippe Reynés, Sophie Rosset, Jean Sibille, and a Lightweight Solution for Less-resourced Lan-
Thomas Lavergne. 2018b. Corpora with Part- guages: Creating a POS Tagger for Alsatian Using
of-Speech Annotations for Three Regional Lan- Voluntary Crowdsourcing. In Proceedings of 11th
guages of France: Alsatian, Occitan and Pi- International Conference on Language Resources
card. In Proceedings of 11th edition of the and Evaluation (LREC’18). Miyazaki, Japan.
Language Resources and Evaluation Conference https://hal.archives-ouvertes.fr/hal-01790615.
(LREC 2018). Miyazaki, Japan. https://hal.archives-
ouvertes.fr/hal-01704806. Slav Petrov, Dipanjan Das, and Ryan McDonald. 2012.
A universal part-of-speech tagset. In Proceed-
Jon Chamberlain, Karën Fort, Udo Kruschwitz, Math- ings of 8th International Conference on Language
ieu Lafourcade, and Massimo Poesio. 2013. Using Resources and Evaluation (LREC 2012). Istanbul,
games to create language resources: Successes and Turkey.
limitations of the approach. In Iryna Gurevych and
Jungi Kim, editors, The People’s Web Meets NLP, Jelena Prokić, Martijn Wieling, and John Nerbonne.
Springer Berlin Heidelberg, Theory and Applica- 2009. Multiple sequence alignments in linguistics.
tions of Natural Language Processing, pages 3–44. In Proceedings of the EACL 2009 Workshop on
https://doi.org/10.1007/978-3-642-35085-6 1. Language Technology and Resources for Cul-
tural Heritage, Social Sciences, Humanities,
Danielle Crévenat-Werner and Edgar Zeidler. 2008. and Education. Association for Computational
Orthographe alsacienne - Bien écrire l’alsacien de Linguistics, Stroudsburg, PA, USA, pages 18–25.
Wissembourg à Ferrette. Jérôme Do Bentzinger. http://dl.acm.org/citation.cfm?id=1642049.1642052.
Lucie Steiblé and Delphine Bernhard. 2016. Towards
Pascal Denis and Benoı̂t Sagot. 2012. Coupling an
an Open Lexicon of Inflected Word Forms for Al-
annotated corpus and a lexicon for state-of-the-art
satian: Generation of Verbal Inflection. In JEP-
pos tagging. Lang. Resour. Eval. 46(4):721–736.
TALN-RECITAL 2016. Paris, France, volume 2 of
https://doi.org/10.1007/s10579-012-9193-0.
Proceedings of la conférence conjointe JEP-TALN-
Izaskun Etxeberria Uztarroz, Iñaki Alegrı́a Loinaz, RECITAL 2016, volume 2 : TALN, pages 547–554.
Mans Hulden, and Larraitz Uria Garin. 2014. Learn- https://hal.archives-ouvertes.fr/hal-01338411.
ing to map variation-standard forms in basque using Pieter Theron and Ian Cloete. 1997. Automatic
a limited parallel corpus and the standard morphol- acquisition of two-level morphological rules.
ogy. Procesamiento del Lenguaje Natural 52:13– In Proceedings of the Fifth Conference on
20. Applied Natural Language Processing. Associ-
ation for Computational Linguistics, Strouds-
Kimmo Koskenniemi. 2013. Finite-state relations be- burg, PA, USA, ANLC ’97, pages 103–110.
tween two historically closely related languages. In https://doi.org/10.3115/974557.974573.
Northern European Association for Language Tech-
nology, editor, Proceedings of the workshop on
computational historical linguistics at NODALIDA
2013. Oslo , Norway, volume 18, pages 43–53.
784
Neural Feature Extraction for Contextual Emotion Detection
785
Proceedings of Recent Advances in Natural Language Processing, pages 785–794,
Varna, Bulgaria, Sep 2–4 2019.
Dini and Bittar (2016) broke down the task focused on domain-specific emotion detection.
of emotion detection from tweets into a cascade They created a dataset of 2107 sentences taken
of decisions: classifying tweets into emotional from online forums on the Cancer Survivors
and non-emotional categories, and then tagging Network website1 . In order to combine the
the emotional tweets with the appropriate emo- strengths of lexicon-based and machine learn-
tion label. For the latter, they compared a sym- ing approaches, they proposed a model that uses
bolic system using gazetteers, regular expressions, word2vec embeddings as input to a Convolutional
and graph transformation, with a machine learning Neural Network (CNN) (LeCun et al., 1999). The
system using a linear classifier with words, lem- CNN generates feature vectors which are then
mas, noun phrases, and dependencies as features. augmented with domain-specific lexical features.
Using their collected corpus of emotional tweets, The combined features are then used as input to
the rule-based approach achieved an F1 score of an LSTM network which classifies the texts into 6
0.41, while the machine learning approach yielded different emotion categories.
an F1 score of 0.58 on 6 emotion classes. While most of the literature has focused on
Mohammad and Bravo-Marquez (2017) made the detection and assessment of emotions in on-
use of an SVM regression model to determine the line textual data, few researchers have investigated
intensity of 4 emotions: anger, fear, joy, and sad- emotion detection in textual conversations. We ar-
ness in a dataset of tweets that they have previ- gue that the detection of emotions from dialogues
ously collected and annotated. As features, they poses new challenges compared to emotion detec-
used word and character n-grams, word embed- tion from monologues, as the utterances made by
dings trained using the word2vec skip-gram model different interlocutors can influence differently the
(Mikolov et al., 2013), and affect-related lexical emotional state of a speaker.
features. Using the Pearson correlation coefficient In this work, we investigate the effectiveness of
as evaluation metric, they demonstrated that word neural feature extraction for the task of emotion
embeddings yield better results than n-gram fea- detection in short dialogues.
tures. They achieved their best average result of
0.66, using a combination of word embeddings
and lexical features. 3 Dataset and Task
Abdul-Mageed and Ungar (2017) also collected
their own dataset of emotional tweets using emo- The dataset used in this work is taken from Chat-
tion hashtags. They trained word embeddings on terjee et al. (2019b). It consists of short 3-turn
the training data, employed a gated recurrent neu- dialogues between two speakers (turn 1 uttered by
ral network (Cho et al., 2014) as a classifier and speaker 1, turn 2 uttered by speaker 2, and turn 3
achieved an average F1 score of 0.87 over 3 emo- uttered by speaker 1 again). Table 1 shows two
tion datasets, labelled with 8 emotions. samples of the dataset2 .
Abdullah and Shaikh (2018) proposed an ap- The goal is to detect the emotion of speaker 1 in
proach to detect the intensity of affect in tweets. turn 3, taking into account the previous turns. The
Their features include feature vectors extracted us- data is annotated with 4 emotions: happy, angry,
ing the AffectiveTweets package of Weka (Holmes sad, and others. In order to simulate a real-life
et al., 1994), as well as word2vec and doc2vec (Le task, the distribution of the labels in the dataset is
and Mikolov, 2014) embeddings. They developed highly imbalanced: 50% of the training data be-
three models using different subsets of the feature longs to the others class, while 14%, 18%, and
set as input to either a dense feed-forward network 18% of the training data is dedicated to classes
or a Long Short-Term Memory (LSTM) (Hochre- happy, angry, and sad, respectively. The test and
iter and Schmidhuber, 1997) network. Using the development sets are even more imbalanced, with
dataset of SemEval 2018 task 1 (Affect in Tweets) 85% of the samples labelled as others. Table 2
(Mohammad et al., 2018), they achieved their best summarizes some statistics of the dataset.
Spearman correlation score of 0.69 over 4 emo-
tions by averaging over the outputs of the three 1
https://csn.cancer.org/
models. 2
Samples are taken from https://competitions.
More recently, Khanpour and Caragea (2018) codalab.org/competitions/19790.
786
ID Turn1 (Speaker1) Turn2 (Speaker2) Turn3 (Speaker1) Label (of Turn3)
156 You are funny LOL I know that. :p , happy
187 Yeah exactly Like you said, like brother like sister ;) Not in the least others
Average # of Tokens
Dataset Label # of Samples Percentage The Recurrent Layer takes as input the token
vectors ([xi,1 , xi,2 , . . . , xi,t , . . . , xi,n ]), and pro-
Turn 1 Turn 2 Turn 3
happy 4243 14% 4.873 7.195 3.825
angry 5506 18% 5.107 6.859 5.457
Train sad 5463 18% 4.608 6.450 4.829 cesses them in a forward and a backward passes.
others 14948 50% 4.232 6.493 4.153
All 30160 100% 4.550 6.650 4.467
In the forward pass, the content value of the hidden
happy 142 5% 4.761 7.444 3.690 layer at a specific time-step is calculated using the
angry 150 5% 4.647 7.347 4.867
Development sad 125 5% 4.624 6.200 5.192 value of the input at the current time-step, and the
others 2338 85% 4.245 6.546 4.143
All 2755 100% 4.311 6.620 4.207
content value of the hidden layer in the previous
happy 284 5% 5.063 6.845 3.493 time-step.
angry 298 5% 4.470 6.456 4.856
Test sad 250 5% 5.000 6.632 4.936 Equation 1 shows how the content value of the
others 4677 85% 4.279 6.601 4.143
All 5509 100% 4.362 6.607 4.184
hidden layer is calculated at a specific time-step t,
happy 4669 12% 4.881 7.181 3.801 where xt represents the input value in the current
angry 5954 16% 5.063 6.851 5.412
All sad 5838 15% 4.626 6.452 4.842 time-step, and ht and ht±1 represent the content
others 21963 57% 4.243 6.521 4.150
All 38424 100% 4.506 6.642 4.408
value of the hidden node in the current and pre-
vious/next time-steps (in the forward or backward
Table 2: Statistics of the EmoContext 2019 pass), respectively, and fh is the function that cal-
dataset. culates the value of ht using xt and ht±1 . Subse-
quently, the output of the hidden layer is calculated
using Equation 2, where yt is the output of the hid-
4 The Model
den layer at time-step t, and fy is the function that
Figure 1 shows the overall architecture of our calculates the output value based on ht .
model for the task of contextual emotion detection.
The model is composed of two main components: ht = fh (xt , ht±1 ) (1)
1) the neural feature extractor and 2) the classifier.
yt = fy (ht ) (2)
4.1 The Neural Feature Extractor
The Attention Layer is a function that auto-
As shown in Figure 1, the neural feature extrac- matically assigns weights to the output of the re-
tor is a recurrent neural network with an attention current layer at each time-step, and calculates the
mechanism. The feature extractor is responsible weighted sum of the outputs using their corre-
for creating dense vector representations for each sponding weights (Vaswani et al., 2017). Follow-
dialogue turn. As a result, the model uses 3 feature ing several works that have shown significant im-
extractors, one for each dialogue turn. provement in text classification with the use of
Each neural feature extractor is composed of attention (e.g. Yang et al., 2016; Zhou et al.,
an input layer, a recurrent layer, and an attention 2016; Wang et al., 2016; Cianflone et al., 2018),
layer, explained below. we incorporated an attention mechanism in our
contextual emotion detection framework. Equa-
The Input Layer takes as input the vector rep-
tion 3 shows the overall mechanism of our atten-
resentations of each word in the correspond-
tion layer, where ωt′ represents the corresponding
ing dialogue turn. Each dialogue turn is a
weight for the output of the recurrent layer at time-
sequence of tokens, represented as a vector
step t′ in, and n is the number of time-steps (i.e.
[xi,1 , xi,2 , . . . , xi,t , . . . , xi,n ], where xi,t is the cor-
the length of the dialogue turn).
responding vector for the t-th word in the i-th dia-
logue turn, and n is the length of the i-th turn. The n
vector representation for each token (xi,t ) is com- Attention = ∑ yt′ ωt′ (3)
posed of the word embedding corresponding to the t′ =1
token, concatenated with a one-hot representation In our model, the weights are calculated by
of the token’s part-of-speech (POS) tag. applying a single N -to-1 feed-forward layer on
787
Figure 1: Architecture of the model.
the output of the recurrent layer at each time-step et al., 2014), which is pretrained on 840B tokens
(where N is the size of the output of the recurrent of web data from Common Crawl, and provides
layer), concatenating the results, and applying a 300d vectors as word embeddings. As our sec-
softmax over them. Equations 4 and 5 show the ond word embedder, we experimented with ELMo
mechanisms used to calculate the weights, where (Peters et al., 2018), which produces word embed-
w corresponds to the weights in the single-layer dings of size 1024, and is pretrained on the 1 Bil-
neural network, and νt is the single value, which lion Word Language Model Benchmark3 (Chelba
is the result of feeding yt to the fully-connected et al., 2014).
layer. The main reason for choosing these two word
embedders was to evaluate the effect of their em-
νt = yt × w (4) bedding mechanisms for our task. As opposed to
ω = Sof tmax([ν1 , ν2 , ν3 , . . . , νn ]) (5) GloVe which assigns a word embedding to each
token, the ELMo word embedder calculates the
4.2 The Classifier embedding for each token from its constituent
As shown in Figure 1, we experimented with two characters by also taking into account its textual
types of classifiers at the output layer: A fully- context. We suspected that this approach would
connected neural network, followed by a softmax lead to better results in our task (see Section 6).
activation function, and an SVM, which takes as 5.2 POS Tags
input the neural representations generated by the 3
latent feature extractors for each dialogue turn. The spaCy library4 was used for tokenization and
The neural classifier is trained jointly with the POS tagging, and the Penn Treebank tagset stan-
neural feature extractors, while the SVM classifier dard (Marcus et al., 1993) was followed for assign-
is completely trained after each training epoch of ing POS tags to tokens. This lead to one-hot vec-
the neural network, using the features extracted by tors for POS information of size 51.
the 3 neural feature extractors. 5.3 Recurrent Units
5 Experimental Setup Long Short-term Memory (LSTM) (Hochreiter
and Schmidhuber, 1997) and Gated Recurrent
The neural network components of our model Units (GRU) (Cho et al., 2014) were both exper-
were developed using PyTorch (Paszke et al., imented with as the building blocks of the recur-
2017) and the SVM was developed using the rent layer. For both LSTM and GRU, 2 layers of
Scikit-learn library (Pedregosa et al., 2011). In this 25 bidirectional recurrent units were stacked.
section, we will explain the different setups that
we experimented for the task of contextual emo- 5.4 Neural Network Optimization
tion detection. The Adam optimizer (Kingma and Ba, 2014) with
a learning rate of 10−4 was used to train the neu-
5.1 Word Embeddings
3
In order to test our model, we experimented with The selected versions of GloVe and ELMo lead to the
best results in our task. We also experimented with other ver-
two different pretrained word embeddings. As the sions of the two, but their performances were inferior.
first word embedder, we chose GloVe (Pennington 4
https://spacy.io/
788
ral network. Cross-entropy with class weights was Figure 2 shows the performance of each model
used as the loss function. In order to handle the on the development dataset, throughout the train-
imbalanced class distribution in the dataset (see ing process. As Figure 2 shows, using both the
Table 2), the corresponding weight for each class neural classifier and the SVM classifier, the mod-
was calculated proportional to the inverse of the els with GRU as the recurrent units and ELMo em-
number of samples for that class in the training beddings as input features were generally superior
data. This way, more penalty was applied to the to the others.
network when an error was made on a sample from The notation <type-of-recurrent-unit>+<type-
a minority class rather than a more frequent one. of-classifier> with <type-of-input-features> is
Minibatches of size 32 were used during train- used in the rest of the paper, to refer to each
ing and testing, and zero-padding was applied in model; for example, LSTM+NN with GloVe refers
order to handle different input sequence lengths. to the model that uses LSTM units in the recur-
In order to minimize padding, samples with sim- rent layer, fully-connected neural layer as clas-
ilar average lengths of tokens over the three turns sifier, and GloVe embeddings as input; whereas
were put in the same batch. GRU+SVM with ELMo+POS denotes the version
Finally, in order to avoid the exploding gradient of our model with GRU units in the recurrent layer,
problem (Pascanu et al., 2012), gradient clipping SVM as the classifier, and ELMo embeddings and
with a norm of 0.5 was applied. one-hot encoded POS tags as input.
As indicated in Section 5.6, the final versions
5.5 SVM Hyperparameters of the models were chosen based on their perfor-
The SVM utilizes a polynomial kernel with de- mance on the development dataset, i.e. for each
gree of 4. To set the parameter γ, the svm.SVC model, the final set of trained parameters were the
model in Scikit-learn was initiated with its param- one that yielded the maximum micro-average F1
eter gamma set to auto, which automatically sets score on the three emotion classes on the develop-
γ to the inverse of the number of features extracted ment dataset.
by the neural feature extractor. In our model, this The results achieved from our models are also
value was set to 1/150, since each of the three neu- compared with the baseline system, provided by
ral feature extractors extracts 50 features from the the EmoContext 2019 shared task (Chatterjee
dialogue turn that it handles. et al., 2019a). The baseline system is composed
of a neural network with 128 LSTM units in the
5.6 Overall Training Process hidden layer, and as input features, uses the GloVe
As indicated in Section 4.2, the neural classifier word embeddings, pretrained on 6 billion tokens
was trained jointly with the neural feature extrac- from Wikipedia 2014 and the Gigaword 5 corpus5 .
tors, while the SVM was trained separately after Table 3 shows the performance of each model6 ,
each epoch, using the extracted features on the where the best micro-average F1 scores are high-
training data. lighted in bold. The results show that the model
The models with either neural or SVM classi- GRU+SVM with ELMo yields the best perfor-
fier were trained for 50 epochs, and the model’s mance of 73.03% on the development data, while
parameters were saved after each training epoch. the model GRU+SVM with ELMo+POS outper-
The optimal parameters were then picked as the forms all the other models on the test dataset with
ones that led to the highest micro-average F1 score a micro-average F1 score of 69.93%, by being
on the three main emotion classes (all except class marginally better than GRU+SVM with ELMo7 .
other) on the development dataset. This final The results also show that, with the exception
model with the optimal trained parameters was of the two models LSTM+NN with GloVe and
then evaluated on the test set. GRU+NN with GloVe which have inferior perfor-
5
https://catalog.ldc.upenn.edu/
6 Results LDC2011T07
6
At the time of writing this paper, the F1 score of the base-
The official evaluation metric used at the Emo- line model had not been released for each emotion class.
7
Context shared task is the micro-average F1 score The result that we submitted to the EmoContext shared
task was slightly higher than the current result (70.72%),
over the three main emotion classes, i.e. happy, which was achieved by also using the features from the de-
angry, and sad (ignoring the 4th class, others). velopment dataset to train the SVM classifier.
789
(a) With the neural classifier. (b) With the SVM classifier.
Figure 2: Performance of the model (in micro-average F1 score over the three main emotion classes) on
the development dataset, as a function of the number of training epochs
mance than the baseline system on the test dataset, 7.2 Effect of the Recurrent Units
all the other models significantly outperform the The results in Table 3 show that for the models that
baseline model on both development and test data. incorporate ELMo embeddings, the ones that use
GRU in their recurrent layer significantly outper-
7 Discussion form the ones with LSTM; however, this behav-
ior cannot be observed in GloVe-based models, as
we can see several cases, where the LSTM-based
To better understand the results, we analyzed the models are slightly better.
effect of different components of the model.
It could be concluded that, for the current
task, since GloVe word embeddings are context-
7.1 Effect of Input Features independent, a stronger recurrent unit is required
to capture the context, while in the case of ELMo,
The results in Table 3 demonstrate that the models where context is already taken into account, a sim-
that use ELMo as word embeddings have a sig- pler recurrent unit such as GRU is enough while
nificantly higher performance than the ones with being less prone to overfitting.
GloVe. We believe that this is due to two main rea-
sons: 1) The ELMo word embedder is character- 7.3 Effect of the Classifiers
based, which allows it to better handle out-of- Table 3 shows that, in almost all cases, the models
vocabulary words, and 2) ELMo takes into ac- with the SVM classifier significantly outperform
count the textual context of the token when ex- the ones with the neural classifier. We believe
tracting the word embedding. that, although the neural network may have been
Table 3 also shows that the use of POS tags able to reach similar results as the SVM, the lat-
leads to a significant improvement with the models ter reached this performance using less fine-tuning
that utilize GloVe word embeddings. On the other due to its explicit design to optimize the margin
hand, for the models that use ELMo embeddings size between classes.
as input, this is not the case. As Table 3 shows, On another note, Figure 2 shows that the models
in several occasions, the use of POS tags has even with the SVM classifier were significantly more
reduced the performance of ELMo-based models. robust and demonstrated much less performance
We believe that, in the case of GloVe, where the fluctuation during training than the ones with the
word embeddings are context-independent, POS neural classifier. We believe that this is due to the
tags can improve token representations to also take more deterministic nature of SVM in comparison
into account the textual context in which the to- to the neural networks.
kens have occurred. However, since ELMo al- However, the most important drawback of the
ready takes into account the textual context when SVM was the training time: since the SVM classi-
extracting the token representations, POS tags do fier was trained separately from feature extractors,
not help much and can even be redundant in some training it entailed additional training time to the
cases. model. All being said, we believe that, if training
790
F1 Score on Development Data F1 Score on Test Data
Model Input Feature
happy angry sad Micro happy angry sad Micro
BASELINE – – – 58.61 – – – 58.68
GloVe 53.45 67.37 59.39 60.40 52.05 67.02 52.89 57.60
Glove+POS 58.33 68.28 62.54 63.15 58.03 68.68 59.77 62.35
LSTM+NN
ELMo 63.88 67.61 69.14 66.74 62.07 65.14 67.37 64.74
ELMo+POS 59.03 65.92 68.42 64.13 63.95 62.06 68.56 64.54
GloVe 47.79 65.67 66.19 59.36 51.34 65.80 58.04 58.26
Glove+POS 57.94 68.39 64.74 63.77 61.24 69.27 57.82 62.99
GRU+NN
ELMo 64.26 67.58 74.71 68.34 63.32 66.21 69.73 66.33
ELMo+POS 65.48 65.15 73.12 67.46 62.20 66.40 71.64 66.53
GloVe 54.85 67.06 66.91 65.83 53.00 68.05 57.30 62.84
Glove+POS 61.30 68.95 66.20 66.46 60.30 67.32 60.84 63.26
LSTM+SVM
ELMo 65.64 65.91 73.44 68.94 63.78 65.25 70.10 66.45
ELMo+POS 62.31 68.44 68.73 67.25 63.37 64.09 68.86 67.28
GloVe 50.21 68.31 71.00 65.08 50.22 70.26 60.13 62.02
Glove+POS 52.77 68.73 66.67 65.42 56.00 69.68 57.96 63.11
GRU+SVM
ELMo 67.33 67.83 75.40 73.03 65.08 68.83 73.07 69.39
ELMo+POS 68.21 66.26 74.46 70.03 64.71 69.13 71.05 69.93
AVERAGE 59.55 67.34 68.82 65.96 59.42 67.07 64.07 64.22
Table 3: The performance of each model on the development and test datasets, in terms of F1 score
on each emotion class, and micro-average F1 over the three main emotion classes. The AVERAGE is
computed over the proposed models and does not include the baseline.
time is not a concern, the SVM classifier is a better 7.5 Quality of the Extracted Neural Features
option than the neural one. To better understand the contribution of the ex-
An interesting finding regarding the SVM clas- tracted neural features from the feature extractors,
sifier is that, in contrast to the neural network we calculated the mutual information between the
where applying class-weights to the loss function values of each neural feature and the classes.
helped improve the performance of the models, Figure 3 shows the average and the standard de-
applying class-weights to the SVM decreased its viation of the mutual information between the fea-
performance. tures extracted from each neural feature extractor
in each model and the classes in the training data.
7.4 A Closer Look at Emotion Classes Since both the neural and the SVM classifiers use
the same set of neural features, we did not differ-
The row labelled AVERAGE in Table 3 provide entiate between models with similar neural feature
information regarding the difficulty of detecting extractors and different classifiers.
each class. Table 3 shows that, among the three
main classes, happy, angry, and sad, the class
happy was the most difficult to detect.
Table 2 shows that the low average F1 score for
class happy is probably due to the significantly
smaller number of samples with this class in the
training data (14%) in comparison to the sam-
ples from the other two emotion classes (18% and
18%). Although the weighted loss functions (see
Section 5.4) somehow managed to handle the im-
balanced class distribution in the data, the optimal Figure 3: The average mutual information be-
weights are not necessarily proportional to the in- tween the features of each dialogue turn, extracted
verse of the frequency of classes. by each neural feature extractor, and the classes.
791
As Figure 3 shows, in all cases, the features ex- 8 Conclusion and Future Work
tracted from the third turn of the conversation have
the highest mutual information with the classes, In this paper, we proposed a model for the task
and the ones from the second turn have the lowest. of contextual emotion detection. We evaluated
This agrees with the nature of the dataset, where our model with the EmoContext 2019 shared task
the label is assigned to the emotion of speaker 1 dataset (Chatterjee et al., 2019b), which consists
(who uttered dialogue turns 1 and 3) after the third of 3 turn conversations tagged with one of the la-
turn is uttered. This also indicates that the nature bels: happy, sad, angry, and others, based on the
of emotion detection in the context of dialogues emotion present in the last dialogue turn.
is different from that in monologues in two ways: The proposed model utilizes an attention-based
not only do the utterances by different speakers recurrent neural network. We experimented with
contribute differently to the emotional state of a GloVe and ELMo embeddings, alongside POS
speaker, but also the timing of the utterances by tags as input, LSTM and GRU as recurrent units,
the same speaker has an impact on the contribu- and a neural or an SVM classifier. The best re-
tion of that utterance to the emotion classification. sult on the test dataset was achieved with ELMo
and POS tags as input, GRU as recurrent units,
Figure 3 shows that the difference in average and SVM as the final classifier. Using this setup,
mutual information between the extracted features we reached a performance of 69.93% in terms of
and the emotional classes are higher for features micro-average F1 score which is a significant im-
from the third (i.e. the most recent) dialogue provement over the baseline of 58.68%.
turn. As expected, the features from the ELMo- Three future directions can be proposed. The
based models have significantly higher mutual in- first is to investigate a more effective way of han-
formation with the classes than the ones from the dling the imbalanced distribution of labels in the
GloVe-based models. The features from the third dataset. As a example, methods for finding the op-
dialogue turn in models with GRU have slightly timal class-weights for training the models can be
higher mutual information than the ones from the investigated.
models with LSTM. However, the standard devia-
Secondly, the use of different number of fea-
tion between the features extracted by the GRU-
tures for each dialogue turn can be studied. As
based models are significantly smaller than the
shown in Figure 3, features extracted from differ-
ones with LSTM, showing that between the mod-
ent dialogue turns had different levels of contribu-
els with LSTM and the ones with GRU, the fea-
tion to the final classification. In that case, more
tures extracted from the models with GRU had
features could be extracted from turns 3 and 1 (ut-
more similar amount of contribution to the clas-
tered by the same speaker) in comparison to turn 2,
sification task than the ones from the LSTM. This
which has the least contribution to the classifica-
has led to the GRU-and-ELMo-based models out-
tion.
performing the others.
Lastly, knowing that the SVM classifier is capa-
A surprising finding is that the neural features ble of outperforming the neural one, studies can be
extracted by the GRU-based models from the sec- performed in order to make the extracted features
ond dialogue turn have the least mutual informa- more suitable for the SVM classifier.
tion with the classes in comparison to the ones
from the other models. Observing this, we ex- Acknowledgments
perimented with our classifiers by disregarding the
neural features from the second turn; however, this We would like to express our gratitude to Parsa
lead to a slight performance drop. We hypoth- Bagherzadeh for insightful discussions through-
esize that, although the neural features from the out the course of this work. We would also like
second dialogue turns bring only a small contribu- to thank the anonymous reviewers for their com-
tion to the classification, the GRU-based models ments on an earlier version of this paper.
tend to focus more on the features from the other This work was financially supported by the Nat-
two turns. This leads to the second feature extrac- ural Sciences and Engineering Research Council
tor being less focused upon, and as a result, being of Canada (NSERC).
less trained than the ones from other models.
792
References Sepp Hochreiter and Jürgen Schmidhuber. 1997.
Long short-term memory. Neural Computation
Muhammad Abdul-Mageed and Lyle Ungar. 2017. 9(8):1735–1780.
Emonet: Fine-grained emotion detection with gated
recurrent neural networks. In Proceedings of Geoffrey Holmes, Andrew Donkin, and Ian H. Wit-
the 55th Annual Meeting of the Association for ten. 1994. Weka: a machine learning workbench.
Computational Linguistics (ACL 2017). Vancouver, In Proceedings of ANZIIS 1994 - Australian New
Canada, pages 718–728. Zealnd Intelligent Information Systems Conference.
Brisbane, Australia, pages 357–361.
Malak Abdullah and Samira Shaikh. 2018. Teamuncc
at SemEval-2018 task 1: Emotion detection in En- Hamed Khanpour and Cornelia Caragea. 2018. Fine-
glish and Arabic tweets using deep learning. In Pro- grained emotion detection in health-related online
ceedings of the 12th International Workshop on Se- posts. In Proceedings of the 2018 Conference on
mantic Evaluation (SemEval 2018). New Orleans, Empirical Methods in Natural Language Process-
Louisiana, USA, pages 350–357. ing (EMNLP 2018). Brussels, Belgium, pages 1160–
1166.
Ankush Chatterjee, Umang Gupta, Manoj Kumar
Chinnakotla, Radhakrishnan Srikanth, Michel Gal- Diederik P. Kingma and Jimmy Ba. 2014. Adam: A
ley, and Puneet Agrawal. 2019a. Understanding method for stochastic optimization. Computing Re-
emotions in text using deep learning and big data. search Repository arXiv:1412.6980.
Computers in Human Behavior 93:309–317.
Quoc Le and Tomas Mikolov. 2014. Distributed repre-
Ankush Chatterjee, Kedhar Nath Narahari, Meghana sentations of sentences and documents. In Proceed-
Joshi, and Puneet Agrawal. 2019b. SemEval-2019 ings of the 31st International Conference on Ma-
task 3: EmoContext: Contextual emotion detection chine Learning (ICML 2014). Bejing, China, pages
in text. In Proceedings of The 13th International 1188–1196.
Workshop on Semantic Evaluation (SemEval-2019).
Minneapolis, Minnesota, USA. Yann LeCun, Patrick Haffner, Léon Bottou, and
Yoshua Bengio. 1999. Object recognition with
Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, gradient-based learning. In Shape, contour and
Thorsten Brants, Phillipp Koehn, and Tony Robin- grouping in computer vision, Springer, pages 319–
son. 2014. One billion word benchmark for 345.
measuring progress in statistical language model-
ing. In 15th Annual Conference of the Interna- Lan Li and Ji-hua Chen. 2006. Emotion recognition
tional Speech Communication Association (INTER- using physiological signals. In Proceedings of 16th
SPEECH 2014). Singapore. International Conference on Artificial Reality and
Telexistence (ICAT 2006). Hangzhou, China, pages
Kyunghyun Cho, Bart van Merrienboer, Caglar Gul- 437–446.
cehre, Dzmitry Bahdanau, Fethi Bougares, Holger
Jasy Suet Yan Liew and Howard R. Turtle. 2016. Ex-
Schwenk, and Yoshua Bengio. 2014. Learning
ploring fine-grained emotion detection in tweets. In
phrase representations using RNN encoder–decoder
Proceedings of the NAACL Student Research Work-
for statistical machine translation. In Proceedings of
shop. San Diego, California, USA, pages 73–80.
the 2014 Conference on Empirical Methods in Nat-
ural Language Processing (EMNLP 2014). Doha, Qi-rong Mao, Xin-yu Pan, Yong-zhao Zhan, and
Qatar, pages 1724–1734. Xiang-jun Shen. 2015. Using kinect for real-time
emotion recognition via facial expressions. Fron-
Andre Cianflone, Yulan Feng, Jad Kabbara, and Jackie tiers of Information Technology & Electronic Engi-
Chi Kit Cheung. 2018. Let’s do it “again”: A first neering 16(4):272–282.
computational approach to detecting adverbial pre-
supposition triggers. In Proceedings of the 56th Mitchell P. Marcus, Mary Ann Marcinkiewicz, and
Annual Meeting of the Association for Computa- Beatrice Santorini. 1993. Building a large annotated
tional Linguistics (ACL 2018). Melbourne, Aus- corpus of English: The Penn Treebank. Computa-
tralia, pages 2747–2755. tional Linguistics 19(2):313–330.
Luca Dini and André Bittar. 2016. Emotion analysis Walaa Medhat, Ahmed Hassan, and Hoda Korashy.
on twitter: The hidden challenge. In Proceedings 2014. Sentiment analysis algorithms and applica-
of the Tenth International Conference on Language tions: A survey. Ain Shams Engineering Journal
Resources and Evaluation (LREC 2016). Portorož, 5(4):1093–1113.
Slovenia.
Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey
Ruchi Hirat and Namita Mittal. 2015. A survey on Dean. 2013. Efficient estimation of word represen-
emotion detection techniques using text in blog- tations in vector space. In Workshop Proceedings of
posts. International Bulletin of Mathematical Re- the International Conference on Learning Represen-
search 2(1):180–187. tations (ICLR 2013). Scottsdale, Arizona, USA.
793
Saif Mohammad and Felipe Bravo-Marquez. 2017. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Emotion intensities in tweets. In Proceedings of the Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
6th Joint Conference on Lexical and Computational Kaiser, and Illia Polosukhin. 2017. Attention is all
Semantics (*SEM 2017). Vancouver, Canada, pages you need. In I. Guyon, U. V. Luxburg, S. Ben-
65–77. gio, H. Wallach, R. Fergus, S. Vishwanathan, and
R. Garnett, editors, Advances in Neural Information
Saif Mohammad, Felipe Bravo-Marquez, Moham- Processing Systems 30 (NIPS 2017), Curran Asso-
mad Salameh, and Svetlana Kiritchenko. 2018. ciates, Inc., Long Beach, California, USA, pages
SemEval-2018 task 1: Affect in tweets. In 5998–6008.
Proceedings of The 12th International Workshop
on Semantic Evaluation (SemEval 2018). Associa- Yequan Wang, Minlie Huang, Xiaoyan Zhu, and
tion for Computational Linguistics, New Orleans, Li Zhao. 2016. Attention-based LSTM for aspect-
Louisiana, pages 1–17. level sentiment classification. In Proceedings of the
2016 Conference on Empirical Methods in Natu-
Razvan Pascanu, Tomas Mikolov, and Yoshua Ben- ral Language Processing (EMNLP 2016). Austin,
gio. 2012. Understanding the exploding gradi- Texas, USA, pages 606–615.
ent problem. Computing Research Repository
arXiv:1211.5063v1. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He,
Alex Smola, and Eduard Hovy. 2016. Hierarchi-
Adam Paszke, Sam Gross, Soumith Chintala, Gre- cal attention networks for document classification.
gory Chanan, Edward Yang, Zachary DeVito, Zem- In Proceedings of the 2016 Conference of the North
ing Lin, Alban Desmaison, Luca Antiga, and Adam American Chapter of the Association for Computa-
Lerer. 2017. Automatic differentiation in PyTorch. tional Linguistics: Human Language Technologies
In NIPS 2017 Autodiff Workshop. Long Beach, Cal- (NAACL-HLT 2016). San Diego, California, USA,
ifornia, USA. pages 1480–1489.
Fabian Pedregosa, Gaël Varoquaux, Alexandre Gram- Mohamed Yassine and Hazem Hajj. 2010. A frame-
fort, Vincent Michel, Bertrand Thirion, Olivier work for emotion mining from text in online social
Grisel, Mathieu Blondel, Peter Prettenhofer, Ron networks. In Proceedings of the 2010 IEEE Interna-
Weiss, Vincent Dubourg, Jake Vanderplas, Alexan- tional Conference on Data Mining Workshops. Syd-
dre Passos, David Cournapeau, Matthieu Brucher, ney, Australia, pages 1136–1142.
Matthieu Perrot, and Édouard Duchesnay. 2011.
Scikit-learn: Machine learning in Python. Journal Yong Zhang, Xiaomin Ji, and Suhua Zhang. 2016. An
of Machine Learning Research 12(Oct):2825–2830. approach to EEG-based emotion recognition using
combined feature extraction method. Neuroscience
Jeffrey Pennington, Richard Socher, and Christopher Letters 633:152–157.
Manning. 2014. GloVe: Global vectors for word
representation. In Proceedings of the 2014 Con- Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen
ference on Empirical Methods in Natural Language Li, Hongwei Hao, and Bo Xu. 2016. Attention-
Processing (EMNLP 2014). Doha, Qatar, pages based bidirectional long short-term memory net-
1532–1543. works for relation classification. In Proceedings
of the 54th Annual Meeting of the Association for
Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Computational Linguistics (ACL 2016). Berlin, Ger-
Gardner, Christopher Clark, Kenton Lee, and Luke many, pages 207–212.
Zettlemoyer. 2018. Deep contextualized word rep-
resentations. In Proceedings of the 2018 Confer-
ence of the North American Chapter of the Associ-
ation for Computational Linguistics: Human Lan-
guage Technologies (NAACL-HLT 2018). New Or-
leans, Louisiana, USA, pages 2227–2237.
794
Empirical Study of Diachronic Word Embeddings for Scarce Data
795
Proceedings of Recent Advances in Natural Language Processing, pages 795–803,
Varna, Bulgaria, Sep 2–4 2019.
scheme and compare the behaviour of word drifts training such as the two previously presented are
exhibited by the models. Finally, we experiment inefficient in the case of low-frequency words and
regularising the models in order to tackle the faults hypothesise that a new set of methods, pooled
detected in the previous analysis. The experiments under the name of dynamic models, may be more
in section 4 are made on the New York Times An- adapted. These models use probabilistic models
notated Corpus (NYT) 2 (Sandhaus, 2008) cover- to learn time-varying word embeddings while
ing two decades. controlling the drift of the word vectors using a
Gaussian diffusion process. Bamler and Mandt
2 Related Work (2017) uses Bayesian word embeddings, which
The first methods to measure semantic evolution makes the algorithm more robust to noise in
rely on detecting changes in word co-occurrences, the case of sparse data; while Rudolph and Blei
and approaches bases on distributional similarity (2018) relies on a Bernoulli distribution to learn
(Gulordava and Baroni, 2011) . The use of auto- the dynamic embeddings jointly across all time
mated learning methods, based on word embed- slices, making the most of the full dataset.
dings (Mikolov et al., 2013), is recent and has un-
dergone a increase in interest these last two years Outside of the framework of diachrony, several
with the successive publication of three articles attempts aim at improving or adapting word em-
dedicated to a literature review of the domain (Ku- beddings to low-volume corpora in the literature.
tuzov et al., 2018; Tahmasebi et al., 2018; Tang, It can involve morphological information (Luong
2018). In this section, we mainly consider this et al., 2013) derived from the character level (San-
second line of work, along with the peculiarities tos and Zadrozny, 2014; Labeau et al., 2015), and
of scarce data. often make use of external resources: semantic
Kim et al. (2014) developed one of the first lexicon (Faruqui et al., 2015), and pre-trained em-
method to learn time-varying word sparse repre- beddings from larger corpora (Komiya and Shin-
sentations. It consists in learning an embedding nou, 2018). However, to our knowledge, no work
matrix for the first time slice t0 , then updating it has attempted to apply similar solutions to the
at each time step t using the matrix at t − 1 as problem of sparse data in temporal corpora, even
initialisation. This method is called incremental thought this situation has been faced by many au-
updating . Another broadly used method it to thors, often in the case of short time steps for so-
learn an embedding matrix for each time slice cial media data (Stewart et al., 2017; Bamler and
independently; due to the stochastic aspect of Mandt, 2017; Kulkarni et al., 2015).
word embeddings, the vectorial space for each
time slice is different, making them not directly
3 Diachronic Models
comparable. Thus, authors perform an alignment This section briefly describes the three models un-
of the embeddings spaces by optimising a ge- der study: the Skip-Gram incremental updating al-
ometric transformation (Hamilton et al., 2016; gorithm from Kim et al. (2014), the dynamic filter-
Dubossarsky et al., 2017; Szymanski, 2017; ing algorithm of Bamler and Mandt (2017), and
Kulkarni et al., 2015)). the dynamic Bernoulli embeddings model from
Rudolph and Blei (2018). We consider a corpus
In the case of sparse data, in addition to the divided into T time slices indiced by t. For each
approximative aspect of the alignment that harms time step t, every word i is associated with two
the robustness of the embeddings, these methods vectors uit (word vector) and vit (context vector).
are sensitive to random noise, which is difficult to
disambiguate from semantic drifts. Moreover, the 3.1 Incremental Skip-Gram (ISG)
second one require large amounts of data for each
This algorithm relies on the skip-gram model es-
time step to prevent overfitting. Tahmasebi (2018)
timated with negative sampling (SGNS) method
shows that low-frequency words have a much
described in (Mikolov et al., 2013) and it can be
lower temporal stability than high-frequency ones.
summarised as follows. The probability of a word
In (Tahmasebi et al., 2018), the authors explain
i to appear in the context of a word j is defined
that usual methods for diachronic embeddings
by σ(uTi,t vj,t ), with σ being the sigmoid function.
2
https://catalog.ldc.upenn.edu/LDC2008T19 Words i and j are represented by their embedding
796
vectors ui,t and vj,t ) at time t. The matrices Ut and variance D0 is added, resulting in the follow-
and Vt gathers all of them for the whole vocab- ing distributions over the embeddings matrices Ut :
ulary. The context is made of a fixed number of
surrounding words and each word in the context p(U1 |U0 ) ∼ N (0, D0 ) (2)
are considered as independent of each other. p(Ut |Ut−1 ) ∼ N (Ut−1 , D) N (0, D0 ).
The negative sampling strategy associates to
each observed word-context pair (the positive ex- The same equation stands for Vt . Training
amples) n+ − this model requires to estimate the posterior dis-
ijt , a set of negative examples nijt . The
negative examples are sampled for a noise distri- tributions over Ut and Vt given n+− t . This
bution following Mikolov et al. (2013). (Bayesian) inference step is unfortunately un-
Let n+− denote for the time step t, the union tractable. In (Bamler and Mandt, 2017), the au-
t
of positive and negative examples. The objec- thors propose to use variational inference (Jordan
tive function can be defined as the following log- et al., 1999) in its online extension (Blei et al.,
likelihood: 2017). The principle of variational inference is to
approximate the posterior distribution with a sim-
log p(n+− pler variational distribution qλ (U, V ) (λ gathers all
t |Ut , Vt ) = Lpos (Ut , Vt )+Lneg (Ut , Vt )
the parameters of q). This variational posterior
L
X
− will be iteratively updated at each time step. The
= (n+ T T
ijt log σ(ui,t vj,t )+nijt log σ(−ui,t vj,t )) final objective function can be written as follows:
i,j=1
(1)
Lt (λ) = Eqλ [log p(n+−
t |Ut , Vt )] (3)
For the first time slice, the matrices U1 and
V1 are initialised using a Gaussian random noise + Eqλ [log p(Ut , Vt )|n+−
1:t−1 ]
N (0, 1) before being trained according to equa- − Eqλ [log qλ (Ut , Vt )].
tion (1). Then, for each successive time slice,
the embeddings are initialised with values of the This loss function is the sum of three terms: the
previous time slice following the methodology of log-likelihood (computed following equation (1)),
(Kim et al., 2014). This way, the word vectors of the log-prior (which enforces the smooth drift of
each time step are all in the same vectorial space embedding vectors, sharing information with the
and directly comparable. previous time step), and the entropy term (approx-
imated as the sum of the variances of the embed-
3.2 Dynamic Filtering of Skip-Gram (DSG) ding vectors).
This second method relies on the Bayesian ex- 3.3 Dynamic Bernoulli Embeddings (DBE)
tension of the SGNS model described by Barkan
(2015). The main idea is to share information from The DBE models extends the Exponential Family
one time step to another, allowing the embeddings Embeddings (EFE)(Rudolph et al., 2016), a prob-
to drift under the control of a diffusion process. A abilistic generalisation of the Continuous Bag-of-
full description of this approach, denoted as the fil- Words (CBOW) model of Mikolov et al. (2013).
tering model, can be found in (Bamler and Mandt, The main idea is that the model predicts the cen-
2017). tral word vector conditionally to its context vector
In this model, the vectors ui,t and vi,t are con- following a Bernoulli distribution. A detailed de-
sidered as latent vectors. Under a Gaussian as- scription of the model can be found in (Rudolph
sumption, they are represented by their means and Blei, 2018).
(µui,t , µvi,t ) and variances (Σui,t , Σvi,t ). They are Each word i has T different embeddings vec-
initialised for the first time slice with respectively tors uit , but this time, the context vectors vi are
a zero mean vector and a identity variance matrix. assumed to be fixed across the whole corpus. The
The temporal drift from one time step to an- embedding vector uit drifts throughout time fol-
other follows a Gaussian diffusion process with lowing a Gaussian random walk, very similarly to
zero mean and variance D. This variance is called equation (2):
the diffusion constant and has to be tuned along U0 , V ∼ N (0, λ−1 (4)
0 I),
with the other hyperparameters. Moreover, at each −1
time step a second Gaussian prior with zero mean Ut ∼ N (Ut−1 , λ I).
797
The drift hyper-parameter λ controls the tem- a small vocabulary is more adequate for sparse
poral evolution of Ut , and is shared across all time data in a temporal analysis in order to avoid hav-
steps. The training process, described more pre- ing time steps were some word does not appear at
cisely by Rudolph and Blei (2018), relies on a vari- all. The total number of words in the corpus after
ant of the negative sampling strategy described by preprocessing is around 38.5 million. It amounts
Mikolov et al. (2013). The goal is to optimise the to around 200k words per time step in the 10 %
model across all time steps jointly, by summing subset of the corpus, thus only 20k in the 1 % sub-
over t the following loss function: set.
To tune the hyperparameters, we use the log-
Lt = Lpos (Ut , V ) + Lneg (Ut , V ) likelihood of positive examples Lpos measured on
+ Lprior (Ut , V ). (5) the validation set. We train each model for 100
epochs, with a learning rate of 0.1, using the Adam
The two first terms are computed as in equation optimiser. For the DSG model, we use a diffusion
(1). The third term is defined as : constant D = 1 and a prior variance D0 = 0.1 for
both corpora. For the DBE model, we use λ = 1
λ0 X
L XL and λ0 = 0.01.
2 λ0
Lprior (Ut , V ) = − kvi k − kui,0 k2 We choose an embedding dimension d = 100,
2 2
i=1 i=1 as the experiments show that a small embedding
λX dimension, as in (Stewart et al., 2017), leads to
− kui,t − ui,t−1 k2 . (6)
2 smoother word drifts and makes the model less
i,t
sensitive to noise when the data is scarce.
The role of Lprior is twofold: it acts as a regu- We use a context window of 4 words and a neg-
larisation term on V and Ut , and as a constraint on ative ratio of 1; we observed that having a higher
the drift of Ut , preventing it from going to far apart number of negative samples artificially increased
from Ut−1 the held-out likelihood, but equalised the drifts of
all the words in the corpus. Thus, in an extreme
4 Experimental Results
scarcity situation, each negative sample has a high
The goal of this study is to compare the behaviour weight during training: the number of negative
of the three algorithms described in section 3 in samples has to be very carefully selected depend-
case of low-volume corpora. We evaluate their ing on the amount of data.
predictive power on different volumes of data to
compare the impact of two initialisation methods, 4.2 Impact of Initialisation on Sparse Data
and analyse the behaviour of the drift of the em- The embedding vectors of the ISG and DBE mod-
beddings. els are initialised using a Gaussian white noise,
4.1 Experimental Setup while the means and variances of DSG are ini-
tialised with null vectors and identity matrices
We use the New York Times Annotated Cor- respectively. However, a good initialisation can
pus (NYT) (Sandhaus, 2008) 3 containing around greatly improve the quality of embeddings, partic-
1 855 000 articles ranging from January 1st 1987 ularly in the case of scarce data.
to June 19th 2007. We divide the corpus into We experiment the impact of two types of
T = 20 yearly time steps (the incomplete last year initialisation on the log-likelihood of positive
is not used in the analysis) and held out 10 % of examples on the test set.
each time step for validation and testing. Then,
we sample several subsets of the corpus : 50 %,
Internal initialisation:
10%, 5% and 1% of the training set. This way, we
We train each model in a static way on the
can compare the models on each subset to eval-
full dataset. Then, we use the resulting vectors
uate their ability to train a model in the case of
as initialisation for the first time step of the
low-volume corpora.
diachronic models. This methods is especially
We remove stopwords and choose a vocabu-
suitable for domain-specific corpora where no
lary of V = 10k most frequent words. Indeed,
external comparable data is available.
3
released by the Linguistic Data Consortium
798
Figure 1: Log-likelihoods for the DSG model on three subsets of the corpus, comparing the baseline (random
initialisation) with the two initialisation methods: internal is the initialisation from the full dataset while external-
backward is the initialisation with the Wikipedia vectors, with training from most recent to oldest time step.
799
Figure 2: Histogram of word drift for each model on two subsets of the NYT corpus. The drifts are computed
from t0 = 1987 to each successive time step, and superposed on the histogram. The lightest colours indicate drifts
calculated until the most recent time steps. The number of words are on logarithmic scale.
haviour (no colour gradient), while the two other subset than on the 10 % subset: they are not able
models do. This is justified by the use of the dif- to discriminate very high drifts from the rest of the
fusion process to link the time steps in equations 2 words in extreme scarcity situation.
and 5: it allows the DSG and DBE models to em-
phasise the directed fashion of drifts even in the 4.4 Regularisation Attempt
situation of scarce data. To tackle the weakness of the DBE and DSG mod-
The second property to highlight is the capac- els on the smallest subset, we attempt to regularise
ity of the models to discriminate words that drift their loss in order to control the weight of the high-
from words that stay stable. From the human point est and lowest drifts. Our goal is to allow the
of view, a majority of words has a stable mean- model to:
ing (Gulordava and Baroni, 2011); especially on a • better discriminate very high drifts;
dataset covering only two decades like the NYT.
The DBE model has a regularisation term (equa- • be less sensitive to noise, giving lower weight
tion 6) to enforce this property, and a majority to very low embedding drifts.
of words have a very low drift on the histogram. We test several possible regularisation terms to
However, on 1 % of the dataset, this model can not be added to the loss. The best result is obtained
discriminate very high drifts from the rest. The with the Hardshrink activation function, which is
ISG and DSG models have a different distribu- defined this way :
tion shape, with the peak having a drift superior
to zero. The only words that do not drift on their HardShrink(x) = x, if x > β (8)
histograms are the one that are absent from a time = −x, if x < −β
step. = 0, otherwise
To conclude, both the DBE and DSG model are
For the DSG and DBE models, we add to the loss
able to detect directed drifts even in the 1 % subset
the following regularisation term, amounting to a
of the NYT corpus, while the ISG can not. How-
thresholding function applied to the drift:
ever, the drift distributions of the DBE and DSG
models have a much shorter shorter tail on the 1 % regβ = α ∗ HardShrink(drift(Ut ), β) (9)
800
Where α is the regularisation constant to be tuned, We proposed two initialisation schemes: the
β is the threshold of the hardshrink function, and internal initialisation, more suited for low volume
the drift is computed according to equation 7. The of data, and the backward external initialisation,
regularisation term is minimised. The activation more suited for higher volumes and long periods
function acts as a threshold to limit the amount of of temporal study. Then, we compared the distri-
words having an important drift. We choose β as butions of the drifts of the models. We conclude
the mean drift for both models. that even in extreme scarcity situations, the DBE
and DSG models can highlight directed drifts
For both DSG and DBE, the right tail of the dis- while the ISG model is too sensitive to noise.
tribution of the drifts with regularisation (Figure Moreover, the DBE model is the best at keeping
3) is much longer than in the original model (Fig- a majority of the words stable. This property,
ure 2). Moreover, in the case of the DSG model, as long as the ability to detect directed drift, are
more words have a drift very close to zero. two important properties of a diachronic model.
To conclude, the regularised DSG model con- However, both have low ability to discriminate
siders more words as temporally stable. Further- the highest drifts on a very small dataset. Thus,
more, regularising the loss of the dynamic models we added a regularisation term to their loss using
allows to better discriminate extreme word embed- the Hardshrink activation function, successfully
ding drifts for very small corpora. getting longer distribution tails for the drifts.
801
David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe. 24th International Conference on World Wide Web,
2017. Variational inference: A review for statisti- WWW ’15, pages 625–635, Republic and Canton of
cians. CoRR, abs/1601.00670. Geneva, Switzerland. International World Wide Web
Conferences Steering Committee.
Arthur Brazinskas, Serhii Havrylov, and Ivan Titov.
2018. Embedding words as distributions with a Andrey Kutuzov, Lilja Øvrelid, Terrence Szymanski,
bayesian skip-gram model. In Proceedings of the and Erik Velldal. 2018. Diachronic word embed-
27th International Conference on Computational dings and semantic shifts: a survey. In Proceed-
Linguistics, pages 1775–1789. Association for Com- ings of the 27th International Conference on Com-
putational Linguistics. putational Linguistics, pages 1384–1397. Associa-
tion for Computational Linguistics.
Haim Dubossarsky, Daphna Weinshall, and Eitan
Grossman. 2017. Outta control: Laws of seman- Matthieu Labeau, Kevin Löser, and Alexandre Al-
tic change and inherent biases in word representa- lauzen. 2015. Non-lexical neural architecture for
tion models. In Proceedings of the 2017 Confer- fine-grained pos tagging. pages 232–237, Lisbon,
ence on Empirical Methods in Natural Language Portugal. Association for Computational Linguis-
Processing, pages 1136–1145. Association for Com- tics.
putational Linguistics.
Bofang Li, Tao Liu, Zhe Zhao, Buzhou Tang, Alek-
Manaal Faruqui, Jesse Dodge, Sujay Kumar Jauhar, sandr Drozd, Anna Rogers, and Xiaoyong Du. 2017.
Chris Dyer, Eduard Hovy, and Noah A. Smith. 2015. Investigating Different Syntactic Context Types and
Retrofitting word vectors to semantic lexicons. In Context Representations for Learning Word Embed-
Proceedings of the 2015 Conference of the North dings. In Proceedings of the 2017 Conference on
American Chapter of the Association for Compu- Empirical Methods in Natural Language Process-
tational Linguistics: Human Language Technolo- ing, pages 2411–2421.
gies, pages 1606–1615. Association for Computa-
tional Linguistics. Thang Luong, Richard Socher, and Christopher Man-
ning. 2013. Better word representations with recur-
Kristina Gulordava and Marco Baroni. 2011. A distri- sive neural networks for morphology. In Proceed-
butional similarity approach to the detection of se- ings of the Seventeenth Conference on Computa-
mantic change in the google books ngram corpus. In tional Natural Language Learning, pages 104–113.
Proceedings of the GEMS 2011 Workshop on GE-
ometrical Models of Natural Language Semantics, Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor-
pages 67–71. Association for Computational Lin- rado, and Jeff Dean. 2013. Distributed representa-
guistics. tions of words and phrases and their composition-
ality. In C. J. C. Burges, L. Bottou, M. Welling,
William L. Hamilton, Jure Leskovec, and Dan Jurafsky. Z. Ghahramani, and K. Q. Weinberger, editors, Ad-
2016. Diachronic word embeddings reveal statisti- vances in Neural Information Processing Systems
cal laws of semantic change. In Proceedings of the 26, pages 3111–3119. Curran Associates, Inc.
54th Annual Meeting of the Association for Compu-
tational Linguistics (Volume 1: Long Papers), pages Maja Rudolph and David Blei. 2018. Dynamic embed-
1489–1501. Association for Computational Linguis- dings for language evolution. In Proceedings of the
tics. 2018 World Wide Web Conference on World Wide
Web, pages 1003–1011. International World Wide
Michael I. Jordan, Zoubin Ghahramani, and et al. 1999. Web Conferences Steering Committee.
An introduction to variational methods for graphical
models. In MACHINE LEARNING, pages 183–233. Maja Rudolph, Francisco Ruiz, Stephan Mandt, and
MIT Press. David Blei. 2016. Exponential family embeddings.
In Advances in Neural Information Processing Sys-
Yoon Kim, Yi-I Chiu, Kentaro Hanaki, Darshan Hegde, tems, pages 478–486.
and Slav Petrov. 2014. Temporal analysis of lan-
guage through neural language models. In Proceed- Evan Sandhaus. 2008. The new york times annotated
ings of the ACL 2014 Workshop on Language Tech- corpus. In Philadelphia : Linguistic Data Consor-
nologies and Computational Social Science, pages tium. Vol. 6, No. 12.
61–65. Association for Computational Linguistics.
Cicero D. Santos and Bianca Zadrozny. 2014. Learning
Kanako Komiya and Hiroyuki Shinnou. 2018. Inves- character-level representations for part-of-speech
tigating effective parameters for fine-tuning of word tagging. In Proceedings of the 31st International
embeddings using only a small corpus. In Proceed- Conference on Machine Learning (ICML-14), pages
ings of the Workshop on Deep Learning Approaches 1818–1826. JMLR Workshop and Conference Pro-
for Low-Resource NLP, pages 60–67, Melbourne. ceedings.
Association for Computational Linguistics.
Ian Stewart, Dustin Arendt, Eric Bell, and Svitlana
Vivek Kulkarni, Rami Al-Rfou, Bryan Perozzi, and Volkova. 2017. Measuring, predicting and visual-
Steven Skiena. 2015. Statistically significant de- izing short-term change in word representation and
tection of linguistic change. In Proceedings of the usage in vkontakte social network. In ICWSM.
802
Terrence Szymanski. 2017. Temporal word analogies:
Identifying lexical replacement with diachronic
word embeddings. In Proceedings of the 55th An-
nual Meeting of the Association for Computational
Linguistics (Volume 2: Short Papers), pages 448–
453. Association for Computational Linguistics.
Nina Tahmasebi. 2018. A study on word2vec on a his-
torical swedish newspaper corpus. In DHN.
Nina Tahmasebi, Lars Borin, and Adam Jatowt. 2018.
Survey of computational approaches to diachronic
conceptual change. CoRR, 1811.06278.
Xuri Tang. 2018. A state-of-the-art of semantic
change computation. Natural Language Engineer-
ing, 24(5):649–676.
803
A Fast and Accurate Partially Deterministic Morphological Analysis
804
Proceedings of Recent Advances in Natural Language Processing, pages 804–809,
Varna, Bulgaria, Sep 2–4 2019.
This paper proposes a partially deterministic We describe CISs with the following three ex-
morphological analysis method for improved pro- amples.
cessing speed. We improve processing speed (A) 床の間
by incorporating maximum matching (the first (alcove, or gap of floor )
group) to a bi-gram-based morphological engine (B) 南京
(the third group). (Nanjing)
In order to alleviate the performance loss of (C) 南 京都 病院
maximum matching, we propose a concept of (South) (Kyoto) (hopital)
Context Independent Strings (CISs), which are
(A) is a string that does not satisfy the first con-
strings having no ambiguity in terms of morpho-
dition of CIS. “床の間” has two grammatical anal-
logical analysis. We also propose an algorithm for
ysis result: “床の間 (alcove)” with one word or “
the building of the CISs dictionary from the large
床 (floor) + の (to) + 間 (gap)” with three words.
amount of automatically analysed texts. The dic-
The analysis depends on the context where the
tionary maps CISs to the results of morphological
string occurred.
analysis (sequence of words and POS tags).
In (B), the string “南京” has only one grammat-
Our method first applies maximum matching to
ical analysis result “南京” (Nanjin). However, the
a sentence using the CIS dictionary. CISs in a
string is a part of (C) “南京都病院”, and the string
sentence are deterministically analyzed by the dic-
corresponts “南” (South) + the first character of “
tionary. Then the method analyze the rest part of
京都” (Kyoto). That violates the second condition
the sentence using a bi-gram-based morphological
of CIS.
analysis. We can use high-performance and com-
When we segment the example (C) “南京都
putationally heavy methods such as neural net-
病院” (South Kyoto Hospital) using word-level
work based methods for the building of the dic-
maximum matching, the example is segmented
tionary. That is, we can partially use their analysis
into “南京+都+病院” (Nanjin metropolitan hospi-
results without recomputing through the CIS dic-
tal) instead of correct segmentation “南+京都+病
tionary.
院” (South Kyoto hospital). This is because the
We obtain CISs from automatically parsed longest word matched from the beginning of the
large texts with a state-of-the-art morpholog- example (C) is CDS (“南京”).
ical analyzer JUMAN++ (Morita et al., 2015; In fact, whole (C) is an example of a CIS. There
Tolmachev et al., 2018). We evaluate this method is no “南京+都+病院” (Nanjin metropolitan hos-
on a Japanese morphological analysis task. The pital)“, and then the string has only one grammat-
experimental results show a 30% reduction of run- ical analysis result.
ning time while maintaining improved accuracy.
3 Proposed Method
2 Context Independent Strings
This section describes the building method of a
The key idea of using maximum matching without CIS dictionary, and deterministic morphological
performance loss is to separate character strings analysis using the CIS dictionary.
into two classes: context independent string (CIS)
and context dependent string (CDS). 3.1 Building a CISs Dictionary
A CIS is a string that has a one-to-one cor- A CIS dictionary is a data structure where each
respondence with its grammatical analysis result, CIS is associated to its morphological analysis re-
while a CDS has two or more grammatical analy- sult.
sis results that depend on its contexts. A CIS has We built the CIS dictionary using the algorithm
to satisfy two conditions: showed in Algorithm 1. In this method, the dictio-
nary is built from a large automatically analyzed
• The string does not have two or more gram- corpus C. We collected word N-grams nj , their
matical analysis results. surfaces surface(nj ) and pointers j of the sen-
tences in which the N-grams occurred, as a set of
• The analysis result is not affected by any candidates (H, N, S).
strings adjacent to the beginning and end of In STEP (1) at the line 14 of Algorithm 1, when
the strings. an identical surface was associated with two or
805
CIS CIS
N N N N 京都 N N P N P V
国立 病院 機構 南 病院 まで バス で 向 か う
National hospital organization south Kyoto to bus by go
(location) hospital
N N N P
病 院 ま で
N N disease true by
国立 N
病 院 京 都 institution
で
N
向
Adj
かう(寡雨)
N
Kunitachi disease institution capital metropolis N N hand opposit dry
(location) ま で
N V V V
N N 南京 duration hand で 向 かう(買う)
国 立 Nanjing go out face buy
Ritsu
Nation (person) (location)
N P N
で 向(出向き) か う
visiting or rain
V N
で 向 か う
visit cormorant
V
で 向 か う
resist
Figure 1: Example of a lattice that is built by our baseline method and our proposed model using a
CIS dictionary. The CIS dictionary contains “国立病院機構南京都” and “バスで向かう”, so the corre-
sponding results are loaded from the dictionary. The word candidates surrounded by dotted line are not
generated in our proposed model. Our implementation builds a CIS dictionary from word sequences of
length 5 or less. Thus, even though “国立病院機構南京都病院” (6 words) is clearly a CIS, it is not in
the dictionary.
more N-grams, or the N-grams occurred less fre- isSensitive function in Algorithm 1.
quently than a predetermined threshold T , the sur- Finally, we build a CIS dictionary D using the
faces are discarded from the candidates. We aim to collected CISs and their analysis result.
discard ambiguous strings similar to example (A).
In STEP (2) at the line 15, we remove CDSs that 3.2 Analysis using a CIS Dictionary
violate the second condition of CIS. As we can see There are three steps to analyse a sentence using a
from Example (B), the surface of CDS appears in CIS dictionary: (1) Deterministic analysis by max-
a part of other word N-gram. The findString gives imum matching with a CIS dictionary, (2) Build-
a set of sentences where the string p occurs in cor- ing lattice by lookup dictionary and (3) Search for
pus C. The step checks there is no occurrence of the least cost path on the lattice. We show our al-
the string p that appears in a part of another N- gorithm in Algorithm 2.
gram. We discarded any surface that occurred in
First, LookUpLongestCIS in Algorithm 2 finds
a sentence in which the corresponding N-gram did
CISs with the maximum match of a CIS dictionary
not occur. This is because the occurrence indicates
from each character of the input sentence. If a CIS
that the surface was a part of another different N-
is found, the algorithm skips the range of matched
gram.
string, and calls LookUpLongestCIS again. If it
In STEP (3) at the line 15, the step is filter- is not found, the step is restarted from the next
ing using a heuristic. If the corpus is sufficiently character. The algorithm calls LookUpDictionary
large, the process considers any string before and and builds a word lattice for each span where any
after the N-gram. However, it is infeasible to cover analysis result is not obtained with the maximum
all contexts of all expressions in a corpus of lim- matching.
ited size. For the purpose of augmenting the cor- Finally, whole sentence is connected as one lat-
pus, we discarded the N-grams where words at tice, and a word sequence with the lowest cost
the beginning or ending of N-gram are sensitive is obtained by Viterbi algorithm. On this lattice,
to preceding and succeeding strings. Specifically, the spans deterministically analyzed are expressed
if the beginning or end of the N-gram is a one- as nodes corresponding to the analysis result ob-
character word and is not in a predefined white tained from the dictionary, and their costs are 0.
list (punctuations, and a part of particles), the N- Other costs are calculated by a bi-gram model.
gram is discarded by this step. We assume that
Figure 1 shows an example of a lattice that is
one-character words are often a suffix or prefix
built by our proposed method.
of other words and these words are sensitive to
the surrounding contexts. We show the process as
806
Algorithm 1 An algorithm for building a CIS dic- Algorithm 2 Morphological analysis using a CIS
tionary dictionary and Regular Expressions
1: T is a threshold for N-gram frequency, 1: S ← {c0 ,. . . ,cn }, i ← 0, j ← 0, k ← 0
2: C ← {s0 , . . . , sM }, 2: L = A lattice structure
3: S ← {} 3: while i ≤ N do
4: H, R, D are associative arrays. H maps sur- 4: result, length ← LookUpLongestCIS(S, i)
face to array of sentence ids, R maps surface
to array of analysis result, D maps surface to 5: if result != ϕ then
an analysis result 6: L[i] ← result
5: for i = 0 to N do 7: i ← i + length
6: for all nj ∈ si do 8: else
7: insert(H[surface(nj )], i) 9: i←i+1
8: insert(R[surface(nj )], nj ) 10: end if
9: insert(S, surface(nj )) 11: end while
10: end for 12: for j = 0 to N –(2) do
11: end for 13: if j is processed by LookUpLongestCIS
12: for all p ∈ S do then
13: # STEP (1) 14: continue
next if |R[p]| > 1 OR |H[p]| < T 15: end if
14: # STEP (2) 16: L[j] ← LookUpDictionary(S,j)
next if H[p] ̸= findString(p, C) 17: end for
15: # STEP (3) 18: return Viterbi(L) –(3)
next if isSensitive(R[p][0])
16: D[p] ← (R[p][0])
17: end for by KNB corpus (Hashimoto et al., 2011) that con-
18: return D sists of 249 articles (4,186 sentences). The corpus
contains manually annotatted word segmentation,
POS tags and dependencies. Then, we evaluated
4 Experiments the efficiency of our models by measuring run-
4.1 Data and Models ning times of analysis on news articles of Yomi-
uri Shimbun in 2013. The text consists of ap-
We build a CIS dictionary using Mainichi News proximately 300 thousand articles (approximately
articles published from 1991 to 2010. The cor- 4 million sentences).
pus contains approximately 2 million articles. We measured the performance of the meth-
We analyzed the corpus using JUMAN++ v1.02 ods by two performance measures Word and
(Morita et al., 2015). We set threshold T to 4, and Word+POS. Word is the F-value of word segmen-
limited the maximum N-gram length to 5. tation, and Word+POS is the F-value of joint eval-
A morphological dictionary used in our models uation of word segmentation and POS tagging.
is stemmed from JUMAN++ v1.02. We trained
our feature parameters of CRF using L-BFGS 4.2 Performance Evaluation
(Kudo et al., 2004; Liu and Nocedal, 1989). We
trained our models using Kyoto University Web We compared our proposed model with the base-
Document Leads Corpus (Hangyo et al., 2012; line and JUMAN++. Partially deterministic analy-
Kawahara et al., 2014) that contains approxi- sis using a CIS dictionary did not lose to their per-
mately 15,000 manually annotated sentences. The formance but slightly improved in both Word and
corpus contains manually annotated word seg- Word+POS. Our baseline is almost a reimplemen-
mentations, POS tags, dependencies, predicate- tation of MeCab (Kudo et al., 2004) and performs
argument structures including zero anaphora, and almost equally. JUMAN++ is an upper bound of
so on. We used sentences with id (w201106- our proposed model because a deterministically
00000–w201106-00023) for training and sen- analyzed part of the analysis result is almost equiv-
tences with id (w201106-00024) for development. alent with the result of JUMAN++.
We evaluated the performance of our methods
807
Bali,Indonesia, pages 535–544.
Table 1: Performance comparison of various sys-
tems. Chikara Hashimoto, Sadao Kurohashi, Keiji Shinzato,
Method Word Word+POS and Nagata Masaaki. 2011. Construction of a blog
(F-val) (F-val) corpus with syntactic, anaphoric, and sentiment an-
Baseline 95.94 95.07 notations (in japanese). Journal of Natural Lan-
Partially deterministic guage Processing 18(2):175–201.
95.99 95.17
(Proposed)
JUMAN++ 96.84 96.15 Nobuhiro Kaji and Masaru Kitsuregawa. 2013. Effi-
cient word lattice generation for joint word segmen-
Table 2: Comparing running times of various sys- tation and pos tagging in japanese. In Proceedings
tems. of the Sixth International Joint Conference on Natu-
Method Time (seconds) ral Language Processing. pages 153–161.
Baseline 315.3
Partially deterministic Daisuke Kawahara, Yuichiro Machida, Tomohide Shi-
223.4 bata, Sadao Kurohashi, Hayato Kobayashi, and
(Proposed)
MeCab 124.5 Manabu Sassano. 2014. Rapid development of a
JUMAN++ 341705.7 corpus with discourse annotations using two-stage
crowdsourcing. In Proceedings of COLING 2014,
the 25th International Conference on Computational
4.3 Running Time Comparison Linguistics: Technical Papers. Dublin City Univer-
We compared our systems with publicly available sity and Association for Computational Linguistics,
Dublin, Ireland, pages 269–278.
implementations. MeCab is the fastest, but our
implementation (Baseline) marks similar running Taku Kudo, Kaoru Yamamoto, and Yuji Matsumoto.
time. Since our baseline model is not largely dif- 2004. Applying conditional random fields to
ferent from MeCab, the difference of running time japanese morphological analysis. In Dekang Lin
and Dekai Wu, editors, Proceedings of EMNLP
is due to the implementation. Partially determin- 2004. Association for Computational Linguistics,
istic analysis reduces running time by 30%. If Barcelona, Spain, pages 230–237.
we implement our proposed method on MeCab or
JUMAN++, we suspect it will improve their effi- Shuhei Kurita, Daisuke Kawahara, and Sadao Kuro-
hashi. 2017. Neural joint model for transition-based
ciency. chinese syntactic analysis. In Proceedings of the
55th Annual Meeting of the Association for Com-
5 Conclusion putational Linguistics (ACL2017). Association for
Computational Linguistics.
We presented a new fast partially deterministic
morphological analysis method. We introduced Sadao Kurohashi and Daisuke Kawa-
the concept of Context Independent Strings and hara. 2014. Juman ver.7.01.
http://nlp.ist.i.kyoto-u.ac.jp/EN/index.php?JUMAN.
presented an algorithm for building a dictionary of
CISs from a large corpus. We checked that the pro- D. C. Liu and J. Nocedal. 1989. On the limited mem-
posed method improved both efficiency and per- ory bfgs method for large scale optimization. Math.
formance of morphological analysis. The method Program. 45(3):503–528.
reduced the running time by 30% and improved Hajime Morita, Daisuke Kawahara, and Sadao Kuro-
the performance 0.1 pt in Word+POS. hashi. 2015. Morphological analysis for unseg-
mented languages using recurrent neural network
language model. In Proceedings of the 2015 Con-
References ference on Empirical Methods in Natural Language
Processing. pages 2292–2297.
Xinchi Chen, Xipeng Qiu, Chenxi Zhu, Pengfei Liu,
and Xuanjing Huang. 2015. Long short-term mem- Masaaki Nagata. 1997. A self-organizing japanese
ory neural networks for chinese word segmentation. word segmenter using heuristic word identification
In Proceedings of the 2015 Conference on Empiri- and re-estimation. In Proceedings of the 5th Work-
cal Methods in Natural Language Processing. pages shop on Very Large Corpora. pages 203–215.
1197–1206.
Graham Neubig, Yosuke Nakata, and Shinsuke Mori.
Masatsugu Hangyo, Daisuke Kawahara, and Sadao 2011. Pointwise prediction for robust, adaptable
Kurohashi. 2012. Building a diverse document japanese morphological analysis. In Proceedings of
leads corpus annotated with semantic relations. In the 49th Annual Meeting of the Association for Com-
Proceedings of the 26th Pacific Asia Conference putational Linguistics: Human Language Technolo-
on Language, Information, and Computation. Fac- gies: Short Papers - Volume 2. HLT ’11, pages 529–
ulty of Computer Science, Universitas Indonesia, 533.
808
Manabu Sassano. 2014. Deterministic word segmen-
tation using maximum matching with fully lexical-
ized rules. In Proceedings of the 14th Conference of
the European Chapter of the Association for Compu-
tational Linguistics, volume 2: Short Papers. pages
79–83.
Richard Sproat, William Gale, Chilin Shih, and Nancy
Chang. 1996. A stochastic finite-state word-
segmentation algorithm for chinese. Comput. Lin-
guist. 22(3):377–404.
Arseny Tolmachev, Daisuke Kawa-
hara, and Sadao Kurohashi. 2018.
Juman++: A morphological analysis toolkit for scriptio continua.
In Proceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing: System
Demonstrations. Association for Computational
Linguistics, Brussels, Belgium, pages 54–59.
https://doi.org/10.18653/v1/D18-2010.
Heng-Jun Wang, Nian-Wen Si, and Cheng Chen. 2016.
An effective joint model for chinese word segmen-
tation and pos tagging. In Proceedings of the 2016
International Conference on Intelligent Information
Processing. ICIIP ’16, pages 35:1–35:6.
Xiaoqing Zheng, Hanyang Chen, and Tianyu Xu. 2013.
Deep learning for Chinese word segmentation and
POS tagging. In Proceedings of the 2013 Confer-
ence on Empirical Methods in Natural Language
Processing. pages 647–657.
809
incom.py – A Toolbox for Calculating Linguistic Distances and
Asymmetries between Related Languages
Marius Mosbach† , Irina Stenger, Tania Avgustinova, Dietrich Klakow†
Collaborative Research Center (SFB) 1102: Information Density and Linguistic Encoding
† Spoken Language Systems, Saarland Informatics Campus
Saarland University, Germany
ira.stenger@mx.uni-saarland.de
avgustinova@coli.uni-saarland.de
{mmosbach, dietrich.klakow}@lsv.uni-saarland.de
1 Introduction
1.1 Related Work Similarly, the research of Jágrová et al. (2016)
Linguistic phenomena may be language specific shows that Czech and Polish, both West Slavic, us-
or shared between two or more languages. With ing the Latin script, are orthographically more dis-
regard to cross-lingual intelligibility, various con- tant from each other than Bulgarian and Russian,
stellations are possible. For example, speakers South and East Slavic respectively using the Cyril-
of language A may understand language B bet- lic script. Both language pairs have similar lexical
ter than language C, i.e. [A(B) > A(C)] while distances, however, the asymmetric conditional
speakers of language B may understand language entropy based measures suggest that Czech read-
C better than language A, i.e. [B(C) > B(A)]. ers should have more difficulties reading Polish
For instance, Ringbom (2007) distinguishes be- text than vice versa. The asymmetry between Bul-
tween objective (established as symmetrical) and garian and Russian is very small with a predicted
perceived (not necessarily symmetrical) cross- minimal advantage for Russian readers (Stenger
linguistic similarities. Asymmetric intelligibility et al., 2017b). Additionally Stenger et al. (2017a)
can be of linguistic nature. This may happen if found that word-length normalized adaptation sur-
language A has more complicated rules and/or ir- prisal appears to be a better predictor than aggre-
regular developments than language B, which re- gated Levenshtein distance when the same stimuli
sults in structural asymmetry (Berruto, 2004). sets in different language pairs are compared.
810
Proceedings of Recent Advances in Natural Language Processing, pages 810–818,
Varna, Bulgaria, Sep 2–4 2019.
1.2 This paper assume the characters ci ∈ w are ordered with c0
To calculate linguistic distances and asymmetries, being the first and c|w−1| being the last character of
perform statistical analyses and visualize the ob- word w, where the length |w| of a word w is given
tained results we developed the linguistic tool- by the number of characters it contains, including
box incom.py. The toolbox is validated on the duplicates. Given two words wi , wj , the align-
Russian-Bulgarian language pair. We focus on ment of wi and wj results in two new words w i , w
j
word-based methods in which segments are com- where |wi | = |w j |. We say character sk ∈ w i is
pared at the orthographic level, since orthography aligned to character tl ∈ w j if k = l. That is, they
is a linguistic determinant of mutual intelligibility occur at the same position.
which may facilitate or impede reading intercom- 2.1.2 Levenshtein distance
prehension. We make the following contributions.
Levenshtein distance (LD) (Levenshtein, 1966) is,
1. We provide implementations of various met- it its basic implementation, a symmetric similar-
rics for computing linguistic distances and ity measure between two strings – in our case
asymmetries between languages. words – wi ∈ L1 and wj ∈ L2 . Leven-
shtein distance quantifies the number of opera-
2. We demonstrate the use of incom.py in tions one has to perform in order to transform
an intercomprehension experiment for the wi into wj . Levenshtein distance allows to mea-
Russian-Bulgarian language pair. sure the orthographic distance between two words
and has been successfully used in previous works
3. We show how incom.py can be used to for measuring the linguistic distance between di-
validate word adaptation surprisal and con- alects (Heeringa et al., 2006) as well as the pho-
ditional entropy as predictors for intercom- netic distance between Scandinavian language va-
prehension and discuss benefits over Leven- rieties (Gooskens, 2007). When computing Lev-
shtein distance. enshtein distance between two words LD(wi , wj ),
three different character transformations are con-
The remainder of this paper is structured as fol-
sidered: character deletion, character insertion,
lows. The considered distance metrics imple-
and character substitution. In the following we use
mented in the incom.py toolbox are introduced
T = {insert, delete, substitute} to denote the
in Section 2. Section 3 describes the linguistic data
set of possible transformations. A cost c(t) is as-
used in the experiments. Section 4 presents the
signed to each transformation t ∈ T and setting
evaluation results of the statistical measures and
c(t) = 1 ∀t ∈ T results in the most simple imple-
compares them with the intelligibility scores ob-
mentation.
tained in a web-based cognates guessing. Finally,
in Section 5, conclusions are drawn and future de- incom.py allows computing LD(wi , wj )
velopments outlined. based on a user-defined cost matrix M, which
contains the complete alphabets A(L1 ), A(L2 ) of
2 Linguistic Distances and Asymmetries two languages L1 , L2 as rows and columns, re-
spectively, as well as the costs for every possi-
2.1 Distance measures ble character substitution. That is, for two char-
We start with the introduction of basic notations acters s ∈ A(L1 ) and t ∈ A(L2 ), M(s, t) is
and present the implemented distance measures. the cost of substituting s by t. This user de-
fined cost matrix allows computing linguistically
2.1.1 Notation motivated alignments by incorporating a linguis-
Let L denote a language such as Russian or Bul- tic prior into the computation of the Levenshtein
garian. Each language L has an associated alpha- distance. For example, we assign a cost costs of
bet – a set of characters – A(L) which includes 0 when mapping a character to itself. In case of
the special symbol ∅1 . We use w ∈ L to denote a M being symmetric, the Levenshtein distance re-
word in language L and ci ∈ w to denote the i-th mains symmetric. Along with the edit distance be-
character in word w. Note that while L is a set, tween the two words wi and wj our implementa-
w is not and may contain duplicates. Further, we tion of the Levenshtein distance returns the align-
1
∅ plays an important role when computing alignments. ments w j of wi and wj , respectively. Given
i , w
We will also refer to it as nothing the length K = |w i | of the alignment, we are fur-
811
ther able to compute the normalized Levenshtein
LD(wi ,wj )
distance nLD(wi , wj ) = . For comput-
K count(L1 = s ∧ L2 = t)
ing both the alignment and the resulting edit dis- P̂ (L2 = t|L1 = s) =
count(L1 = s)
tance incom.py uses the Needleman-Wunsch al-
(5)
gorithm (Needleman and Wunsch, 1970) follow-
ing a dynamic-programming approach. ≈ P (L2 = t|L1 = s) (6)
2.1.3 Word adaptation surprisal Certainly the quality of the estimate P̂ (L2 =
Given two aligned words w i and w
j , we can com- t|L1 = s) depends on the size of the corpus C.
pute the Word Adaptation Surprisal (WAS) be- In addition to the corpus based estimated charac-
tween w i and w
j . Intuitively, word adaptation sur- ter surprisals, incom.py provides functionality
prisal measures how confused a reader is when to modify the computed CAS values in a manual
trying to translate wi to wj character by charac- post-processing step. Based on this, the modified
ter. In order to define WAS formally, we intro- word adaptation surprisal can be computed as:
duce the notation of Character Adaptation Sur-
prisal (CAS). Given a character s ∈ A(L1 ) and
K−1
another character t ∈ A(L2 ), the character adap-
i , w
mWAS(w j ) = mCAS(sk , tk ) (7)
tation surprisal between s and t is defined as fol-
k=0
lows:
where mCAS denotes the modified character
adaptation surprisal. Similar to using a user de-
CAS(s, t) = − log2 (P (t | s)) (1) fined cost matrix M when computing the Lev-
enshtein distance, using modified character sur-
Now, the word adaptation surprisal between w
i ∈ prisal allows to incorporate linguistic priors into
L1 and w j ∈ L2 can be computed straightfor- the computation of word adaptation surprisal.
wardly by summing over all characters of the
aligned word pair, i.e. 2.1.4 Conditional entropy
Another asymmetric measure that is supported by
K−1
our incom.py toolbox is Conditional Entropy
i , w
WAS(w j ) = CAS(sk , tk ) (2) (CE) (Shannon, 1948). Formally, the entropy of
k=0
a discrete random variable X is defined as the
weighted average of the surprisal values of this
where K = |w i | = |wj |. Similarly, the normal- distribution. As discussed above, we can obtain
ized word adaptation surprisal is computed as the character surprisals based on the alignments
obtained when computing the Levenshtein dis-
tance. Using these surprisal values we can com-
K−1
1 pute the entropy of a language L as
i , w
nWAS(w j ) = CAS(sk , tk ) (3)
K
k=0
1
= i , w
WAS(w j ) (4) H(L) = − P (L = c) log2 P (L = c) (8)
K
c∈L
Note that in contrast to Levenshtein distance, word
adaptation surprisal is not symmetric. In this work we are interested the entropy of a lan-
Computing CAS (and hence also WAS) de- guage L1 , e.g. Russian, that we compare to the
pends on the conditional probability P (t|s), which entropy of another language L2 , e.g. Bulgarian.
is usually unknown. incom.py estimates P (t|s) Thus we compute the conditional entropy between
by P̂ (t|s) which is based on corpus statistics. two languages L1 and L2 .
Given the alignments of a corpus C of word pairs
produced by the Levenshtein algorithm, we com-
pute P (t|s) by counting the number of times t is CE(L1 |L2 ) = − P (c2 )H(L1 |L2 = c2 )
aligned with s and divide over the total number of c2 ∈L2
812
.xls files .ipynb files .ipynb files
Intuitively, CE(L1 |L2 ) measures the difficulty for ple, BG–RU ние–мы (nie–my) ‘we’ was removed
a L1 reader reading language L2 . Note that similar and the BG вр (zvjar) ‘beast’ instead of и-
to the word adaptation surprisal, both entropy and вотно (životno) ‘animal’ was added to its RU
conditional entropy highly depend on the num- cognate вер (zver’) ‘animal, beast’. In a sec-
ber of available word pairs and only serve as an ond step, a cross-linguistic rule set was designed
approximation to the true (unknown) entropy and taking into account diachronically motivated or-
conditional entropy, respectively. thographic correspondences, e.g. BG–RU: б:бл,
:д, :е, ла:оло etc. Following (Fischer et al.,
2.2 incom.py toolbox
2015) we apply the rule set to the parallel word
A high-level overview of the imcom.py toolbox lists in a computational transformation experiment
is shown in Figure 1. The toolbox is a collec- and categorized all cognates in the respective pairs
tion of jupyter notebooks based on the pandas and as either (i) identical, or (ii) successfully trans-
NumPy libraries2 . To foster reproducibility and formed, or (iii) non-transformable by this rule set.
provide a resource for other researchers to eas- The stimuli selection for the online experiments
ily compute linguistic distances and asymmetries (Section 3.2) is based on the successfully trans-
we make imcom.py available online https: formed ones: 128 items of a total of 935 (from
//github.com/uds-lsv/incompy. In ad- all lists, excluding doublets). In this way we
dition to computing distances and asymmetries could exclude possible different derivational mor-
based on a corpus of word pairs, incom.py read- phemes between related languages (e.g. BG–RU
ily supports visualizing the obtained results. хладен–холодный (chladen–cholodnyj) ‘cold’)
in order to focus on the impact of mismatched or-
3 Data Sources
thographic correspondences for cognate intelligi-
3.1 Language material bility. Even though it may seem artificial to test
The Bulgarian (BG) and Russian (RU) data used in isolated words, the underlying assumption here is
this work comes from a collection of parallel word that correct cognate recognition is a precondition
lists consisting of internationalisms, Pan-Slavic of success in reading intercomprehension. If the
vocabulary, and cognates from Swadesh lists. The reader correctly recognizes a minimal proportion
words belong to different parts of speech, mainly of words, he or she will be able to piece the writ-
nouns, adjectives, and verbs. We chose to use ten message together.
vocabulary lists instead of parallel sentences or
3.2 Web-based experiments
texts in order to exclude the influence of other
linguistic factors. The lists, each containing 120 The orthographic intelligibility between BG
words, were manually adjusted by removing non- and RU was tested in web-based experiments
cognates by possibly substituting them with ety- (http://intercomprehension.coli.
mologically related items, if such could be found, uni-saarland.de/en) in which 71 native
and adding further cognates3 . Thus, for exam- speakers of BG and 94 native speakers of RU took
2
https://jupyter.org, https://pandas. definition; partial cognates are pairs of words which have
pydata.org, https://www.numpy.org the same meaning in both languages only in some contexts,
3
Shared inherited words from Proto-Slavic, shared loans, for example, BG м (măž) ‘man, husband’ and RU му
for example, internationalisms. Cognates are included in the (muž) ‘husband’.
813
Stimuli RU: the BG participants understand a slightly
larger number of the RU words (74.67%) than
Bulgarian Russian
the RU participants understand the BG words they
Bulgarian – 74.67% are presented with (71.33%). This can be ex-
Native Russian 71.33% – plained by the fact that there are only slight differ-
ences between the two languages on the graphic-
Table 1: Intercomprehension scores from free orthographical level (for more details see (Stenger
translation tasks performed by humans. et al., 2017b)).
4 Results
part. The participants started with registration
and then completed a questionnaire in their native 4.1 Levenshtein distance and intelligibility
language. The challenges were presented: 2 with score
each 60 different BG stimuli in each group for Using incom.py we compute the orthographic
RU speakers and 2 with 60 different RU stimuli LD in both directions and further consider the nor-
in each group for BG speakers. The order of malized Levenshtein distance nLD between the
the stimuli were randomized. The participants 120 BG and RU cognates motivated by the as-
saw the stimuli on their screen, one by one, and sumption that a segmental difference in a word
were given 10 seconds4 to translate each word of two segments has a stronger impact on intel-
into RU or into BG. It was also possible to finish ligibility than a segmental difference in a word of
before the 10 seconds were over by either clicking ten segments (Beijering et al., 2008; Stenger et al.,
on the ‘next’ button or pressing ‘enter’ on the 2017a). There is a general assumption that the
keyboard. After 10 seconds the participants saw higher the normalized LD, the more difficult it is
the next stimulus on their screen. During the to translate a given word (Gooskens, 2007; Van-
experiment the participants received feedback in hove and Berthele, 2015; Vanhove, 2016). Thus,
form of emoticons for their answers. The results we correlate the normalized LD and the intelligi-
were automatically categorized as ‘correct’ or bility scores from our experiments for both lan-
‘wrong’ via pattern matching with pre-defined guage pairs. The correlation results are presented
answers: some stimuli had more than one possible in Figure 2. We find a correlation between ortho-
translation and we also provided a list of so-called graphic distance (normalized LD) and the intelli-
alternative correct answers. For example, the gibility of BG words for RU readers of r = –0.57
BG word пт (păt) ‘way’ can be translated in (p = 1.4e − 11) and r = –0.36 (p = 6.3e − 05)
RU as пут (put’) or дорога (doroga), so both for BG readers. Both correlations are significant
translations were counted as correct. and confirm the above hypothesis. However, the
The analysis of the collected material5 is based LD accounts for only 32% (R2 = 0.32) of the
on the answers of 37 native speakers of Bulgarian variance in the intelligibility scores for RU read-
(31 women and 6 men between 18 and 41 years of ers and for only 13% (R2 = 0.13) of the variance
age, average 27 years) and 40 native speakers of in the intelligibility scores for BG readers, leav-
Russian (32 women and 8 men between 18 and 71 ing the majority of variance unexplained. Recall
years of age, average 33 years). The mean percent- from Section 2 that LD is a symmetric measure,
age of correctly translated items constitutes the in- and therefore it does not capture any asymmetries
telligibility score of a given language (Table 1). between correspondences. If, for instance, the RU
The results show that there is virtually no asym- vowel character a always corresponds to a for a
metry in written intelligibility between BG and BG reader, but in the other direction, BG a can
4
The time limit is chosen based on the experience from correspond to a, o or for a RU reader, then a
other reading intercomprehension experiments. The allocated measure of linguistic distance is required to re-
time is supposed to be sufficient for typing even the longest flect both this difference in adaptation possibili-
words, but not long enough for using a dictionary or an online
translation tool. ties and the uncertainty involved in transforming
5
For the present study we exclude those participants who a. Such asymmetries are effectively captured by
have indicated knowledge of the stimuli language(s) in the the next two intelligibility measurements of word
questionnaire and analyze the results only of the initial chal-
lenge for each participant in order to avoid any learning ef- adaptation surprisal and conditional entropy, both
fects. of which are implemented in the incom.py tool-
814
(a) BG-RU (b) RU-BG
Figure 2: Normalized Levenshtein distance as a predictor for intelligibility. 2a Shows Russian native
speakers reading Bulgarian. 2b Shows Bulgarian native speakers reading Russian.
815
(a) BG-RU (b) RU-BG
Figure 3: Normalized word adaptation surprisal as a predictor for intelligibility. 3a Shows Russian native
speakers reading Bulgarian. 3b Shows Bulgarian native speakers reading Russian.
are very similar to the stimulus, the number of ues of 5 RU characters о, е, , у, л for BG read-
mismatched orthographic correspondences in the ers (on the right). Note that the alignment of ∅
stimulus and their position, the word frequency in to any other character c corresponds to the case
one’s own language, the word length of stimulus where Russian readers have to fill in a character.
etc. The estimated character values seem not to The entropy calculations reveal that, for example,
exactly reflect this constellation. BG readers should have more uncertainty with the
RU vowel character о, while RU readers should
4.3 Conditional entropy and intelligibility have more difficulties with the adaptation of the
score BG vowel character e. This means that the map-
ping of the RU о to possible BG characters is more
For the BG–RU language pair the difference in complex than the opposite direction. More pre-
the full conditional entropies (CE) is very small: cisely, the RU о can map into 4 BG vowel charac-
0.4853 bits for the BG to RU transformation and ters (о, а, , е) or to nothing (∅), the BG e can
0.4689 bits for the RU to BG transformation, with map into 3 RU vowel characters (е, ё, or ). Cer-
a very small amount of asymmetry of 0.0164 bits. tainly, in an intercomprehension scenario a BG or
These results predict that speakers of RU reading a RU reader does not know these mappings and the
BG words are more uncertain than speakers of BG respective probabilities. However, the assumption
reading RU words. This is in accordance with is that the measure of complexity of the mapping
the experimental results where the language com- can be used as an indicator for the degree of intel-
bination with the slightly higher CE (RU speak- ligibility (Moberg et al., 2007), because it reflects
ers reading BG) had a slightly lower intelligibility the difficulties with which a reader is confronted in
score (see Table 1). Thus, CE can be a reliable ‘guessing’ the correct correspondence. Our exper-
measure when explaining even the small asymme- imental results indeed show that BG readers have
try in the mutual intelligibility. greater problems with the RU o than RU readers
Using incom.py we calculated entropy val- with the BG character a or nothing (∅) in cognate
ues of BG and RU characters in order to analyse pairs like RU–BG холод – хлад (cholod – chlad)
asymmetries on the written (orthographic) level in ‘cold’, борода – брада (boroda – brada) ‘beard’,
more details. Figure 4 shows the entropy values ворона – врана (vorona – vrana) ‘crow’.
of 6 BG characters е, , а, щ, и, , and the spe-
cial symbol ∅ for RU readers and the entropy val-
816
there is room for improvement in our orthographic
distance algorithm. Word adaptation surprisal
measures the complexity of a mapping, in par-
ticular, how predictable the particular correspon-
dence in a language pair is. The surprisal values
of correspondences are indeed different. How-
ever, they depend on their frequency and distri-
bution in the particular cognate set. Most impor-
Figure 4: Character entropy values when translat-
tant and in contrast to Levenshtein distance, sur-
ing from Russian to Bulgarian and vice versa.
prisal can be asymmetric. The character adapta-
tion surprisal values between language A and lan-
5 Discussion and Outlook guage B are not necessarily the same as between
language B and language A. This indicates an ad-
Previous research in reading intercomprehension vantage of the surprisal-based method compared
has shown that (closely) related languages may be to Levenshtein distance. Our results show that the
differently distant from each other and their mu- predictable potential of word adaptation surprisal
tual intelligibility may be asymmetric. In this pa- was rather weak despite its modification. We as-
per we present incom.py – a toolbox for com- sume that word adaptation surprisal should to a
puting linguistic distances and asymmetries. With larger extent take into account relevant factors in
incom.py we perform experiments on measur- reading intercomprehension, for example, ortho-
ing and predicting the mutual intelligibility of graphic neighbors (words that are very similar to
Slavic languages, as exemplified by the language the stimulus word and differ only in one charac-
pair Bulgarian-Russian by means of the Leven- ter). Something we keep as future work.
shtein distance, word adaptation surprisal, and
conditional entropy. Using a small corpus of par-
allel cognate lists we validated linguistic distances Conditional entropy can reflect the difficul-
and asymmetries as predictors of mutual intelligi- ties humans encounter when mapping one ortho-
bility based on stimuli obtained from written intel- graphic system on another. The underlying hy-
ligibility tests. The results of our statistical anal- pothesis is that high predictability improves intel-
yses clearly support normalized Levenshtein dis- ligibility, and therefore a low entropy value should
tance as a reliable predictor of orthographic intel- correspond to a high intelligibility score. This re-
ligibility at the word level for both language pairs sult is as we expected. We have calculated condi-
tested. However, we find that only 32% (for RU tional entropy for Bulgarian and Russian using a
readers) and 13% (for BG readers) of the variance cognate word list from intelligibility tests. In our
in the intelligibility data is explained by the or- experiments, conditional entropy – like the intel-
thographic similarity quantified by means of the ligibility task – reveals asymmetry between Bul-
normalized Levenshtein distance. We find that garian and Russian on the orthographic level: the
the predictive power of the Levenshtein distance conditional entropy in Bulgarian for Russian read-
is different within the two language pairs. It must ers is slightly higher than the conditional entropy
be mentioned here that the RU stimuli are in gen- in Russian for Bulgarian readers. This means that
eral longer (5.09 characters) than the BG stimuli the slightly higher entropy is found in the lan-
(4.61 characters). Thus, the BG readers should in- guage pair where there is slightly lower intelli-
tuitively delete more characters while the RU read- gibility. Thus, we were able to show that con-
ers should add more characters in order to guess ditional entropy can be a reliable measure when
the correct cognate. explaining small asymmetries in intelligibility. In
Previous research has shown that deletions and future work we plan to extend incom.py with
additions, the basic operations performed when additional functionality to compute distances and
computing Levenshtein distance, are not of equal asymmetries on the phonological level. Addition-
value in the mutual intelligibility: it appears that ally, it might be interesting to consider the mor-
deletions are more transparent for the participants phological level which has been shown to be help-
in terms of subjective similarity than additions ful when processing words for humans with lim-
(Kaivapalu and Martin, 2017). This means that ited reading abilities (Burani et al., 2008).
817
Acknowledgments Jens Moberg, Charlotte Gooskens, John Nerbonne,
and Nathan Vaillette. 2007. Conditional entropy
We would like to thank Klára Jágrová and Volha measures intelligibility among related languages.
Petukhova for their helpful feedback on this pa- Proceedings of Computational Linguistics in the
Netherlands .
per. Furthermore, we thank the reviewers for their
valuable comments. This work has been funded Saul B Needleman and Christian D Wunsch. 1970. A
by Deutsche Forschungsgemeinschaft (DFG) un- general method applicable to the search for simi-
der grant SFB 1102: Information Density and Lin- larities in the amino acid sequence of two proteins.
Journal of molecular biology 48(3).
guistic Encoding.
Håkan Ringbom. 2007. Cross-linguistic similarity in
foreign language learning, volume 21. Multilingual
References Matters.
Karin Beijering, Charlotte Gooskens, and Wilbert Claude Elwood Shannon. 1948. A mathematical the-
Heeringa. 2008. Predicting intelligibility and per- ory of communication. Bell system technical jour-
ceived linguistic distance by means of the Leven- nal 27(3).
shtein algorithm. Linguistics in the Netherlands
25(1). Irina Stenger, Tania Avgustinova, and Roland Marti.
2017a. Levenshtein distance and word adaptation
Gaetano Berruto. 2004. Sprachvarietät-sprache surprisal as methods of measuring mutual intel-
(Gesamtsprache, historische Sprache) . ligibility in reading comprehension of slavic lan-
guages. In Computational Linguistics and Intellec-
tual Technologies: International Conference ‘Dia-
Cristina Burani, Stefania Marcolini, Maria De Luca,
logue 2017’Proceedings. volume 16.
and Pierluigi Zoccolotti. 2008. Morpheme-based
reading aloud: Evidence from dyslexic and skilled Irina Stenger, Klára Jágrová, Andrea Fischer, Tania
italian readers. Cognition 108(1). Avgustinova, Dietrich Klakow, and Roland Marti.
2017b. Modeling the impact of orthographic cod-
Andrea Fischer, Klára Jágrová, Irina Stenger, Tania ing on Czech–Polish and Bulgarian–Russian reading
Avgustinova, Dietrich Klakow, and Roland Marti. intercomprehension. Nordic Journal of Linguistics
2015. An orthography transformation experiment 40(2).
with Czech-Polish and Bulgarian-Russian parallel
word sets. Natural Language Processing and Cog- Jan Vanhove. 2016. The early learning of interlingual
nitive Science 2015 Proceedings, Libreria Editrice correspondence rules in receptive multilingualism.
Cafoscarina, Venezia . International Journal of Bilingualism 20(5).
Charlotte Gooskens. 2007. The contribution of linguis- Jan Vanhove and Raphael Berthele. 2015. Item-related
tic factors to the intelligibility of closely related lan- determinants of cognate guessing in multilinguals.
guages. Journal of multilingual and multicultural Crosslinguistic influence and crosslinguistic inter-
development 28(6). action in multilingual language learning .
818
A Holistic Natural Language Generation Framework for the Semantic
Web
819
Proceedings of Recent Advances in Natural Language Processing, pages 819–828,
Varna, Bulgaria, Sep 2–4 2019.
Duma and Klein, 2013; Ell and Harth, 2014; Biran of the restrictions on the individuals’ characteris-
and McKeown, 2015). However, the subject of re- tics. OWL offers existential (SOME) or universal
search has only recently gained significant momen- (ONLY) qualifiers and a variety of typical logical
tum due to the great number of published works in constructs, such as negation (NOT), other Boolean
the WebNLG (Colin et al., 2016) challenge along operators (OR, AND), and more constructs such as
with deep learning techniques (Sleimi and Gardent, cardinality restriction (MIN, MAX, EXACTLY)
2016; Mrabet et al., 2016). RDF has also been and value restriction (VALUE), to create class ex-
showing promising benefits to the generation of pressions. Such constructs can be combined in
benchmarks for evaluating NLG systems (Gardent arbitrarily complex class expressions CE according
et al., 2017; Perez-Beltrachini et al., 2016). to the following grammar
Despite the plethora of recent works written on
CE = A | C AND D | C OR D | NOT C | R
handling RDF data, only a few have exploited the SOME C | R ONLY C | R MIN n | R MAX
generation of NL from OWL and SPARQL. For n | R EXACTLY n | R VALUE a | {a1
instance, Androutsopoulos et al. (2013) generates ,...,am }
sentences in English and Greek from OWL ontolo-
gies. Also, SPARQL2NL (Ngonga Ngomo et al., where A is an atomic class, C and D are class ex-
2013) uses rules to verbalize atomic constructs and pressions, R is an object property, a as well as a1
combine their verbalization into sentences. There- to am with m ≥ 1 are individuals, and n ≥ 0 is an
fore, our goal with LD2NL is to provide a complete integer.
framework to verbalize SW concepts rather than Axioms are statements that are asserted to be
become the state of the art on the respective tasks. true in the domain being described. Usually,
one distinguish between (1) terminological and
3 Background (2) assertional axioms. (1) terminological ax-
ioms are used to describe the structure of the
3.1 OWL domain, i.e., the relationships between classes
OWL2 (OWL Working Group, 2009) is the de-facto resp. class expressions. For example, using
standard for machine processable and interoperable a subclass axiom (SubClassOf:), one can state
ontologies on the SW. In its second version, OWL that the class :Koala is a subclass of the class
is equivalent to the description logic SROIQ(D). :Animal. Classes can be subclasses of other
Such expressiveness has a higher computational classes, thus creating a taxonomy. In addi-
cost but allows the development of interesting ap- tion, axioms can arrange properties in hierarchies
plications such as automated reasoning (Bühmann (SubPropertyOf:) and can assign various char-
et al., 2016). OWL 2 ontologies consist of the acteristics (Characteristics:) such as transitivity
following three different syntactic categories: or reflexivity to them. (2) Assertional axioms
Entities, such as classes, properties, and individ- formulate facts about individuals, especially the
uals, are identified by IRIs. They form the primi- classes they belong to and their mutual relation-
tive terms and constitute the basic elements of an ships. OWL can be expressed in various syntaxes
ontology. Classes denote sets of individuals and with the most common computer readable syntax
properties link two individuals or an individual and being RDF/XMLA more human-readable format
a data value along a property. For example, a class is the Manchester OWL Syntax (MOS) (Horridge
:Animal can be used to represent the set of all et al., 2006). For example, the class expression
animals. Similarly, the object property :childOf that models people who work at a university that is
can be used to represent the parent-child relation- located in Spain could be as follows in MOS:
ship and the data property :birthDate assigns Person AND worksAt SOME (University AND
a particular birth date to an individual. Finally, locatedIn VALUE Spain)
the individual :Alice can be used to represent a
particular person called ”Alice”. Likewise, expressing that every professor works at
Expressions represent complex notions in the a university would read as
domain being described. For example, a class ex-
pression describes a set of individuals in terms Class: Professor
SubClassOf: worksAt SOME University
2
www.w3.org/TR/owl2-overview/
820
3.2 RDF representation for the most common used SW mod-
RDF (RDF Working Group, 2014) uses a graph- eling languages RDF and OWL, and SPARQL. In
based data model for representing knowledge. terms of the standard model of NL generation pro-
Statements in RDF are expressed as so-called posed by Reiter & Dale (Reiter and Dale, 2000b),
triples of the form (subject, predicate, our steps mainly play the role of the micro-planner,
object). RDF subjects and predicates are In- with focus on aggregation, lexicalization, referring
ternationalized Resource Identifierss (IRIs) and ob- expressions and linguistic realization. In the fol-
jects are either IRIs or literals.3 RDF literals always lowing, we present our approach to formalizing NL
have a datatype that defines its possible values. A sentences for each of the supported languages.
predicate denotes a property and can also be seen as 4.1 From RDF to NL
a binary relation taking subject and object as argu-
ments. For example, the following triple expresses 4.1.1 Lexicalization
that Albert Einstein was born in Ulm: The lexicalization of RDF triples must be able to
deal with resources, classes, properties and literals.
:Albert_Einstein :birthPlace :Ulm .
Classes and resources The lexicalization of
classes and resources is carried out as follows:
3.3 SPARQL Given a URI u we ask for the English label of
Commonly, the selection of subsets of RDF is u using a SPARQL query.5 If such a label does not
performed using the SPARQL query language.4 exist, we use either the fragment of u (the string
SPARQL can be used to express queries across di- after #) if it exists, else the string after the last oc-
verse data sources. Query forms contain variables currence of /. Finally this NL representation is
that appear in a solution result. They can be used realized as a noun phrase, and in the case of classes
to select all or a subset of the variables bound in is also pluralized. As an example, :Person is
a pattern match. They exist in four different in- realized as people (its label).
stantiations, i.e., SELECT, CONSTRUCT, ASK and Properties The lexicalization of properties re-
DESCRIBE. The SELECT query form is the most lies on the insight that most property labels are
commonly used and is used to return rows of vari- either nouns or verbs. While the mapping of a
able bindings. Therefore, we use this type of query particular property p can be unambiguous, some
in our explanation. CONSTRUCT allows to cre- property labels are not as easy to categorize. For
ate a new RDF graph or modify the existing one examples, the label crosses can either be the
through substituting variables in a graph templates plural form of the noun cross or the third person
for each solution. ASK returns a Boolean value singular present form of the verb to cross. To
indicating whether the graph contains a match or automatically determine which realization to use,
not. Finally, DESCRIBE is used to return all triples we relied on the insight that the first and last word
about the resources matching the query. For exam- of a property label are often the key to determining
ple, 1 represents the following query “Return all the type of the property: properties whose label
scientists who were born in Ulm”. begins with a verb (resp. noun or gerund) are most
to be realized as verbs (resp. nouns). We devised
SELECT ?person a set of rules to capture this behavior, which we
WHERE {
?person a dbo:Scientist;
omit due to space restrictions. In some cases (such
dbo:birthPlace dbr:Ulm. as crosses) none of the rules applied. In these
} cases, we compare the probability of P (p|noun)
Listing 1: All scientists who were born in Ulm and P (p|verb) by measuring
P
log2 (f (t))
t∈synset(p|X)
P (1)
4 LD2NL Framework P (p|X) =
log2 (f (t0 ))
,
t0 ∈synset(p)
821
the syntactic class X ∈ {noun, verb} and f (t) is 4.2 Realization - RDF Triples to NL
the frequency of use of p in the sense of the synset The same procedure of generating a single triple
t according to WordNet. For can be applied for the generation of each triple in a
P (p|verb) set of triples. However, the NL output would con-
≥ θ, (2)
P (p|noun) tain redundant information and consequently sound
we choose to realize p as a noun; else we realized it very artificial. Thus, the goal is to transform the
as a verb. For θ = 1, for example, dbo:crosses generated description to sound more natural. To
is realized as a verb. this end, we focus on two types of transformation
Literals Literals in an RDF graph usually con- rules (cf. (Dalianis and Hovy, 1996)): ordering and
sist of a lexical form LF and a datatype IRI DT, clustering and grouping. In the following, we de-
represented as "LF"ˆˆ<DT>. Optionally, if the scribe the transformation rules we employ in more
datatype is rdf:langString, a non-empty lan- detail. Note that clustering and ordering (4.2.1) is
guage tag is specified and the literal is denoted applied before grouping (4.2.2).
as language-tagged string6 . The realization of 4.2.1 Clustering and ordering rules
language-tagged strings is done by using simply
We process the input trees in descending order with
the lexical form, while omitting the language tag.
respect to the frequency of the variables they con-
For example, "Albert Einstein"@en is re-
tain, starting with the projection variables and only
alized as Albert Einstein. For other types
after that turning to other variables. As an example,
of literals, we further differentiate between built-in
consider the following triples about two of the most
and user-defined datatypes. For the former, we also
known people in the world:
use the lexical form, e.g. "123"ˆˆxsd:int ⇒
123, while the latter are processed by using the :William_Shakespeare rdf:type :Writer .
:Albert_Einstein :birthPlace :Ulm .
literal value with its representation of the datatype :Albert_Einstein :deathPlace :Princeton
IRI, e.g., "123"ˆˆdt:squareKilometre as :Albert_Einstein rdf:type :Scientist .
123 square kilometres. :William_Shakespeare :deathDate
"1616-04-23"ˆˆxsd:date .
4.1.2 Realizing single triples
The five triples are verbalized as given in 3a–
The realization ρ of a triple (s p o) depends
3e. Clustering and ordering first take all sentences
mostly on the verbalization of the predicate p. If p
containing the subject :Albert Einstein, i.e.
can be realized as a noun phrase, then a possessive
3b –3d, which are ordered such that copula-
clause can be used to express the semantics of (s
tive sentences (such as Albert Einstein is a scien-
p o), more formally
tist) come before other sentences, and then takes
1. ρ(s p o) ⇒ all sentences containing the remaining subject
poss(ρ(p),ρ(s)) ∧ subj(BE,ρ(p))
∧ dobj(BE,ρ(o)) :William Shakespeare in 3a and 3e result-
ing in a sequence of sentences as in 4.
For example, if ρ(p) is a relational noun
like birth place e.g. in the triple 3. (a) William Shakespeare is a writer.
(b) Albert Einstein’s birth place is Ulm.
(:Albert Einstein :birthPlace
(c) Albert Einstein’s death place is Princeton.
:Ulm), then the verbalization is Albert
(d) Albert Einstein is a scientist.
Einstein’s birth place is Ulm. Note (e) William Shakespeare’s death date is 23 April
that BE stands for the verb “to be”. In case 1616.
p’s realization is a verb, then the triple can be
4. Albert Einstein is a scientist. Albert Einstein’s
verbalized as follows: birth place is Ulm. Albert Einstein’s death place
2. ρ(s p o) ⇒ is Princeton. William Shakespeare’s is a writer.
subj(ρ(p),ρ(s)) ∧ dobj(ρ(p),ρ(o)) William Shakespeare’s death date is 23 April 1616.
822
phrases and verb phrases (verbs together with ob- it would lead to a non-human understandable text
ject noun phrases), leading to subject grouping and in many cases. For example, the intersection of two
object grouping. To maximize the grouping ef- classes :A and :B can be represented in RDF by
fects, we collapse common prefixes and suffixes the six triples
of sentences, irrespective of whether they are full _:x rdf:type owl:Class .
subject noun phrases or complete verb phrases. In _:x owl:intersectionOf _:y1 .
the following we use X1 , X2 ,. . . XN as variables _:y1 rdf:first :A .
_:y1 rdf:rest _:y2 .
for the root nodes of the input sentences and Y as _:y2 rdf:first :B .
variable for the root node of the output sentence. _:y2 rdf:rest rdf:nil .
Furthermore, we abbreviate a subject subj(Xi , si )
The verbalization of these triples would re-
as si , an object dobj(Xi , oi ) as oi , and a verb
sult in Something that is a class
root(ROOTi , vi ) as vi .
and the intersection of something
Subject grouping collapses the predicates (i.e.
whose first is A and whose rest
verb and object) of two sentences if their subjects
is something whose first is B and
are the same, as specified in 5 (abbreviations as
whose rest ist nil., which is obviously
above).
far away from how a human would express it
5. ρ(s1 ) = ρ(s2 ) ∧ cc(v1 , coord) in NL. Therefore, generating NL from OWL
⇒ root(Y, coord(v1 , v2 )) ∧ subj(v1 , s1 ) ∧
dobj(v1 , o1 ) ∧ subj(v2 , s1 ) ∧ dobj(v1 , o2 ) requires a different procedure based on its syntactic
categories, OWL expressions and OWL axioms.
An example are the sentences given in 6, which We show the general rules for each of them in the
share the subject Albert Einstein and thus can be following.
collaped into a single sentence.
6. Albert Einstein is a scientist and Albert Einstein is 4.3.1 OWL Class Expressions
known for general relativity. In theory, class expressions can be arbitrarily com-
⇒ Albert Einstein is a scientist and known for
general relativity.
plex, but as it turned out in some previous analy-
sis (Power and Third, 2010), in practice they sel-
Object grouping collapses the subjects of two dom arise and can be seen as some corner cases.
sentences if the realizations of the verbs and ob- For example, an ontology could contain the follow-
jects of the sentences are the same, where the ing class expression about people and their birth
coord ∈ {and, or} is the coordination combining place:
the input sentences X1 and X2 , and coord ∈ Person AND birthPlace SOME (City AND
{conj, disj} is the corresponding coordination locatedIn VALUE France)
combining the subjects.
Class expressions do have a tree-like structure and
7. ρ(o1 ) = ρ(o2 ) ∧ ρ(v1 ) = ρ(v2 ) ∧ cc(v1 , coord)
⇒ root(Y, PLURAL(v1 )) ∧ can simply be parsed into a tree by means of the
subj(v1 , coord(s1 , s2 )) ∧ dobj(v1 , o1 ) binary OWL class expressions constructors con-
For example, the sentences in 8 share their verb tained in it. For our example, this would result in
and object, thus they can be collapsed into a single the following tree:
AND
sentence. Note that to this end the singular aux-
iliary was needs to be transformed into its plural
SOME
form were.
8. Benjamin Franklin was born in Boston. Leonard
Nimoy was born in Boston. ⇒ Benjamin Franklin AND
and Leonard Nimoy were born in Boston.
823
class expression in our example, i.e. we omit assertional axioms assert individuals to atomic
birthPlace, locatedIn, and France. classes or relate individuals to another individual
Moreover and again for simplicity, we’ll explain resp. literal value. For example, axioms about the
the transformation process by starting from type as well as birth place and birth date of Albert
the right-hand side of the tree. Thus, in our Einstein can be expressed by
example we begin with the class expression Individual: Albert_Einstein
City which is transformed to everything Types: Person
that is a city and locatedIn VALUE Facts: birthPlace Ulm, birthDate "
1879-03-14"ˆˆxsd:date
France resulting in everything that is
located in France by application of a rule. Those axioms can simply be rewritten as triples,
Both class expressions are used in the conjunction thus, we can use the same procedure as we do
City AND locatedIn VALUE France. for triples (Section 4.1.2). Converting them
Thus, the next step would be to merge both into NL gives us Albert Einstein is a
phrases. An easy way is to use the coordinating person whose birth place is Ulm
conjunction and, i.e. everything that and whose birth date is 14 March
is a city and everything that is 1879. OWL also allows for assigning an
located in France. Although the output individual to a complex class expression. In
of this transformation is correct, it still contains that case we’ll use our conversion of OWL class
unnecessarily redundant information. Therefore, expressions as described in Section 4.3.1.
we apply the aggregation procedure described Terminological Axioms (TBox Axioms) - Ac-
in Section 4.2.2, i.e. we get everything that cording to Power and Third (2010), most of the ter-
is a city and located in France. minological axioms used in ontologies are subclass
Yet, the aggregation can still be improved: if there axioms. By definition, subclass and superclass can
is any atomic class in the conjunction, we know be arbitrarily complex class expressions CE1 and
that this is more specific than the placeholder CE2 , i.e. CE1 SubClassOf CE2 , but in praxis it is
everything. Thus, we can replace it by the quite often only used with atomic classes as sub-
plural form of the class, finally resulting in class or even more simple with the superclass also
cities that are located in France. beeing an atomic class. Nevertheless, we support
The same procedure is applied for its parent class any kind of subclass axiom and all other logical
expression being the existential restriction OWL axioms in LD2NL. For simplicity, we outline
birthPlace SOME (City AND locatedIn here how we verbalize subclass axioms in LD2NL.
VALUE France) The semantics of a subclass axiom denotes that
every individual of the subclass also belongs to the
This will be transformed to everything superclass. Thus, the verbalization seems to be
whose birth place is a city that relatively straightforward, i.e. we verbalize both
is located in France. Note, that we used class expressions and follow the template : every
the singular form here, assuming that the property ρ(CE1 ) is a ρ(CE2 ). Obviously, this works pretty well
birthPlace is supposed to be functional for subclass axioms with atomic classes only. For
in the ontology. In the last step, we process example, the axiom
the class expression Person, which gives us
Class: Scientist
everything that is a person. Again, SubClassOf: Person
due to the conjunction we merge this result with
with the previous one, such that in the end we is verbalized as every scientist is a
get people whose birth place is a person.
city that is located in France.
4.4 From SPARQL to NL
4.3.2 OWL Axioms A SPARQL SELECT query can be regarded as con-
As we described in Section 4.3, OWL axioms can sisting of three parts: (1) a body section B, which
roughly be categorized into terminological and as- describes all data that has to be retrieved, (2) an
sertional axioms. Therefore, we have different pro- optional section O, which describes the data items
cedures for processing each category: that can be retrieved by the query if they exist,
Assertional Axioms (ABox Axioms) - Most and (3) a modifier section M, which describes all
824
solution sequences, modifiers and aggregates that regarding the two following measures according
are to be applied to the result of the previous two to Gardent et al. (2017): (1) Adequacy: Does the
sections of the query. Let Var be the set of all text contain only and all the information from the
variables that can be used in a SPARQL query. In data? (2) Fluency: Does the text sound fluent and
addition, let R be the set of all resources, P the set natural?. For both measures the volunteers were
of all properties and L the set of all literals con- asked to rate on a scale from 1 (Very Bad) to 5
tained in the target knowledge base of the SPARQL (Very Good). The experiment was carried out us-
queries at hand. We call x ∈ Var ∪ R ∪ P ∪ L ing 41 axioms of the Koala ontology.9 Because
an atom. The basic components of the body of of the complexity of OWL axioms, only domain
a SPARQL query are triple patterns (s, p, o) ∈ experts were asked to perform this experiment.
(Var ∪ R) × (Var ∪ P ) × (Var ∪ R ∪ L). Let In the second experiment, a set of triples describ-
W be the set of all words in the dictionary of our ing a single resource and their verbalization were
target language. We define the realization func- shown to the volunteers. The experts were asked
tion ρ : Var ∪ R ∪ P ∪ L → W ∗ as the function to rate the verbalization regarding adequacy, flu-
which maps each atom to a word or sequence of ency and completeness, i.e., whether all triples have
words from the dictionary. The extension of ρ to all been covered. The non-experts were only asked to
SPARQL constructs maps all atoms x to their real- rate the fluency. The experiment was carried out
ization ρ(x) and defines how these atomic realiza- using 6 DBpedia resources.In the third experiment,
tions are to be combined. We denote the extension the verbalization of an OWL class and 5 resources
of ρ by the same label ρ for the sake of simplicity. were shown to the human raters. For non-experts,
We adopt a rule-based approach to achieve this goal, the resources have been verbalized as well, while
where the rules extending ρ to all valid SPARQL for domain experts the resources were presented
constructs are expressed in a conjunctive manner. as triples. The task of the raters was to identify
This means that for premises P1 , . . . , Pn and con- the resource that fits the class description and, thus,
sequences K1 , . . . , Km we write P1 ∧ . . . ∧ Pn ⇒ is an instance of the class. We used 4 different
K1 ∧. . .∧Km . The premises and consequences are OWL axioms and measured the amount of correct
explicated by using an extension of the Stanford identified class instances.
dependencies8 . Results In our first series of experiments, the ver-
For example, a possessive dependency between balization of OWL axioms, we achieved an average
two phrase elements e1 and e2 is represented as adequacy of 4.4 while the fluency reached 4.38. In
poss(e1 , e2 ). For the sake of simplicity, we addition, more than 77% of the verbalizations were
slightly deviate from the Stanford vocabulary by assigned the maximal adequacy (i.e., were assigned
not treating the copula to be as an auxiliary, but a score of 5, see Fig. 1). The maximal score for flu-
denoting it as BE. Moreover, we extend the vo- ency was achieved in more than 69% of the cases
cabulary by the constructs conj and disj which (see Fig. 1). This clearly indicates that the ver-
denote the conjunction resp. disjunction of two balization of axioms generated by LD2NL can be
phrase elements. In addition, we sometimes reduce easily understood by domain experts and contains
the construct subj(y,x) ∧ dobj(y,z) to the all the information necessary to access the input
triple (x,y,z) ∈ W 3 . OWL class expression.
Experiments on the verbalization of summaries
5 Experiments for RDF resources revealed that verbalizing re-
We evaluated our approach in three different exper- source summaries is a more difficult task. While
iments based on human ratings. We divided the the adequacy of the verbalization was assigned an
volunteers into two groups—domain experts and average score of 3.92 by experts (see Fig. 2), the
non-experts. The group of domain experts com- fluency was assigned a average score of 3.47 by
prised 66 persons while there were 20 non-experts experts and 3.0 by non-experts (see Fig. 2). What
forming the second group. In the first experiment, these results suggest is that (1) our framework
an OWL axiom and its verbalization were shown to generates sentences that are close to that which
the experts who were asked to rate the verbalization a domain expert would also generate (adequacy).
However (2) while the sentence is grammatically
8
For a complete description of the vocabulary, see https:
9
//stanford.io/2EzMjmo. https://bit.ly/2K8BWts
825
5 5
4 4
Adequacy
Fluency
3 3
2 2
1 1
0 100 200 300 400 500 0 100 200 300 400 500
Number of ratings Number of ratings
Experts Experts
5 5 5
Completeness
4 4 4
Adequacy
Fluency
3 3 3
2 2 2
1 1 1
Figure 2: Experiment II: adequacy (left), fluency (middle) and completeness (left) results
sufficient for the experts, it is regarded by non- our first conclusion is that our framework clearly
domain experts (which were mostly linguists, i.e., serves its purpose. Still, potential improvements
the worst-case scenario for such an evaluation) as can be derived from the results achieved during the
being grammatically passably good but still worthy experiments. In particular, we will consider the
of improvement. The completeness rating achieves used of attention-based encoder-decoder networks
a score of 4.31 on average (see Fig. 2). This was to improve the fluency of complex sentences.
to be expected as we introduced a rule to shorten
the description of resources that contain more than 6 Conclusion and Future Work
5 triples which share a common subject and predi- In this paper, we presented LD2NL, a framework
cate. Finally, we measured how well the users and for verbalizing SW languages, especially on RDF
experts were able to understand the meaning of the and OWL while including the SPARQL verbal-
text generated by our approach. As expected, the ization provided by SPARQL2NL. Our evaluation
domain experts outperform the non-expert users by with 86 persons revealed that our framework gener-
being able to find the answers to 87.2% of the ques- ates NL that can be understood by lay users. While
tions. The score achieved by non-domain experts, the OWL verbalization was close to NL, the RDF
i.e., 80%, still suggest that our framework is able to was less natural but still sufficient to convey the
bridge the gap pertaining to understand RDF and meaning expressed by the corresponding set of
OWL for non-experts from 0% to 80%, which is triples. In future work, we aim to extend LD2NL
more than 91.8% of the performance of experts. to verbalize the languages SWRL (Horrocks et al.,
Discussion Our evaluation results suggest that 2004) and SHACL (Knublauch and Kontokostas,
the verbalization of these languages is a non-trivial 2017).
task that can be approached by using a bottom-up
Acknowledgments
approach. As expected, the verbalization of short
expressions leads to sentences which read as if This work was supported by the German Fed-
they have been generated by a human. However, eral Ministry of Transport and Digital Infrastruc-
due to the complexity of the semantics that can ture (BMVI) through the projects LIMBO (no.
be expressed by the languages at hand, long ex- 19F2029I) and OPAL (no. 19F2028A). This work
pressions can sound mildly artificial. Our results was supported by the German Federal Ministry
however also suggest that although the text gener- of Economics and Technology (BMWI) in the
ated can sound artificial, it is still clear enough to projects RAKI (no. 01MD19012D) as well as by
enable non-expert users to achieve results that are the BMBF project SOLIDE (no. 13N14456).
comparable to those achieved by experts. Hence,
826
References Ian Horrocks, Peter F. Patel-Schneider, Harold Boley,
Said Tabet, Benjamin Grosofand, and Mike Dean.
Ion Androutsopoulos, Gerasimos Lampouras, and Dim- 2004. SWRL: A semantic web rule language com-
itrios Galanis. 2013. Generating natural language bining OWL and RuleML. W3C Member Submis-
descriptions from OWL ontologies: The natural owl sion.
system. J. Artif. Int. Res. 48(1):671–715.
Or Biran and Kathleen McKeown. 2015. Discourse Holger Knublauch and Dimitris Kontokostas. 2017.
planning with an n-gram model of relations. In Shapes constraint language (shacl). W3C Candidate
EMNLP. pages 1973–1977. Recommendation 11(8).
Lorenz Bühmann, Jens Lehmann, and Patrick West- Jens Lehmann, Tim Furche, Giovanni Grasso, Axel-
phal. 2016. Dl-learnera framework for inductive Cyrille Ngonga Ngomo, Christian Schallhart, An-
learning on the semantic web. Journal of Web Se- drew Jon Sellers, Christina Unger, Lorenz Bühmann,
mantics 39:15–24. Daniel Gerber, Konrad Höffner, David Liu, and
Sören Auer. 2012. DEQA: Deep Web Extraction for
Philipp Cimiano, Janna Lüker, David Nagel, and Question Answering. In The Semantic Web - ISWC
Christina Unger. 2013. Exploiting ontology lex- 2012 - 11th International Semantic Web Conference,
ica for generating natural language texts from rdf Boston, MA, USA, November 11-15, 2012, Proceed-
data. In Proceedings of the 14th European Work- ings, Part II. pages 131–147.
shop on Natural Language Generation. ACL, Sofia,
Bulgaria, pages 10–19. Yassine Mrabet, Pavlos Vougiouklis, Halil Kilicoglu,
Claire Gardent, Dina Demner-Fushman, Jonathon
Emilie Colin, Claire Gardent, Yassine Mrabet, Shashi Hare, and Elena Simperl. 2016. Aligning texts and
Narayan, and Laura Perez-Beltrachini. 2016. The knowledge bases with semantic sentence simplifica-
webnlg challenge: Generating text from dbpedia tion. WebNLG 2016 .
data. In Proceedings of the 9th INLG conference.
pages 163–167. Axel-Cyrille Ngonga Ngomo, Lorenz Bühmann,
Christina Unger, Jens Lehmann, and Daniel Gerber.
H. Dalianis and E.H. Hovy. 1996. Aggregation in
2013. Sorry, I Don’t Speak SPARQL: Translating
natural language generation. In G. Adorni and
SPARQL Queries into Natural Language. In Pro-
M. Zock, editors, Trends in natural language genera-
ceedings of the 22Nd International Conference on
tion: an artificial intelligence perspective, Springer,
World Wide Web. ACM, New York, NY, USA, pages
volume 1036 of Lecture Notes in Artificial Intelli-
977–988.
gence, pages 88–105.
Daniel Duma and Ewan Klein. 2013. Generating nat- Axel-Cyrille Ngonga Ngomo, Lorenz Bühmann,
ural language from linked data: Unsupervised tem- Christina Unger, Jens Lehmann, and Daniel Gerber.
plate extraction. In IWCS. pages 83–94. 2013. Sorry, i don’t speak sparql: translating sparql
queries into natural language. In Proceedings of the
Basil Ell and Andreas Harth. 2014. A language- 22nd international conference on World Wide Web.
independent method for the extraction of rdf verbal- ACM, pages 977–988.
ization templates. In INLG. pages 26–34.
Axel-Cyrille Ngonga Ngomo, Lorenz Bühmann,
Basil Ell, Denny Vr, and Elena Simperl. 2012. SPAR- Christina Unger, Jens Lehmann, and Daniel Gerber.
TIQULATION Verbalizing SPARQL queries. In In 2013. SPARQL2NL: Verbalizing SPARQL queries.
Proceedings of ILD Workshop, ESWC. In 22nd International World Wide Web Conference,
Rio de Janeiro, Brazil, May 13-17. pages 329–332.
Basil Ell, Denny Vrandecic, and Elena Paslaru Bon-
tas Simperl. 2011. Labels in the web of data. In W3C OWL Working Group. 2009. OWL 2 Web Ontol-
Proceedings of ISWC. Springer, volume 7031, pages ogy Language: Document Overview. W3C Recom-
162–176. mendation.
Claire Gardent, Anastasia Shimorina, Shashi Narayan,
and Laura Perez-Beltrachini. 2017. Creating train- Laura Perez-Beltrachini, Rania Sayed, and Claire Gar-
ing corpora for nlg micro-planning. In Proceedings dent. 2016. Building rdf content for data-to-text gen-
of ACL. eration. In COLING. pages 1493–1502.
Albert Gatt and Emiel Krahmer. 2017. Survey of the Richard Power and Allan Third. 2010. Expressing
state of the art in natural language generation: Core OWL Axioms by English Sentences: Dubious in
tasks, applications and evaluation. arXiv preprint Theory, Feasible in Practice. In Proceedings of the
arXiv:1703.09902 . 23rd International Conference on Computational
Linguistics: Posters. ACL, Stroudsburg, PA, USA,
Matthew Horridge, Nick Drummond, John Goodwin, pages 1006–1013.
Alan L Rector, Robert Stevens, and Hai Wang. 2006.
The Manchester OWL syntax. In OWLed. volume W3C RDF Working Group. 2014. RDF 1.1 Concepts
216. and Abstract Syntax. W3C Recommendation.
827
Ehud Reiter and Robert Dale. 2000a. Building natural
language generation systems. Cambridge university
press.
Ehud Reiter and Robert Dale. 2000b. Building natu-
ral language generation systems. Cambridge Uni-
versity Press, New York, NY, USA.
Amin Sleimi and Claire Gardent. 2016. Generat-
ing paraphrases from dbpedia using deep learning.
WebNLG 2016 page 54.
828
Building a Comprehensive Romanian Knowledge Base for Drug
Administration
Bogdan Nicula1 , Mihai Dascalu1 , Maria-Dorinela Sirbu1 , Stefan Trăus, an-Matu1 and Alexandru Nuta2
1
University Politehnica of Bucharest, 313 Splaiul Independentei, 060042, Bucharest, Romania
2
All Business Management, Str. Vasile Conta 19, sc. A, et. 1, ap. 8, Bucharest, Romania
829
Proceedings of Recent Advances in Natural Language Processing, pages 829–836,
Varna, Bulgaria, Sep 2–4 2019.
The second section of this paper covers simi- Washington, which focuses on the structure of the
lar knowledge bases, as well as systems provid- human body. Generic ontologies are also avail-
ing similar information for English language. The able, such as DBpedia (Bizer et al., 2009) built by
third section describes the data extraction and the using data from Wikipedia, but they do not offer
architecture of our knowledge base. The last two information relevant for the task at hand.
sections focus on the current results, shortcom-
ings and ways of further improving our knowledge 2.2 Medical Applications
base.
Several applications offering drug-related infor-
2 Related Work mation exist for English Language. They allow
users to search a drug by name, illness or medi-
2.1 Medical Ontologies cal procedure, and offer the possibility of testing
Multiple knowledge bases for English language whether two drugs are incompatible due to harm-
exist, covering different medical areas of interest. ful interactions between their active substances.
Part of them were developed by authors of the One of the most known portals making medi-
Open Biological and Biomedical Ontology (OBO) cal information available and easy to interpret for
Foundry (Smith et al., 2007) which focuses on non-professional users is WebMD. WebMD offers
collaborations and on defining a common set of two applications, Medscape and the WebMD app.
principles for developing medical ontologies. The Medscape is focused on the more practical aspects
foundry’s mission is to develop a set of inter- of healthcare, offering features such as identifying
operable ontologies that are well formed and sci- pills based on a set of physical features, comput-
entifically accurate. These ontologies are built us- ing the body mass index (BMI) or other relevant
ing semantic web technologies (Berners-Lee et al., metrics based on user’s input, and searching for
2001) and they are usually made available in OWL nearby medical professionals and hospitals. The
(McGuinness et al., 2004) format. Over 150 on- WebMD app is focused more on offering theoreti-
tologies are currently listed on the OBO webpage cal insights regarding drugs and diseases. It offers
(http://www.obofoundry.org/). the possibility of searching a disease based on a
Out of all the ontologies developed by OBO list of symptoms, searching for remedies based on
members, the following knowledge bases are rele- age, gender and severity of the symptoms, and it
vant for the functionalities presented in this paper: can notify the user when the US Food and Drug
Administration (FDA) has published new infor-
• CHEBI (Chemical Entities of Biological In- mation regarding a drug from the user’s treatment
terest), containing information regarding a profile.
diverse set of chemical compounds relevant
Other applications, such as Drugs.com, offer
for biological interests (Hastings et al., 2015).
similar features, but also take into account com-
• DINTO (The Drug-Drug Interactions Ontol- munity feedback as a mechanism for informing
ogy), covering information on how 2 active users. The application facilitates communication
ingredients interact one with another; DINTO within its user-base, allowing customers to find
is integrated with CHEBI. relevant and useful information from peers who
underwent similar experiences. This application
• DOID (Human Disease Ontology), a taxon- also differentiates itself from the rest by offering
omy of human diseases (Kibbe et al., 2014). two different views based on the user’s medical
proficiency. Users with little medical knowledge
• SYMP (Symptom Ontology), including
are directed towards pages with simple and easy-
symptoms which may be indicative of a
to-grasp information, while experts are provided
disease. SYMP was developed as part of the
access to more complex content, which includes
Gemina project (Schriml et al., 2009) and is
more scientific terms.
integrated with DOID.
A specific sub-category of applications is fo-
Besides OBO, other detailed medical ontologies cused on providing drug-administration assistants.
have been created, such as FMA (The Founda- One such example is CareZone. These systems al-
tional Model of Anatomy Ontology) (Rosse and low users to register all drugs on their current treat-
Mejino Jr, 2003) developed by the University of ment scheme, and offer the possibility of entering
830
and keeping track of different medical parameters, 3.1.1 Information Regarding Drugs
such as blood sugar levels. Apart from that, the Both ANMDM structured web information re-
application can also be used to set up reminders garding drugs, as well as medical leaflets obtained
for administering drugs. from ANMDM and private drug producers were
used to create a drugs ontology. A total of 220
3 Method leaflets from Biofarm (Figure 2) and 1138 leaflets
from ANMDM were parsed (Figures 3-4), con-
3.1 Corpus taining information on 15,093 drugs having 1330
different active substances.
There is no established medical or biological on-
tology for Romanian language. However, pub-
lic information is readily available in both struc-
tured and unstructured format. Medical leaflets
are available either as .pdf files or integrated in
web pages made available by both private drug
producers (e.g., Biofarm) and by state authorities
(e.g., The National Agency for Drugs and Medical
Devices - ANMDM). The web page of ANMDM
also contains structured information (e.g., active
substances, concentrations, therapeutical role) for
all approved drugs. Figure 1 contains an overview
of extracted information from the considered data Figure 2: Content extracted from Biofarm leaflets.
sources; specific details are provided in the follow-
up subsections.
831
”acidum ascorbicum” in the Romanian ontology is
equivalent with ”ascorbic acid” in DINTO). This
alignment was done in 2 phases. First, 500 ac-
tive substances were merged because they either
represented perfect matches, or they matched af-
ter applying a small set of conventional changes
(e.g., removing the ”-um” prefix from the Roma-
nian version). Second, the remaining 800 active
substances were matched by analyzing the clos-
est correspondent in the other ontology in terms
of Levenshtein distance (Levenshtein, 1966), fol-
lowed by a manual validation of the match.
832
3.1.3 Connecting Symptoms to Drugs by adding semantic information. For a given set
There are no datasets containing information on of candidate texts from leaflets, each correspond-
what drugs should be administered for certain ing to a different drug, a word2vec (Mikolov et al.,
symptom in Romanian language. However, a us- 2013) model trained on a 1-billion word Romanian
age section exists in each medical leaflet describ- corpus is used to compute the semantic similarity
ing cases in which the drug should be taken (e.g., between it and the given set of symptoms. This is
to numb pain, to reduce coughing, to lower body done by computing aggregate embeddings for the
temperature and eliminate fever, etc.). Thus, the two text representations, and then computing co-
usage section for each of the 1138 leaflets was sine similarity. The Elasticsearch query score and
extracted from the ANMDM website and was in- the cosine similarity are both min-max normalized
dexed into Elasticsearch (Divya and Goyal, 2013), with regard to the set of candidate symptom-drug
allowing users to find drugs with the usage de- pairs extracted with Elasticsearch, their average is
scription that most closely fits the set of provided computed, and the candidate having the best aver-
symptoms. age score is selected in the end.
Extra functionalities are added on top of the
3.2 Architecture and Processing Pipeline Fuseki repository and Elasticsearch server, such as
The knowledge base can be hosted on a single suggesting possible matches when the string rep-
server consisting of two different applications (see resenting a searched drug/disease/symptom was
Figure 7). First, an instance of a Fuseki seman- not found. These suggestions are made by us-
tic repository server (Jena, 2019) was used to ing a Levenshtein edit distance. In the case of
host the aforementioned ontologies: the Roma- symptoms, however, the same sensation can be ex-
nian drug ontology, DINTO, and a merge between pressed in multiple ways; thus, we rely on a very
DOID and SYMP called DOID-merged, to which strict Levenshtein distance search to account for
we have added extra labels for Romanian. This small typing errors. If this fails, a semantic search
repository allows users to query the ontologies via is used, looking for the symptom in the knowledge
the SPARQL query language (Prud’hommeaux base for which the word2vec embedding is closest
and Seaborne, 2008). Second, an Elasticsearch to the embedding of the input text.
instance stores unstructured leaflets and can be
queried in order to find drugs that are most likely 4 Results
helpful in relieving one or more symptoms. 4.1 Drug Information and Drug-Drug
Interactions
Our knowledge base offers access to structured in-
formation regarding the 15,093 drugs and 1,330
active substances. The information regarding
drugs includes both numeric attributes (e.g., time
until expiration), as well as text attributes (e.g.,
usage recommendations which were not standard-
ized as format).
Figure 7: Knowledge base architecture. Users can search for the list of drugs with which
any given one may interact because the Roma-
On top of the two applications that act as infor- nian drug ontology was aligned with DINTO. This
mation sources, a user interface allows users lack- search is done at active substance level. In case
ing experience on semantic web technologies to of drugs containing a combination of active sub-
easily access the information. The user requests stances, we consider that drugs A and B may in-
made at these endpoints are transformed into valid teract if, for at least one active substance from
SPARQL and Elasticsearch queries. Furthermore, A, there is at least one active substance from B
this interface can act as an autocorrect, by suggest- with which it interacts. For example, if the user
ing symptoms and drugs present in our dataset. wants to check the interaction between ”OTOTIS”
In the case of querying for the most helpful which is based on a combination of two active in-
drugs given a set of symptoms, the relevance of gredients (namely ”ciprofloxacinum” and ”fluoci-
the Elasticsearch information retrieval is improved nolonum”) and ”ENAFILZIL” which is based on
833
”sildenafilum”, the knowledge base will search if In most simple use cases, the first entries are rel-
either ”ciprofloxacinum” or ”fluocinolonum” in- evant. However, the symptom of a disease can be,
teract with ”sildenafilum”. As ”ciprofloxacinum” in some cases, the effect of a drug that is targeted
interacts with ”sildenafilum”, the system will con- against a totally different disease. For instance,
clude that the 2 drugs interact. If the same user if the user searches symptoms related to diarrhea,
wants to find the list of all the drugs which interact such as ”scaun moale” (eng, ”loose stool”), some
with ”OTOTIS”, a list of all the drugs containing of the first drugs to be recommended are laxatives,
at least one active substance that interacts with ei- but this would not be the wisest choice of med-
ther of ”ciprofloxacinum” or ”fluocinolonum” is ication. This happens because the effect of the
generated and it will contain 35 entries. medicine is, in some cases, mentioned alongside
with the symptoms it should alleviate.
4.2 Symptom-Disease Information
Our consolidated knowledge base contains 900 5 Conclusions
symptoms and 10,000 diseases from DOID and This paper presents a medical knowledge base
SYMP with names translated into Romanian lan- for Romanian language focused on drugs. It is
guage. The user can enter a list of symptoms to built using medical leaflets in Romanian and struc-
search for the disease that matches most symp- tured information regarding the drugs that was ex-
toms. As mentioned before, if a symptom does tracted from the ANMDM website. The knowl-
not exist, two sequential attempts are performed edge base is also integrated with English on-
to find its closest correspondent in our knowledge tologies in order to make more powerful infer-
base. First, symptoms with a Levenshtein distance ences, such as searching for drug-drug interac-
of 2 or less are searched. Second, if no result is tions. The provided information can be structured
found in the previous step, a word2vec embedding into three main categories: drug related infor-
of the input string is computed, and the symptom mation, disease-symptom information, and drug-
from the knowledge base having the closest em- symptom information. To our knowing, this is the
bedding to it is considered its equivalent. most comprehensive effort of building a knowl-
For example, if a user searches for diseases that edge base for Romanian drugs, their counter-
have ”mic de statură” (eng, ”short stature”) as a indications, as well as potential relations to exhib-
symptom, the results would be ”trichorhinopha- ited conditions.
langeal syndrome type II”, ”Albright’s hered- The drug related information was extracted di-
itary osteodystrophy”, ”spondyloepimetaphyseal rectly from official Romanian sources, thus it can
dysplasia, strudwick type” or ”Renpenning syn- be considered reliable. In order to keep the knowl-
drome”, as these are the only diseases linked to edge base up to date, the sources need to be
that symptom according to DOID and SYMP. If crawled periodically in order to ensure that new
users make 1-2 typos when writing the symptom, information is always added, and that deprecated
the most relevant symptoms is suggested, and they records are eliminated promptly. Information con-
can redo the search with the correct version. If cerning drug-drug interactions is based on the
they enter ”corp mic” (eng, ”small body”) or DINTO ontology, which was last updated in 2016
”scund” (eng, ”short”), the first recommenda- and is still relevant. Nevertheless, we warn users
tion would fail, but the second one would suggest that the information presented by our services is
”short stature” as the most similar symptom based not a valid substitute for the opinion of a medical
on a semantic similarity search. professional or a pharmacist. In the future, apart
from drug-drug interactions, the knowledge base
4.3 Connecting Symptoms to Relevant Drugs could also take into account pre-existing condi-
Considering a search for a combination of ”tuse, tions or dietary choices which may interact with a
febră, durere de cap” (eng, ”coughing, fever, certain treatment scheme. Part of this information
headache”), several types of analgesics, aspirin, is already available in DINTO, but it would need
paracetamol, and 2 types of cough syrup are rec- to be translated and integrated in our knowledge
ommended. The bottom results focus on yellow base.
fever or other diseases that contain only a part of The disease-symptom information is based on
the symptoms specified as input. the DOID and SYMP ontologies. The name of the
834
diseases and symptoms were automatically trans- 7 Acknowledgements
lated using Google Translate, and then manually
The work presented in this paper has been
corrected, if necessary. The two ontologies are
funded by the “Intelligent platform for drugs ad-
still actively maintained; thus, the knowledge base
ministration – PIAM”, subsidiary contract no.
needs to refresh this information from time to time
1267/22.01.2018, from the NETIO project ID:
in order to get the latest version. As is the case
P 40 270, MySMIS Code: 105976.
with the previous category, this information is not
exhaustive and cannot substitute the knowledge of
a professional. References
The drug-symptom information is based only Tim Berners-Lee, James Hendler, and Ora Lassila.
on medical leaflets crawled from the ANMDM 2001. The Semantic Web. Scientific American
website, which needs to be updated every several 284(5):34–43.
weeks. In some cases, the drug-symptom queries Christian Bizer, Jens Lehmann, Georgi Kobilarov,
are very effective, for instance when searching for Sören Auer, Christian Becker, Richard Cyga-
drugs targeting flu-like symptoms, such as fever niak, and Sebastian Hellmann. 2009. DBpe-
and coughing. In other cases, the queries mistake dia - A crystallization point for the Web of
Data. Web Semantics: Science, Services and
the symptoms that the drug should address, with Agents on the World Wide Web 7(3):154–165.
the drug’s effect. These types of mistakes cannot https://doi.org/10.1016/j.websem.2009.07.002.
be avoided for the time being due to manner in
Olivier Bodenreider. 2004. The unified medical lan-
which the information was indexed. If more struc- guage system (umls): integrating biomedical termi-
tured information regarding the drug-symptom re- nology. Nucleic acids research 32(suppl 1):D267–
lation could be extracted from the leaflets, either D270.
manually or by using different NLP techniques,
Manda Sai Divya and Shiv Kumar Goyal. 2013. Elas-
these outlier cases would be addressed. ticsearch: An advanced and quick search technique
Our knowledge base provides real aid for to handle voluminous data. Compusoft 2(6):171.
Romanian users requiring drug-related informa- Janna Hastings, Gareth Owen, Adriano Dekker, Mar-
tion, and no similar initiatives exist at national cus Ennis, Namrata Kale, Venkatesh Muthukrish-
level. The system cannot substitute the knowl- nan, Steve Turner, Neil Swainston, Pedro Mendes,
and Christoph Steinbeck. 2015. Chebi in 2016:
edge of a professional, and there are still prob-
Improved services and an expanding collection of
lems to be addressed, but it is still an easy- metabolites. Nucleic acids research 44(D1):D1214–
to-use and useful tool for informing a user on D1219.
medical treatments. Further improvements will
Marı́a Herrero-Zazo, Isabel Segura-Bedmar, Janna
be explored, including the orientation towards a Hastings, and Paloma Martı́nez. 2015. Dinto: us-
personal health assistant for drug administration, ing owl ontologies and swrl rules to infer drug–drug
similar in some degree to Babylon Health AI interactions and their mechanisms. Journal of chem-
(https://www.babylonhealth.com/ai). ical information and modeling 55(8):1698–1707.
Apache Jena. 2019. Apache jena fuseki documenta-
tion. http://jena.apache.org/documentation/fuseki2/.
6 Future Work
Warren A Kibbe, Cesar Arze, Victor Felix, Elvira
Mitraka, Evan Bolton, Gang Fu, Christopher J
In the future, we aim to expand even further our Mungall, Janos X Binder, James Malone, Drashtti
Vasant, et al. 2014. Disease ontology 2015 update:
knowledge base. This can be done by index-
an expanded and updated database of human dis-
ing medical leaflets from other drug producers, as eases for linking biomedical knowledge through dis-
well as extracting more complex information from ease data. Nucleic acids research 43(D1):D1071–
leaflets - for instance, contraindications expressed D1078.
as rules (e.g. do not take certain antibiotics, such Vladimir I Levenshtein. 1966. Binary codes capable
as tetracycline, with milk, other dairy products, of correcting deletions, insertions, and reversals. In
calcium supplements, or antacids). Furthermore, Soviet physics doklady. volume 10, pages 707–710.
we aim to integrate our knowledge base with other Deborah L McGuinness, Frank Van Harmelen, et al.
information sources, such as the Unified Medical 2004. Owl web ontology language overview. W3C
Language System (Bodenreider, 2004). recommendation 10(10):2004.
835
Tomas Mikolov, Kai Chen, Greg Corrado, and
Jeffrey Dean. 2013. Efficient Estimation of
Word Representations in Vector Space. arXiv
http://arxiv.org/abs/1301.3781.
Eric Prud’hommeaux and Andy Seaborne.
2008. SPARQL Query Language for RDF.
W3C Recommendation 2009(January):1–106.
http://www.w3.org/TR/rdf-sparql-query/.
Cornelius Rosse and José LV Mejino Jr. 2003. A refer-
ence ontology for biomedical informatics: the foun-
dational model of anatomy. Journal of biomedical
informatics 36(6):478–500.
Lynn M Schriml, Cesar Arze, Suvarna Nadendla, Anu
Ganapathy, Victor Felix, Anup Mahurkar, Katherine
Phillippy, Aaron Gussman, Sam Angiuoli, Elodie
Ghedin, et al. 2009. Gemina, genomic meta-
data for infectious agents, a geospatial surveil-
lance pathogen database. Nucleic acids research
38(suppl 1):D754–D764.
Barry Smith, Michael Ashburner, Cornelius Rosse,
Jonathan Bard, William Bug, Werner Ceusters,
Louis J Goldberg, Karen Eilbeck, Amelia Ireland,
Christopher J Mungall, et al. 2007. The obo
foundry: coordinated evolution of ontologies to sup-
port biomedical data integration. Nature biotechnol-
ogy 25(11):1251.
836
Summary Refinement through Denoising
837
Proceedings of Recent Advances in Natural Language Processing, pages 837–843,
Varna, Bulgaria, Sep 2–4 2019.
to a different domain (Isabelle, 2007). 3.2 Types of Noise
In (Xie et al., 2018), authors synthesize gram- We experiment with three simple types of noise,
matically incorrect sentences from correct ones all of which introduce information redundancy
using backtranslation (Sennrich et al., 2016a), into a summary. Our aim is to train denoising
which they use for grammar error correction. They models that minimize repetitive or peripheral in-
enforce hypothesis variety during decoding by formation within summaries.
adding noise to beam search. Another work that
is close to ours is (Fevry and Phang, 2018), where Repeat picks random sentences from the sum-
authors introduce redundancy on the word level in mary and repeats them at the end. Repetition
order to build an unsupervised sentence compres- of phrases or even whole sentences is a problem
sion system. In this work, we take a similar ap- commonly observed in text generation with RNNs
proach, but instead focus on generating informa- (See et al., 2017), which motivates efforts to detect
tion redundancy errors on the sentence rather than and minimize repetitions.
the word level. Replace picks random sentences from the sum-
mary, and replaces them with the closest sentence
3 Approach from the article. This type of noise helps the model
Our approach to summary refinement consists of to learn to refine sentences from the generated
two steps. First, we use a dataset of clean ground summaries, paraphrasing sentences when they are
truth summaries to generate noisy summaries us- too long or contain redundant information.
ing several different types of synthetic noise. Sec- Extra picks random sentences from the article,
ond, we train text rewriting models to correct and paraphrases them, and inserts them into the sum-
denoise the noisy summaries, restoring them to mary, preserving the order of the sentences as
their original form. The learned denoising models they appear in the article. With this type of
are then used to post-process and refine the outputs noise, a model learns to delete sentences which
of a summarization system. are out of context or contain redundant informa-
tion. To paraphrase the sentences, we use the sen-
3.1 Generating Noisy Summaries
tence paraphrasing model from (Chen and Bansal,
To generate noisy datasets, we rely on an exist- 2018), trained on matching sentence pairs from the
ing parallel dataset of articles and clean ground CNN/Daily Mail dataset.
truth summaries S = {s0 , , ..., sj }. We iterate
over each of the summaries and perturb them with Mixture mixes all the above noise types uni-
noise, according to a sentence noise distribution formly into a single dataset, keeping the same
pnoise = [p0 , p1 , ..., pN ]. pnoise defines the proba- dataset size as for the individual noise types. With
bility of adding noise to a specific number of sen- mixture, we explore whether the benefits of each
tences within each summary (fromP 0 up to a max- noise type can be combined into a single model.
imum of N noisy sentences), with pnoise = 1.
4 Experimental Setup
For all experiments in this work, we use
pnoise = [0.15, 0.85] in order to ensure consis- Dataset We use the CNN/Daily Mail dataset2
tency, meaning that ˜15% of our noisy summaries (Hermann et al., 2015) of news articles and sum-
contain no noisy sentences, while ˜85% contain maries in the form of bullet points, and follow
one noisy sentence. Initial experiments showed the preprocessing pipeline from (Chen and Bansal,
that distributions which enforce larger or smaller 2018). We use the standard split of the dataset,
amounts of noise lead to stronger or weaker de- consisting of 287k news-summary pairs for train-
noising effects. Our choice of noise distribution ing and 13k pairs for validation. We follow Sec-
showed good results on the majority of systems tion 3.1 to generate noisy versions of the datasets
that we tested; we leave a more rigorous investiga- to be used during training. During testing, instead
tion of the choice of distribution to future work. of clean summaries that contain noisy sentences,
In addition to adding noise, we generate 3 noisy we input summaries produced by existing extrac-
summaries for each clean summary by picking tive or abstractive summarization systems.
multiple random sentences to noise. This step in- 2
https://github.com/abisee/
creases the dataset size while introducing variety. cnn-dailymail
838
13 18
34 39
12 17
37
Rouge-1
Rouge-2
Rouge-1
Rouge-2
32 16
11
35
30 10 15
33
28 9 31 14
2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6
50 36 50
30
40 34 40
Repeat rate
Repeat rate
Rouge-L
Rouge-L
28 30 no denoising 30 no denoising
overlap 32 overlap
replace replace
26 20 repeat 20 repeat
extra 30 extra
mixture mixture
10 human 10 human
24 28
2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6
(a) Denoising the LexRank system. (b) Denoising the RNN-Ext system.
Figure 2: Metric results (Rouge-1/2/L and Repeat rate) on denoising extractive summarization systems. The x-axis in all
plots is the number of extracted sentences. human is the result of the ground truth summaries (only for the Repeat rate).
839
System Denoising approach ROUGE-1 ROUGE-2 ROUGE-L Repeat #Sent #Tok
Human - - - - 28.86 3.88 61.21
Article - 14.95 8.54 14.41 70.5 26.9 804
Article Mixture 30.47 13.97 28.24 53.43 10.67 304.7
RNN - 35.61 15.04 32.7 51.9 2.93 58.46
RNN Overlap 36.41 15.92 33.73 26.84 2.39 47.31
RNN Repeat 36.5 15.94 33.79 27.65 2.41 48.34
RNN Replace 35.2 14.86 32.4 51.51 2.98 57.0
RNN Extra 33.95 14.58 31.2 37.19 2.21 42.82
RNN Mixture 35.08 15.3 32.44 27.27 2.2 42.14
RNN-RL - 40.88 17.8 38.54 39.29 4.93 72.82
RNN-RL Overlap 40.76 17.69 38.43 37.71 4.83 71.02
RNN-RL Repeat 40.84 17.76 38.49 38.78 4.86 71.69
RNN-RL Replace 40.78 17.72 38.46 39.24 4.93 72.2
RNN-RL Extra 39.12 16.7 36.76 34.04 3.84 55.43
RNN-RL Mixture 40.11 17.33 37.76 35.45 4.18 61.15
Table 1: Results on denoising abstractive summarization. Repeat is the Repeat rate, while #Sent and #Tok are the average
numbers of sentences or tokens in the summaries. Best ROUGE results for each model are in bold. Human is the result of the
ground truth summaries, while Article uses the original article as the summary.
tional denoising operations learned by our models Our metric results from denoising abstractive
(see Figure 4a) are beneficial and can lead to more summarization are in Table 1. In Figure 3, we also
polished summaries that also may contain abstrac- compute the approximate number of sentence rep-
tive elements. etitions on the test set, by calculating the number
The gains from denoising are greater for longer of sentences that overlap significantly (> 80%)
summaries of more than two sentences. Long with at least one other sentence in the summary.
summaries are more likely to be affected by redun- For the RNN model, the repeat noise helps
dancy. For shorter summaries, denoising might to remove repetition, halving our repetition met-
lead to deletion of important information, thus de- ric, while boosting the ROUGE scores. This re-
noising needs to be applied more carefully in such sult is similar to our much simpler overlap
cases. Furthermore, for all sentence lengths and baseline based on sentence deletion. The other
noise types, we observe a reduction in the Re- noise types help to reduce redundancy, bringing
peat rate after denoising, demonstrating that our the Repeat rate closer to that of Human sum-
approach is effective at reducing redundancy. maries. This, however, comes at the cost of a
In Table 1, we additionally include the result decrease in ROUGE. For RNN-RL, while denois-
from using the whole articles (Article) as input ing helps to reduce repetition, none of our noise
to our mixture model. Denoising is effective types managed to yield ROUGE improvements.
in this case, indicating that our approach may be One reason for this may be that this model al-
promising for developing abstractive summariza- ready comes with a built-in mechanism for reduc-
tion systems that are fully unsupervised, similar to ing redundancy which relies on sentence rerank-
recent work in unsupervised sentence compression ing (Chen and Bansal, 2018). However, as shown
(Fevry and Phang, 2018). in Figure 3 (and in our example in Table 2), this
model still generates many more sentence repeti-
5.2 Abstractive Summarization tions than found in human summaries. In over-
For abstractive summarization, we test two sys- all, our approach is effective at reducing redun-
tems. The first is a standard LSTM encoder- dant information in abstractive summaries, how-
840
extra repeat extra repeat
(a) RNN-Ext extractive system, extracting 5 sentences. (b) RNN abstractive system.
Figure 4: Types of denoising operations applied to an extractive (left) and an abstractive (right) system (averaged over our
test set).
ever this comes with a potential loss of informa- system, Replace noise, where ”dinorah santana ,
tion, which can lead to a reduction in ROUGE. the player s agent , said her client had rejected
Thus, our denoising methods are currently better the offer of a three-year contract extension” is
suited for extractive than for absctractive summa- paraphrased to ”the player s agent said she had
rization. Our work therefore calls for the develop- rejected the offer of a three-year contract”), or
ment of novel types of synthetic noise that target even a combination of deletion and rewriting (e.g.
abstractive summarization. RNN-RL system, Repeat noise).
841
Ground truth (Rep=38.38):
1. dani alves has spent seven seasons with the catalan giants
2. alves has four spanish titles to his name with barcelona
3. the brazil defender has also won the champions league twice with barca
RNN-Ext-4 RNN RNN-RL
No denoising (R-1=33.6,Rep=45): No denoising (R-1=34,Rep=79.6): No denoising (R-1=31,Rep=51.6):
1. dani alves looks set to leave 1. dani alves has been unable to agree 1. dani alves looks set to leave
barcelona this summer after his repre- a new deal with catalan club barcelona this summer
sentative confirmed the brazilian right- 2. the brazilian has been unable to 2. alves has enjoyed seven successful
back had rejected the club ’s final con- agree a new deal with catalan club years at barcelona
tract offer 3. alves has been unable to agree a 3. alves has been unable to agree a
2. alves has enjoyed seven successful new deal with catalan club deal with the catalan club
years at barcelona , winning four span- 4. alves has been linked with a num- 4. the 31-year-old has been unable to
ish titles and the champions league ber of clubs including manchester agree a new deal
twice united and manchester city 5. dinorah santana , the player ’s agent
3. but the 31-year-old has been unable , said her client had rejected the offer
to agree a new deal with the catalan of a three-year contract extension
club and will leave the nou camp this
summer
4. dinorah santana , the player ’s agent
and ex-wife , said at a press confer-
ence on thursday that her client had
rejected the offer of a three-year con-
tract extension , which was dependent
on the player taking part in 60 per cent
of matches for the club
Replace (R-1=36.6,Rep=46.6): Replace (R-1=34, Rep=79.6): Replace (R-1=31, Rep=52.6):
1. Same 1. Same 1. Same
2. Same 2. Same 2. Same
3. Same 3. Same 3. Same
4. the player ’s agent and ex-wife said 4. Same 4. Same
at a press conference on thursday that 5. the player ’s agent said she had re-
her client had rejected the offer of a jected the offer of a three-year contract
three-year contract extension
Repeat (R-1=33.6,Rep=45): Repeat (R-1=28, Rep=41.4): Repeat (R-1=24.2, Rep=36.1):
1. Same 1. Same 1. Same
2. Same 2. Deleted 2. Same
3. Same 3. Deleted 3. Deleted
4. Same 4. Same 4. alves has been unable to agree a
new deal
5. Same
Extra (R-1=43.6,Rep=36.8): Extra (R-1=37.2,Rep=92.8): Extra (R-1=37, Rep=60.8):
1. Same 1. Same 1. Same
2. Same 2. Same 2. Same
3. the 31-year-old has been unable to 3. Same 3. Same
agree a new deal with the catalan club 4. Deleted 4. Same
and will leave the nou camp this sum- 5. Deleted
mer
4. Deleted
Mixture (R-1=43, Rep=36.2): Mixture (R-1=28, Rep=41.43): Mixture (R-1=37, Rep=60.8):
1. Same 1. Same 1. Same
2. Same 2. Deleted 2. Same
3. Same 3. Deleted 3. Same
4. Deleted 4. Same 4. Same
5. Deleted
Table 2: Examples for denoising extractive and abstractive summarization. Same indicates a summary sentence has been
unchanged, while Deleted indicates sentence deletion. In brackets, R-1 denotes the Rouge-1 score, while Rep denotes the
Repeat rate.
842
Acknowledgments Shashi Narayan, Shay B. Cohen, and Mirella Lap-
ata. 2018. Ranking sentences for extractive sum-
We acknowledge support from the Swiss National marization with reinforcement learning. In Pro-
Science Foundation (grant 31003A 156976). ceedings of the 2018 Conference of the North
American Chapter of the Association for Com-
putational Linguistics: Human Language Tech-
References nologies, Volume 1 (Long Papers). Association
for Computational Linguistics, pages 1747–1759.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- https://doi.org/10.18653/v1/N18-1158.
gio. 2014. Neural machine translation by jointly
learning to align and translate. arXiv preprint Ani Nenkova, Sameer Maskey, and Yang Liu. 2011.
arXiv:1409.0473 . Automatic summarization. In Proceedings of the
49th Annual Meeting of the Association for Com-
Rajen Chatterjee, Matteo Negri, Raphael Rubino, and putational Linguistics: Tutorial Abstracts of ACL
Marco Turchi. 2018. Findings of the wmt 2018 2011. Association for Computational Linguistics,
shared task on automatic post-editing. In Proceed- page 3.
ings of the Third Conference on Machine Transla-
tion: Shared Task Papers. pages 710–725. Nikola Nikolov, Michael Pfeiffer, and Richard Hahn-
loser. 2018. Data-driven summarization of scien-
Yen-Chun Chen and Mohit Bansal. 2018. Fast abstrac- tific articles. In Proceedings of the Eleventh Interna-
tive summarization with reinforce-selected sentence tional Conference on Language Resources and Eval-
rewriting. In Proceedings of ACL. uation (LREC 2018). European Language Resources
Association (ELRA), Paris, France.
Günes Erkan and Dragomir R Radev. 2004. Lexrank:
Graph-based lexical centrality as salience in text Alla Rozovskaya and Dan Roth. 2016. Grammatical
summarization. Journal of artificial intelligence re- error correction: Machine translation and classifiers.
search 22:457–479. In Proceedings of the 54th Annual Meeting of the
Association for Computational Linguistics (Volume
Thibault Fevry and Jason Phang. 2018. Unsuper- 1: Long Papers). volume 1, pages 2205–2215.
vised sentence compression using denoising auto-
encoders. In Proceedings of the 22nd Conference Abigail See, Peter J Liu, and Christopher D Manning.
on Computational Natural Language Learning. As- 2017. Get to the point: Summarization with pointer-
sociation for Computational Linguistics, pages 413– generator networks. In Proceedings of the 55th An-
422. http://aclweb.org/anthology/K18-1040. nual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers). volume 1,
Sebastian Gehrmann, Yuntian Deng, and Alexander pages 1073–1083.
Rush. 2018. Bottom-up abstractive summarization.
In Proceedings of the 2018 Conference on Empiri- Rico Sennrich, Barry Haddow, and Alexandra Birch.
cal Methods in Natural Language Processing. Asso- 2016a. Improving neural machine translation mod-
ciation for Computational Linguistics, pages 4098– els with monolingual data. In Proc. of ACL. Asso-
4109. http://aclweb.org/anthology/D18-1443. ciation for Computational Linguistics, pages 86–96.
https://doi.org/10.18653/v1/P16-1009.
Karl Moritz Hermann, Tomas Kocisky, Edward
Grefenstette, Lasse Espeholt, Will Kay, Mustafa Su- Rico Sennrich, Barry Haddow, and Alexandra Birch.
leyman, and Phil Blunsom. 2015. Teaching ma- 2016b. Neural machine translation of rare words
chines to read and comprehend. In Advances in Neu- with subword units. In Proc. of ACL. Association
ral Information Processing Systems. pages 1693– for Computational Linguistics, pages 1715–1725.
1701. https://doi.org/10.18653/v1/P16-1162.
P Isabelle. 2007. Domain adaptation of mt systems Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. 2014.
through automatic postediting. Proc. 10th Machine Sequence to sequence learning with neural net-
Translation Summit (MT Summit XI), 2007 . works. In Advances in neural information process-
ing systems. pages 3104–3112.
Diederick P Kingma and Jimmy Ba. 2015. Adam: A
method for stochastic optimization. In International Ziang Xie, Guillaume Genthial, Stanley Xie, Andrew
Conference on Learning Representations (ICLR). Ng, and Dan Jurafsky. 2018. Noising and denoising
natural language: Diverse backtranslation for gram-
Chin-Yew Lin. 2004. Rouge: A package for auto- mar correction. In Proceedings of the 2018 Confer-
matic evaluation of summaries. Text Summarization ence of the North American Chapter of the Associ-
Branches Out . ation for Computational Linguistics: Human Lan-
Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. guage Technologies, Volume 1 (Long Papers). As-
Summarunner: A recurrent neural network based se- sociation for Computational Linguistics, pages 619–
quence model for extractive summarization of docu- 628. https://doi.org/10.18653/v1/N18-1057.
ments. In Thirty-First AAAI Conference on Artificial
Intelligence.
843
Large-Scale Hierarchical Alignment for Data-driven Text Rewriting
Document alignment
Abstract Sentence alignment
Source Target
Target 1 Target 2
We propose a simple unsupervised method Source
document document
document
for extracting pseudo-parallel monolin-
gual sentence pairs from comparable cor- Target 5
quire a seed parallel corpus, but instead Figure 1: Illustration of large-scale hierarchical align-
relies solely on hierarchical search over ment (LHA). For each document in a source dataset, doc-
ument alignment retrieves matching documents from a tar-
pre-trained embeddings of documents and get dataset. In turn, sentence alignment retrieves matching
sentences. We demonstrate the effective- sentence pairs from within each document pair.
ness of our method through automatic and
extrinsic evaluation on text simplification and Langlais (2018)) assume access to some par-
from the normal to the Simple Wikipedia. allel training data. This impedes their application
We show that pseudo-parallel sentences to cases where there is no parallel data available
extracted with our method not only sup- whatsoever, which is the case for the majority of
plement existing parallel data, but can text rewriting tasks, such as style transfer.
even lead to competitive performance on In this paper, we propose a simple unsuper-
their own.1 vised method, Large-scale Hierarchical Align-
ment (LHA) (Figure 1; Section 3), for extract-
1 Introduction ing pseudo-parallel sentence pairs from two raw
Parallel corpora are indispensable resources for monolingual corpora which contain documents
advancing monolingual and multilingual text in two different author styles, such as scientific
rewriting tasks. Due to the scarce availability of papers and press releases. LHA hierarchically
parallel corpora, and the cost of manual creation, searches for document and sentence nearest neigh-
a number of methods have been proposed that bors within the two corpora, extracting sentence
can perform large-scale sentence alignment: auto- pairs that have high semantic similarity, yet pre-
matic extraction of pseudo-parallel sentence pairs serve the stylistic characteristics representative of
from raw, comparable2 corpora. While pseudo- their original datasets. LHA is robust to noise,
parallel data is beneficial for machine translation fast and memory efficient, enabling its application
(Munteanu and Marcu, 2005), there has been little to datasets on the order of hundreds of millions
work on large-scale sentence alignment for mono- of sentences. Its generality makes it relevant to a
lingual text-to-text rewriting tasks, such as simpli- wide range of monolingual text rewriting tasks.
fication (Nisioi et al., 2017) or style transfer (Liu We demonstrate the effectiveness of LHA on
et al., 2016). Furthermore, the majority of existing automatic benchmarks for alignment (Section 4),
methods (e.g. Marie and Fujita (2017); Grégoire as well as extrinsically, by training neural ma-
1 chine translation (NMT) systems on the task of
Code available at https://github.com/
ninikolov/lha. text simplification from the normal Wikipedia to
2
Corpora that contain documents on similar topics. the Simple Wikipedia (Section 5). We show that
844
Proceedings of Recent Advances in Natural Language Processing, pages 844–853,
Varna, Bulgaria, Sep 2–4 2019.
pseudo-parallel datasets obtained by LHA are not documents in the corpora are parallel (labelled
only useful for augmenting existing parallel data, document pairs, such as identical articles in two
boosting the performance on automatic measures, languages), the task is to identify suitable sen-
but can even be competitive on their own. tence pairs from each document. This problem
has been extensively studied both in the multi-
2 Background lingual (Brown et al., 1991; Moore, 2002) and
monolingual (Hwang et al., 2015; Kajiwara and
2.1 Data-Driven Text Rewriting
Komachi, 2016; Štajner et al., 2018) case. The
The goal of text rewriting is to transform an input limited availability of parallel corpora led to the
text to satisfy specific constraints, such as simplic- development of large-scale sentence alignment
ity (Nisioi et al., 2017) or a more general author methods, which is also the focus of this work. The
style, such as political (e.g. democratic to republi- aim of these methods is to extract pseudo-parallel
can) or gender (e.g. male to female) (Prabhumoye sentence pairs from raw, non-aligned corpora. For
et al., 2018; Shen et al., 2017). Rewriting systems many tasks, millions of examples occur naturally
can be valuable when preparing a text for multi- within existing textual resources, amply available
ple audiences, such as simplification for language on the internet.
learners (Siddharthan, 2002) or people with read- The majority of previous work on large-scale
ing disabilities (Inui et al., 2003). They can also be sentence alignment is in machine translation,
used to improve the accessibility of technical doc- where adding pseudo-parallel pairs to an existing
uments, e.g. to simplify terms in clinical records parallel dataset has been shown to boost the trans-
for laymen (Abrahamsson et al., 2014). lation performance (Munteanu and Marcu, 2005;
Text rewriting can be cast as a data-driven task Uszkoreit et al., 2010). The work that is most
in which transformations are learned from large closely related to ours is (Marie and Fujita, 2017),
collections of parallel sentences. Limited avail- where authors use pre-trained word and sentence
ability of high-quality parallel data is a major embeddings to extract rough translation pairs in
bottleneck for this approach. Recent work on two languages. Subsequently, they filter out low-
Wikipedia and the Simple Wikipedia (Coster and quality translations using a classifier trained on
Kauchak, 2011; Kajiwara and Komachi, 2016) and parallel translation data. More recently, (Grégoire
on the Newsela dataset of simplified news arti- and Langlais, 2018) extract pseudo-parallel trans-
cles for children (Xu et al., 2015) explore super- lation pairs using a Recurrent Neural Network
vised, data-driven approaches to text simplifica- (RNN) classifier. Importantly, these methods as-
tion. Such approaches typically rely on statisti- sume that some parallel training data is already
cal (Xu et al., 2016) or neural (Štajner and Nisioi, available, which impedes their application in set-
2018) machine translation. tings where there is no parallel data whatsoever,
Recent work on unsupervised approaches to text which is the case for many text rewriting tasks
rewriting without parallel corpora is based on vari- such as style transfer.
ational (Fu et al., 2017) or cross-aligned (Shen There is little work on large-scale sentence
et al., 2017) autoencoders that learn latent repre- alignment focusing specifically on monolingual
sentations of content separate from style. In (Prab- tasks. In (Barzilay and Elhadad, 2003), authors
humoye et al., 2018), authors model style trans- develop a hierarchical alignment approach of first
fer as a back-translation task by translating input clustering paragraphs on similar topics before per-
sentences into an intermediate language. They use forming alignment on the sentence level. They ar-
the translations to train separate English decoders gue that, for monolingual data, pre-clustering of
for each target style by combining the decoder loss larger textual units is more robust to noise com-
with the loss of a style classifier, separately trained pared to fine-grained sentence matching applied
to distinguish between the target styles. directly on the dataset level.
2.2 Large-Scale Sentence Alignment
3 Large-Scale Hierarchical Alignment
The goal of sentence alignment is to extract from (LHA)
raw corpora sentence pairs suitable as training ex-
amples for text-to-text rewriting tasks such as ma- Given two datasets that contain comparable doc-
chine translation or text simplification. When the uments written in two different author styles: a
845
source dataset Sd consisting of NS documents get sentence. We denote the nearest neighbours
Sd = {sd1 , ..., sdNS } (e.g. all Wikipedia articles) of ssi as N N (ssi ) = {tsi1 , . . . , tsiK } and the near-
and a target dataset Td consisting of NT docu- est neighbours of tsj as N N (tsj ) = {ssj1 , . . . , ssjK }.
ments Td = {td1 , ..., tdNT } (e.g. all articles from We remove all sentence pairs with similarity be-
the Simple Wikipedia), our approach to large-scale low a manually set threshold θs . We then merge all
alignment is hierarchical, consisting of two con- overlapping sets of nearest sentences in the doc-
secutive steps: document alignment followed by uments to produce pseudo-parallel sentence sets
sentence alignment (see Figure 1). (e.g. ({sse , ssi }, {tsj , tsk , tsl }) when source sentence
i is closes to target sentences j, k, and l and target
3.1 Document Alignment sentence j is closest to source sentences e and i).
For each source document sdi , document align- This approach, inspired from (Štajner et al., 2018),
ment retrieves K nearest neighbours {tdi1 , ..., tdiK } provides the flexibility to model multi-sentence in-
from the target dataset. In combination, teractions, such as sentence splitting or compres-
these form K pseudo-parallel document pairs sion, as well as individual sentence-to-sentence re-
{(sdi , tdi1 ), ..., (sdi , tdiK )}. Our aim is to select doc- formulations. Note that when K = 1, we only
ument pairs with high semantic similarity, po- retrieve individual sentence pairs.
tentially containing good pseudo-parallel sentence The final output of sentence alignment is a list
pairs representative of the document styles of each of pseudo-parallel sentence pairs with high seman-
dataset. tic similarity and preserved stylistic characteristics
To find nearest neighbours, we rely on two of each dataset. The pseudo-parallel pairs can be
components: document embedding and approx- used to either augment an existing parallel dataset
imate nearest neighbour search. For each (as in Section 5), or independently, to solve a new
dataset, we pre-compute document embeddings author style transfer task for which there is no par-
ed () as Is = [ed (sd1 ), ..., ed (sdNS )] and It = allel data available (see the supplementary mate-
[ed (td1 ), ..., ed (tdNT )]. We employ nearest neigh- rial for an example).
bour search methods3 to partition the embedding 3.3 System Variants
space, enabling fast and efficient nearest neigh-
The aforementioned framework provides the flexi-
bour retrieval of similar documents across Is and
bility of exploring diverse variants, by exchanging
It . This enables us to find K nearest target docu-
document/sentence embeddings or text similarity
ment embeddings in It for each source embedding
metrics. We compare all variants in an automatic
in Is . We additionally filter document pairs whose
evaluation in Section 4.
similarity is below a manually selected threshold
θd . In Section 4, we evaluate a range of different Text embeddings We experiment with four text
document embedding approaches, as well as alter- embedding methods:
native similarity metrics.
1. Avg, is the average of the constituent word
3.2 Sentence Alignment embeddings of a text4 , a simple approach that
has proved to be a strong baseline for many
Given a pseudo-parallel document pair
text similarity tasks.
(sd , td ) that contains a source document
sd = {ss1 , ..., ssNJ } consisting of NJ sen- 2. In Sent2Vec5 (Pagliardini et al., 2018), the
tences and a target document td = {ts1 , ..., tsNM } word embeddings are specifically optimized
consisting of NM sentences, sentence alignment towards additive combinations over the sen-
extracts pseudo-parallel sentence pairs (ssi , tsj ) tence using an unsupervised objective func-
that are highly similar. tion. This approach performs well on many
To implement sentence alignment, we first em- unsupervised and supervised text similarity
bed each sentence in sd and td and compute an tasks, often outperforming more sophisti-
inter-sentence similarity matrix P among all sen- cated supervised recurrent or convolutional
tence pairs in sd and td . From P we extract K architectures, while remaining very fast to
nearest neighbours for each source and each tar- compute.
3 4
We use the Annoy library https://github.com/ We use the Google News 300-dim Word2Vec models.
5
spotify/annoy. We use the public unigram Wikipedia model.
846
3. InferSent6 (Conneau et al., 2017) is a super- number of false positives. To evaluate document
vised sentence embedding approach based on alignment, we add 1000 randomly sampled arti-
bidirectional LSTMs, trained on natural lan- cles from Wikipedia and the Simple Wikipedia as
guage inference data. noise, resulting in 1046 article pairs in total. The
goal of document alignment is to identify the orig-
4. BERT 7 (Devlin et al., 2019) is a state-of-the- inal 46 document pairs out of 1046×1046 possible
art supervised sentence embedding approach document combinations.
based on the Transformer architecture.
This set-up additionally enables us to jointly
Word Similarity We additionally test four evaluate document and sentence alignment, which
word-based approaches for computing text simi- best resembles the target effort of retrieving good
larity. Those can be used either on their own, or sentence pairs from noisy documents. The two
to refine the nearest neighbour search across doc- aims of the joint alignment task are to identify the
uments or sentences. good sentence pairs from within either 1M doc-
ument or 125M sentence pairs, in the latter case
1. We compute the unigram string overlap without relying on any document-level informa-
o(x, y) = |{y}∩{x}|
|{y}| between source tokens x tion whatsoever.
and target tokens y (excluding punctuation,
numbers and stopwords). 4.1 Results
2. We use the BM25 ranking function (Robert- Our results are summarized in Tables 1 and 2.
son et al., 2009), an extension of TF-IDF. For all experiments, we set K = 1 and report
the maximum F1 score (F1max ) obtained from
3. We use the Word Mover’s Distance (WMD) varying the document threshold θd and the sen-
(Kusner et al., 2015), which measures the tence threshold θs . We also report the percentage
distance the embedded words of one docu- of true positive (TP) document or sentence pairs
ment need to travel to reach the embedded that were retrieved when the F1 score was at its
words of another document. WMD has re- maximum, as well as the average speed of each
cently achieved good results on text retrieval approach (doc/s and sent/s). The speed becomes
(Kusner et al., 2015) and sentence alignment of a particular concern when working with large
(Kajiwara and Komachi, 2016). datasets consisting of millions of documents and
hundreds of millions of sentences.
4. We use the Relaxed Word Mover’s Distance
On document alignment, (Table 1, left) the
(RWMD) (Kusner et al., 2015), which is a fast
Sent2Vec approach achieved the best score, outper-
approximation of the WMD.
forming the other embedding methods including
4 Automatic Evaluation the word-based similarity measures. On sentence
alignment (Table 1, right), the WMD achieves the
We perform an automatic evaluation of LHA using best performance, matching the result from (Ka-
an annotated sentence alignment dataset (Hwang jiwara and Komachi, 2016). When evaluating
et al., 2015). The dataset contains 46 article pairs document and sentence alignment jointly (Table
from Wikipedia and the Simple Wikipedia. The 2), we compare our hierarchical approach (LHA)
67k potential sentence pairs were manually la- to global alignment applied directly on the sen-
belled as either good simplifications (277 pairs), tence level (Global). Global computes the simi-
good with a partial overlap (281 pairs), par- larities between all 125M sentence pairs in the en-
tial (117 pairs) or non-valid. We perform three tire evaluation dataset. LHA significantly outper-
comparisons using this dataset: evaluating docu- forms Global, successfully retrieving three times
ment and sentence alignment separately, as well more valid sentence pairs, while remaining fast to
as jointly. compute. This result demonstrates that document
For sentence alignment, the task is to retrieve alignment is beneficial, successfully filtering some
the 277 good sentence pairs out of the 67k possi- of the noise, while also reducing the overall num-
ble sentence pairs in total, while minimizing the ber of sentence similarities to be computed.
6
We use the GloVe-based model provided by the authors. The Sent2Vec approach to LHA achieves good
7
We use the base 12-layer model provided by the authors. performance on document and sentence align-
847
Table 1: Automatic evaluation of Document (left) and Sentence alignment (right). EDim is the embedding dimensionality.
TP is the percentage of true positives obtained at F1max . Speed is calculated on a single CPU thread.
Document alignment Sentence alignment
Approach EDim F1max TP θd doc/s F1max TP θs sent/s
Average word embeddings (Avg) 300 0.66 43% 0.69 260 0.675 46% 0.82 1458
Word sim Embedding
Sent2Vec (Pagliardini et al., 2018) 600 0.78 61% 0.62 343 0.692 48% 0.69 1710
InferSent† (Conneau et al., 2017) 4096 - - - - 0.69 49% 0.88 110
BERT† (Devlin et al., 2019) 768 - - - - 0.65 43% 0.89 25
Overlap - 0.53 29% 0.66 120 0.63 40% 0.5 1600
BM25 (Robertson et al., 2009) - 0.46 16% 0.257 60 0.52 27% 0.43 20K
RWMD (Kusner et al., 2015) 300 0.713 51% 0.67 60 0.704 50% 0.379 1050
WMD (Kusner et al., 2015) 300 0.49 24% 0.3 1.5 0.726 54% 0.353 180
(Hwang et al., 2015) - - - - - 0.712 - - -
(Kajiwara and Komachi, 2016) - - - - - 0.724 - - -
†: These models are specifically designed for sentence embedding, hence we do not test them on document alignment.
Table 2: Evaluation on large-scale sentence alignment: et al., 2015) with an attention mechanism (Bah-
identifying the good sentence pairs without any document- danau et al., 2015). We train our models on the
level information. We pre-compute the embeddings and use
the Annoy ANN library. For the WMD-based approaches, subword level (Sennrich et al., 2015), capping the
we re-compute the top 50 sentence nearest neighbours of vocabulary size to 50k. We re-learn the subword
Sent2Vec.
rules separately for each dataset, and train until
Approach F1max TP time
LHA (Sent2Vec) 0.54 31% 33s
convergence using the Adam optimizer (Kingma
LHA (Sent2Vec + WMD) 0.57 33% 1m45s and Ba, 2015). We use beam search with a beam
Global (Sent2Vec) 0.339 12% 15s of 5 to generate all final outputs.
Global (WMD) 0.291 12% 30m45s
Evaluation metrics We report a diverse range of
ment, while also being the fastest to compute. We
automatic metrics and statistics. SARI (Xu et al.,
therefore use it as the default approach for the fol-
2016) is a recently proposed metric for text sim-
lowing experiments on text simplification.
plification which correlates well with simplicity
in the output. SARI takes into account the total
5 Empirical Evaluation
number of changes (additions, deletions) of the in-
To test the suitability of pseudo-parallel data ex- put when scoring model outputs. BLEU (Papineni
tracted with LHA, we perform empirical exper- et al., 2002) is a precision-based metric for ma-
iments on text simplification from the normal chine translation commonly used for evaluation of
Wikipedia to the Simple Wikipedia. We chose text simplification (Xu et al., 2016; Štajner and Ni-
simplification because some parallel data are al- sioi, 2018) and of style transfer (Shen et al., 2017).
ready available for this task, allowing us to ex- Recent work has indicated that BLEU is not suit-
periment with mixing parallel and pseudo-parallel able for assessment of simplicity (Sulem et al.,
datasets. In the supplementary material8 we exper- 2018), it correlates better with meaning preserva-
iment with an additional task for which there is no tion and grammaticality, in particular when using
parallel data: style transfer from scientific journal multiple references. We also report the average
articles to press releases. Levenshtein distance (LD) from the model out-
We compare the performance of neural machine puts to the input (LDsrc ) or the target reference
translation (NMT) systems trained under three dif- (LDtgt ). On simplification tasks, LD correlates
ferent scenarios: 1) using existing parallel data well with meaning preservation and grammatical-
for training; 2) using a mixture of parallel and ity (Sulem et al., 2018), complementing BLEU.
pseudo-parallel data extracted with LHA; and 3)
using pseudo-parallel data on its own. Extracting pseudo-parallel data We use LHA
with Sent2Vec (see Section 3) to extract pseudo-
5.1 Experimental Setup parallel sentence pairs for text simplification. To
ensure some degree of lexical similarity, we ex-
NMT model For all experiments, we use a clude pairs whose string overlap (defined in Sec-
single-layer LSTM encoder-decoder model (Cho tion 3.3) is below 0.4, and pairs in which the tar-
8
Available in our arXiv paper at https://arxiv. get sentence is more than 1.5 times longer than the
org/abs/1810.08237 source sentence. We use K = 5 in all of our align-
848
Table 3: Datasets used to extract pseudo-parallel monolingual sentence pairs in our experiments.
Dataset Type Documents Tokens Sentences Tok. per sent. Sent. per doc.
Wikipedia Articles 5.5M 2.2B 92M 25 ± 16 17 ± 32
Simple Wikipedia Articles 134K 62M 2.9M 27 ± 68 22 ± 34
Gigaword News 8.6M 2.5B 91M 28 ± 12 11 ± 7
Table 4: Example pseudo-parallel pairs extracted by our Large-scale hierarchical alignment (LHA) method.
Dataset Source Target
wiki- However, Jimmy Wales, Wikipedia’s co-founder, de- But the co-founder Wikipedia, Jimmy Wales, did not
simp- nied that this was a crisis or that Wikipedia was run- believe that this was a crisis. He also did not believe
65 ning out of admins, saying, ”The number of admins Wikipedia was running out of admins.
has been stable for about two years, there’s really
nothing going on.”
wiki- Prior to World War II, Japan’s industrialized econ- Until Japan ’s defeat in World War II , the economy
news- omy was dominated by four major zaibatsu: Mit- was dominated by four conglomerates , known as “
74 subishi, Sumitomo, Yasuda and Mitsui. zaibatsu ” in Japanese . These were the Mitsui , Mit-
subishi , Sumitomo and Yasuda groups .
Table 5: Statistics of the pseudo-parallel datasets extracted Wikipedia using document and sentence sim-
tgt
with LHA. µsrc
tok and µtok are the mean src/tgt token counts, ilarity thresholds θd = 0.5 and θs = {0.72, 0.65},
tgt
src
while %s>2 and %s>2 report the percentage of items that
contain more than one sentence. producing two datasets: wiki-simp-72 and
Dataset Pairs µsrc µtgt %src %tgt wiki-simp-65. Because LHA uses no
tok tok s>2 s>2
wiki-simp-72 25K 26.72 22.83 16% 11% document-level information in this dataset, align-
wiki-simp-65 80K 23.37 15.41 17% 7% ment leads to new sentence pairs, some of which
wiki-news-74 133K 25.66 17.25 19% 2% may be distinct from the pairs present in the exist-
wiki-news-70 216K 26.62 16.29 19% 2% ing parallel dataset. We monitor for and exclude
pairs that overlap with the testing dataset. Second,
ment experiments, which enables extraction of up we align Wikipedia to the Gigaword news ar-
to 5 sentence nearest neighbours. ticle corpus (Napoles et al., 2012), using θd = 0.5
and θs = {0.74, 0.7}, resulting in two additional
Parallel data As a parallel baseline dataset, we pseudo-parallel datasets: wiki-news-74 and
use an existing dataset from (Hwang et al., 2015). wiki-news-70. With these datasets, we inves-
The dataset consists of 282K sentence pairs ob- tigate whether pseudo-parallel data extracted from
tained after aligning the parallel articles from a different domain can be beneficial for text sim-
Wikipedia and the Simple Wikipedia. This dataset plification. We use slightly higher sentence align-
allows us to compare our results to previous work ment thresholds for the news articles because of
on data-driven text simplification. We use two the domain difference.
versions of the dataset in our experiments: full
We find that the majority of the pairs extracted
contains all 282K pairs, while partial contains
contain a single sentence, and 15-20% of the
71K pairs, or 25% of the full dataset.
source examples and 5-10% of the target exam-
Evaluation data We evaluate our simplification ples contain multiple sentences (see Table 5 for
models on the testing dataset from (Xu et al., additional statistics). Most multi-sentence exam-
2016), which consists of 358 sentence pairs from ples contain two sentences, while 0.5-1% contain
the normal Wikipedia and the Simple Wikipedia. 3 to 5 sentences. Two example aligned outputs
In addition to the ground truth simplifications, are in Table 4 (additional examples are available
each input sentence comes with 8 additional refer- in the supplementary material). They suggest that
ences, manually simplified by Amazon Meachan- our method is capable of extracting high-quality
ical Turkers. We compute BLEU and SARI on the pairs that are similar in meaning, even spanning
8 manual references. across multiple sentences.
Pseudo-parallel data We align two dataset Randomly sampled pairs We also experiment
pairs, obtaining pseudo-parallel sentence pairs for with adding random sentence pairs to the par-
text simplification (statistics of the datasets we allel dataset (rand-100K, rand-200K and
use for alignment are in Table 3). First, we rand-300K datasets, containing 100K, 200K
align the normal Wikipedia to the Simple and 300K random pairs, respectively). The
849
Table 6: Empirical results on text simplification from Wikipedia to the Simple Wikipedia. The highest SARI/BLEU results
from each category are in bold. input and reference are not generated using Beam Search.
Total pairs Beam hypothesis 1 Beam hypothesis 2
Method or Dataset
(% pseudo) SARI BLEU µtok LDsrc LDtgt SARI BLEU µtok LDsrc LDtgt
input - 26 99.37 22.7 0 0.26 - - - - -
reference - 38.1 70.21 22.3 0.26 0 - - - - -
NTS 282K (0%) 30.54 84.69 - - - 35.78 77.57 - - -
Parallel + Pseudo-parallel or Randomly sampled data (Using full parallel dataset, 282K parallel pairs)
baseline-282K 282K (0%) 30.72 85.71 18.3 0.18 0.37 36.16 82.64 19 0.19 0.36
+ wiki-simp-72 307K (8%) 30.2 87.12 19.43 0.14 0.34 36.02 81.13 19.03 0.19 0.36
+ wiki-simp-65 362K (22%) 30.92 89.64 19.8 0.13 0.33 36.48 83.56 19.37 0.18 0.35
+ wiki-news-74 414K (32%) 30.84 89.59 19.67 0.13 0.33 36.57 83.85 19.13 0.18 0.35
+ wiki-news-70 498K(43%) 30.82 89.62 19.6 0.13 0.33 36.45 83.11 18.98 0.19 0.36
+ rand-100K 382K (26%) 30.52 88.46 19.7 0.14 0.34 36.96 82.86 19 0.2 0.36
+ rand-200K 482K (41%) 29.47 80.65 19.3 0.18 0.36 34.36 74.67 18.93 0.23 0.38
+ rand-300K 582K (52%) 28.68 75.61 19.57 0.23 0.4 32.34 68.9 18.35 0.3 0.43
Parallel + Pseudo-parallel data (Using partial parallel dataset, 71K parallel pairs)
baseline-71K 71K (0%) 31.16 69.53 17.45 0.29 0.44 32.92 67.29 19.14 0.3 0.44
+ wiki-simp-65 150K (52%) 31.0 81.52 18.26 0.21 0.38 35.12 77.38 18.16 0.25 0.39
+ wiki-news-70 286K(75%) 31.01 80.03 17.82 0.23 0.4 34.14 76.44 17.31 0.28 0.43
Pseudo-parallel data only
wiki-simp-all 104K (100%) 29.93 60.81 18.05 0.36 0.47 30.13 57.46 18.53 0.39 0.49
wiki-news-all 348K (100%) 22.06 28.51 13.68 0.6 0.63 23.08 29.62 14.01 0.6 0.64
pseudo-all 452K (100%) 30.24 71.32 17.82 0.3 0.43 31.41 65.65 17.65 0.33 0.45
random pairs are uniformly sampled from the lines trained solely on parallel data. The BLEU
Wikipedia and the Simple Wikipedia, respectively. gains are larger when using the smaller paral-
With the random pairs, we aim to investigate how lel dataset, consisting of 71K sentence pairs. In
model performance changes as we add an increas- terms of SARI, the scores remain either sim-
ing number of sentence pairs that are non-parallel ilar or slightly better than the baselines, in-
but are still representative of the two dataset styles. dicating that simplicity in the output is pre-
served. The second Beam hypothesis yields higher
5.2 Automatic Evaluation SARI scores than the first one, in agreement
The simplification results in Table 6 are organized with (Štajner and Nisioi, 2018). Interestingly,
in several sections according to the type of dataset adding out-of-domain pseudo-parallel news data
used for training. We report the results of the (wiki-news-* datasets) results in an increase
top two beam search hypotheses produced by our in BLEU despite the potential change in style of
models, considering that the second hypothesis of- the target sequence.
ten generates simpler outputs (Štajner and Nisioi, Larger pseudo-parallel datasets can lead to big-
2018). ger improvements, however noisy data can result
In Table 6, input is copying the normal in a decrease in performance, motivating careful
Wikipedia input sentences, without making any data selection. In our parallel and random set-
changes. reference reports the score of the up, we find that an increasing number of random
original Simple Wikipedia references with respect pairs added to the parallel data progressively de-
to the other 8 references available for this dataset. grades model performance. However, those mod-
NTS is the previously best reported result on els still manage to perform surprisingly well, even
text simplification using neural sequence models when over half of the pairs in the dataset are ran-
(Štajner and Nisioi, 2018). baseline-{282K, dom. Thus, neural machine translation can suc-
71K} are our parallel LSTM baselines, trained on cessfully learn target transformations despite sub-
282K and 71K parallel pairs, respectively. stantial data corruption, demonstrating robustness
The models trained on a mixture of paral- to noisy or non-parallel data for certain tasks.
lel and pseudo-parallel data generate longer out- When training solely on pseudo-parallel data,
puts on average, and their output is more sim- we observe lower performance on average in com-
ilar to the input, as well as to the original parison to parallel models. However, the re-
Simple Wikipedia reference, in terms of the sults are encouraging, demonstrating the poten-
LD. Adding pseudo-parallel data frequently yields tial of our approach in tasks for which there
BLEU improvements on both Beam hypotheses: is no parallel data available. As expected, the
over the NTS system, as well as over our base- out-of-domain news data (wiki-news-all) is
850
less suitable for simplification than the in-domain Table 7: Human evaluation of the Grammaticality (G),
data (wiki-simp-all), because of the change Meaning preservation (M) and Simplicity (S) of model out-
puts (on the first Beam hypothesis).
in output style of the former. Results are best Method G M S
when mixing all pseudo-parallel pairs into a single reference 4.53 4.34 0.69
dataset (pseudo-all). Having access to a small baseline-272K 4.51 3.68 0.9
amount of in-domain pseudo-parallel data, in ad- + wiki-simp-65 4.39 3.76 0.74
pseudo-all 4.02 2.96 0.77
dition to out-of-domain pairs, seems to be benefi-
cial to the success of our approach.
Table 8: Example model outputs (first Beam hypothesis).
Method Example
5.3 Human Evaluation input jeddah is the principal gateway to mecca , is-
lam ’ s holiest city , which able-bodied mus-
Due to the challenges of automatic evaluation of lims are required to visit at least once in their
text simplification systems (Sulem et al., 2018), lifetime .
reference jeddah is the main gateway to mecca , the holi-
we also perform a human evaluation. We asked est city of islam , where able-bodied muslims
8 fluent English speakers to rate the grammatical- must go to at least once in a lifetime .
ity, meaning preservation, and simplicity of model baseline- it is the highest gateway to mecca , islam .
282K
outputs produced for 100 randomly selected sen- + wiki- jeddah is the main gateway to mecca , islam ’s
tences from our test set. We exclude any model sim-65 holiest city .
outputs which leave the input unchanged. Gram- + wiki- it is the main gateway to mecca , islam ’ s holi-
news- est city .
maticality and meaning preservation are rated on 74
a Likert scale from 1 (Very bad) to 5 (Very good). pseudo- islam is the main gateway to mecca , islam ’s
Simplicity of the output sentences, in compari- all holiest city .
son to the input, is rated following (Štajner et al.,
2018), between: −2 (much more difficult), −1
(somewhat more difficult), 0 (equally difficult), 1 of ’Jeddah’ as a city better than our parallel base-
(somewhat simpler) and 2 (much simpler). line, while correctly simplifying principal to main.
The results are reported in Table 7, where we The model trained solely on pseudo-parallel data
compare our parallel baseline (baseline-272K produces a similar output, apart from wrongly re-
in Table 6) to our best model trained on a placing jeddah with islam.
mixture of parallel and pseudo-parallel data
(wiki-simp-65) and our best model trained
on pseudo-parallel data only (pseudo-all).
6 Conclusion
We also evaluate the original Simple Wikipedia
references (reference) for comparison. In We developed a hierarchical method for extracting
terms of simplicity, our pseudo-parallel sys- pseudo-parallel sentence pairs from two mono-
tems are closer to the result of reference lingual comparable corpora composed of differ-
than is baseline-272K, indicating that ent text styles. We evaluated the performance
they better match the target sentence style. of our method on automatic alignment bench-
baseline-272K and wiki-simp-65 per- marks and extrinsically on automatic text simplifi-
form similarly to the references in terms of cation. We find improvements arising from adding
grammaticality, with baseline-272K having pseudo-parallel sentence pairs to existing parallel
a small edge. In terms of meaning preser- datasets, as well as promising results when using
vation, both do worse than the references, the pseudo-parallel data on its own.
with wiki-simp-65 having a small edge. Our results demonstrate that careful engineer-
pseudo-all performs worse on both grammat- ing of pseudo-parallel datasets can be a successful
icality and meaning preservation, but is on par approach for improving existing monolingual text-
with the simplicity result of wiki-simp-65. to-text rewriting tasks, as well as for tackling novel
In Table 8, we also show example outputs of tasks. The pseudo-parallel data could also be a
our best models (additional examples are avail- useful resource for dataset inspection and analy-
able in the supplementary material). The models sis. Future work could focus on improvements of
trained on parallel plus additional pseudo-parallel our system, such as refined approaches to sentence
data produced outputs that preserve the meaning pairing.
851
Acknowledgments William Hwang, Hannaneh Hajishirzi, Mari Osten-
dorf, and Wei Wu. 2015. Aligning sentences from
We acknowledge support from the Swiss National standard wikipedia to simple wikipedia. In HLT-
Science Foundation (grant 31003A 156976). NAACL, pages 211–217.
852
Matteo Pagliardini, Prakhar Gupta, and Martin Jaggi. Wei Xu, Chris Callison-Burch, and Courtney Napoles.
2018. Unsupervised learning of sentence embed- 2015. Problems in current text simplification re-
dings using compositional n-gram features. In Pro- search: New data can help. Transactions of the As-
ceedings of the 2018 Conference of the North Amer- sociation for Computational Linguistics, 3:283–297.
ican Chapter of the Association for Computational
Linguistics: Human Language Technologies, Vol- Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze
ume 1 (Long Papers), pages 528–540. Association Chen, and Chris Callison-Burch. 2016. Optimizing
for Computational Linguistics. statistical machine translation for text simplification.
Transactions of the Association for Computational
Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Linguistics, 4:401–415.
Jing Zhu. 2002. Bleu: a method for automatic eval-
uation of machine translation. In Proceedings of
the 40th annual meeting on association for compu-
tational linguistics, pages 311–318. Association for
Computational Linguistics.
853
Dependency-Based Relative Positional Encoding for Transformer NMT
854
Proceedings of Recent Advances in Natural Language Processing, pages 854–861,
Varna, Bulgaria, Sep 2–4 2019.
cerpt Corpus (ASPEC) Japanese-to-English trans- dependency-based NMT model that uses depen-
lation task. The experimental results demonstrate dency trees for both source and target languages.
that our approach achieves a 0.5 point gain in Their model encodes source sentences with two
BLEU over baseline Transformers (Vaswani et al., extra sequences linearized from source depen-
2017; Shaw et al., 2018). dency trees and jointly generates both target sen-
tences and their dependency trees. They applied
2 Related Work their model not only to bi-directional RNNs but
also to the Transformer, but did not improve the
NMT performance has been improved by using Transformer’s architecture. In contrast, we im-
the syntactic information of source language sen- prove the Transformer model so that it incorpo-
tences, target language sentences, or both. rates source dependency information by encod-
Some researchers have focused on phrase struc- ing pair-wise relative depths on a source depen-
tures as syntactic information. Aharoni and Gold- dency tree, which are the differences between the
berg (2017) incorporated target-side phrase struc- depths of two source words, in the encoder’s self-
tures into NMT, and Eriguchi et al. (2016) and Ma attention.
et al. (2018) incorporated source-side phrase struc- Ma et al. (2019) proposed several strategies
tures. Our work is different from their research in for improving NMT with neural syntax distance
that we focus on dependency structures rather than (NSD), which has been used for constituent pars-
phrase structures. In addition, while their models ing (Shen et al., 2018), and dependency-based
are based on RNN-based NMT models, we aim to NSD, which is an extension of the original NSD
improve a Transformer NMT model. for dependency trees. In their work, they pro-
Other researchers have focused on dependency posed a syntactic PE for Transformer NMT in
structures as syntactic information. Chen et al. order to incorporate positions on a dependency
(2017) proposed a hybrid NMT model of RNNs tree for each word via an absolute PE mechanism.
and CNNs to incorporate syntactic information In contrast, our model uses relative dependency-
into an encoder. Their model first learns source de- based distances between two words via a relative
pendency representations to compute dependency PE mechanism in the encoder’s self-attention.
context vectors by using CNNs. The RNN-based
encoder-decoder model learns a translation model, 3 Background
which is provided with the CNNs’ syntactic infor-
mation. Sennrich and Haddow (2016) proposed In this section, we first describe the baseline of our
an RNN-based NMT model that combines embed- proposed model, the Transformer model. Then,
ding vectors of linguistic features such as part- we describe a Transformer model that employs rel-
of-speech tags and dependency relation labels on ative PE.
a source sentence with the embedded representa-
tions of the source words. Eriguchi et al. (2017) 3.1 Transformer
proposed a hybrid model, called NMT+RNNG, Transformer (Vaswani et al., 2017) is an encoder-
that learns parsing and translation by combining decoder model that has a distinct architecture
recurrent neural network grammar into an RNN- based on self-attention. Figure 1 shows the ar-
based NMT. chitecture of the model. Unlike RNN-based NMT
Most existing dependency-based NMT mod- and CNN-based NMT, Transformer does not have
els, including the above-mentioned models, are a recurrent or convolutional configuration of net-
improvements over RNN-based NMT models, works. Instead, it encodes source sentences as in-
which, in terms of structure, differ greatly from termediate representations by using self-attention
the Transformer model. Because we make the pro- and decodes them by using self-attention and
posed model consider dependency information in encoder-decoder attention.
self-attention, which is the Transformer’s charac- The encoder maps an input sequence
teristic structure, the usage of dependency infor- (x1 , . . . , xn ) to a sequence of vector representa-
mation is different from their models. tions Z = (z1 , . . . , zn ). Given Z, the decoder
Recently, Wu et al. (2018) and Ma et al. generates an output sequence (y1 , . . . , yn′ ). In
(2019) incorporated syntactic information into both the encoder and decoder, the embedding
Transformer NMT. Wu et al. (2018) proposed a layer converts input tokens (source tokens in the
855
Output
Probabilities
first sub-layer is a multi-head self-attention mech-
anism, and the second layer is a simple, position-
Softmax
wise fully connected feed-forward network (FFN).
The decoder’s layer has three sub-layers. The first
Linear
sub-layer is a masked multi-head self-attention
mechanism, the second sub-layer is a multi-head
encoder-decoder attention mechanism, and the
Add & Norm third sub-layer is the FFN.
Residual connection (He et al., 2016) is applied
Feed Forward
to the sub-layers, followed by layer normaliza-
tion (Ba et al., 2016), i.e., the output of each sub-
Add & Norm layer is LayerN orm(x + Sublayer(x)), where
Encoder-Decoder
Sublayer(x) is the output of the original sub-
Add & Norm
Attention
layer.
Feed Forward The self-attention and the encoder-decoder at-
tention employ a multi-head attention mechanism.
N×
N×
Add & Norm Add & Norm The multi-head attention first computes h dot-
Masked
product attentions after linearly mapping three in-
put vectors, q, k, v *1 ∈ R1×dmodel , from dmodel
Self-Attention
Self-Attention
856
where eij is computed: bought
(xi W Q )(xj W K )T
eij = √ , (7) father car .
dz
where dz is the dimension of zi .
The decoder’s self-attention computes Equation My a red
4 by substituting the intermediate states of the de-
coder for q, k, v. During inference, however, it Figure 2: Example of Dependency Tree
is not possible for the decoder to get the infor-
mation on the words that will be generated later
the input to the next layer. Specifically, the follow-
when predicting a word, i.e., only the intermedi-
ing equation is used instead of Equation 5.
ate states of the sub-sequence that has been gener-
ated can be used for self-attention. Hence, masked n
∑
self-attention is introduced to the decoder’s self- zi = αij (xj W V + aVij ). (11)
attention so as not to calculate the self-attention j=1
between a predicted word and succeeding words.
The following equation is also used for the sub-
Masked self-attention is calculated by changing
stitution of Equation 7 in order to consider relative
Equation 7:
{ position relationships between words in calculat-
(xi W Q )(xj W K )T
√ (i ≥ j), ing eij :
eij = dz (8)
−∞ (otherwise).
xi W Q (xj W K + aK
ij )
T
eij = √ . (12)
The coefficient representing the strength of the re- dz
lationship between a certain word and the word lo-
cated behind it (i < j) becomes zero, and it can be Shaw et al. (2018) assume that relative posi-
controlled so as not to consider the relationship. tion information is not useful when the distance is
Hence, Equation 6 is changed: long. They define the maximum relative position
{ as a constant k. In addition, the relative position
∑nexp (eij ) (i ≥ j), relationships between two words are captured by
αij = k=1 exp (eik ) (9)
0 (otherwise). 2k + 1 unique labels as follows, considering that
succeeding words are in a positive direction and
In the encoder-decoder attention, the intermedi- preceding words are in a negative direction.
ate states of the decoder are used for q, and the
outputs of the encoder are used for k, v. aK K
ij = wclip(j−i,k) , (13)
The FFN for input x compute as follows:
aVij = wclip(j−i,k)
V
, (14)
F F N (x) = max(0, xW1 + b1 )W2 + b2 , (10) clip(x, k) = max(−k, min(k, x)), (15)
where W1 ∈ Rdmodel ×df f , W2 ∈ Rdf f ×dmodel are where wK = (w−k K , . . . , w K ) and w V =
k
parameter matrices, and b1 , b2 are biases.
(w−k , . . . , wk ) (wk , wk ∈ R ) are relative po-
V V K V d k
tions are added to the output of the sub-layer to be dependency-based encoding is incorporated into
857
My father bought a red car . We also describe a hybrid model that learns both
My 0 -1 -2 0 0 -1 -1 relative position representations on dependency
father 1 0 -1 1 1 0 0 structures and relative position representations for
bought 2 1 0 2 2 1 1 linear relations in sentences, i.e., the relative posi-
a 0 -1 -2 0 0 -1 -1
tional encoding explained in Section 3.2. This hy-
red 0 -1 -2 0 0 -1 -1
brid method is called T ransf ormerdep+rel . The
car 1 0 -1 1 1 0 0
. 1 0 -1 1 1 0 0 T ransf ormerdep+rel model uses the sum of aVij
and bVij and the sum of aK ij and bij as relative po-
K
Table 1: Examples of Dependency-based Inter- sition information between two words. zi and eij
Word Distances are defined in T ransf ormerdep+rel as follows:
n
∑
the self-attention mechanism, following the idea zi = αij (xj W V + aVij + bVij ), (21)
of relative positional encoding (Shaw et al., 2018). j=1
The inter-word distance on dependency trees is xi W Q (xj W K + aK K T
ij + bij )
defined as the relative depth between two words eij = √ . (22)
dz
in dependency trees. The relative depth distij be-
tween node ni and node nj corresponding to word 5 Experiments
wi and word wj is defined as follows:
5.1 Experimental Setup
distij = depth(nj ) − depth(ni ), (16)
We experimented on the WAT’18 Asian Scien-
where depth(n) is the depth of node n in a depen- tific Paper Excerpt Corpus (ASPEC) (Nakazawa
dency tree. For example, in Figure 2, the depth of et al., 2016) by using the Japanese-to-English lan-
“bought” (w3 ) relative to “My” (w1 ) is calculated guage pair. We tokenized English sentences by us-
by dist1,3 = 0 − 2 = −2. Table 1 shows a list ing Moses (Koehn et al., 2007) and Japanese sen-
of the inter-word distances on the dependency tree tences by using KyTea (Neubig et al., 2011). We
shown in Figure 2. also parsed the dependency of the Japanese sen-
The relative position between node ni and node tences by using EDA*2 .
nj in a source dependency tree is represented by For model learning, we used 1,341,417 sentence
vectors, bVij , bK pairs of 50 words or less for both the English and
ij ∈ R , and the following equa-
dk
tions are used instead of Equations 11 and 12. Japanese sentences from the first 1.5 million sen-
tence pairs of the training data (train-1.txt, train-
n
∑ 2.txt). The Japanese dictionary was comprised of
zi = αij (xj W V + bVij ), (17)
words that appeared 7 times or more in the training
j=1
data, and the English dictionary was comprised
xi W Q (xj W K + bK ij )
T of words that appeared 10 times or more in the
eij = √ . (18) training data. The other words were replaced with
dz
⟨U N K⟩ tags representing unknown words. We
We assume that the influence of a distance de-
used 1,790 sentences (dev.txt) as validation data
creases if the distance is longer than some certain
and 1,812 sentences (test.txt) as test data.
threshold. We limit the maximum distance to a
We compared our models, T ransf ormerdep
constant l. The relative position representations
and T ransf ormerdep+rel , with two baseline
bVij and bK
ij between node ni and node nj in a de- Transformer NMT models, T ransf ormerabs
pendency tree are defined with inter-word distance
(Vaswani et al., 2017), which learns absolute posi-
labels:
tion representations, and T ransf ormerrel (Shaw
bK K
ij = wclip(distij ,l) , (19) et al., 2018), which learns relative position repre-
sentations in a sentence.
bVij = V
wclip(dist . (20)
ij ,l) Hyper-parameters of all Transformer mod-
Using these expressions, the encoder’s self- els were determined, following the settings of
attention networks learn the relative position rep- Vaswani et al. (2017). We set the number of stacks
resentations on a source dependency structure. We *2
http://www.ar.media.kyoto-u.ac.jp/
call this model T ransf ormerdep . tool/EDA/
858
Model BLEU bVij bK
ij BLEU
T ransf ormerabs 25.91 ✓ ✓ 16.71
T ransf ormerrel 26.72 × ✓ 16.60
T ransf ormerdep 26.10 ✓ × 15.62
T ransf ormerdep+rel 27.22 × × 8.69
859
bKij , while T ransf ormerdep using only bij
K Kehai Chen, Rui Wang, Masao Utiyama, Lemao
was 0.11 points slightly lower than the base- Liu, Akihiro Tamura, Eiichiro Sumita, and Tiejun
Zhao. 2017. Neural machine translation with
line T ransf ormerdep . Table 3 also shows that
source dependency representation. In Proceed-
the T ransf ormerdep that used neither bVij and ings of the 2017 Conference on Empirical Meth-
bKij was 8.02 points lower than the baseline ods in Natural Language Processing. Association
T ransf ormerdep , which was significantly worse. for Computational Linguistics, pages 2846–2852.
https://doi.org/10.18653/v1/D17-1304.
These results were consistent with the exper-
imental results in Shaw et al. (2018). The Yuan Ding and Martha Palmer. 2005. Machine
dependency-based relative position representa- translation using probabilistic synchronous depen-
dency insertion grammars. In Proceedings of
tions, bK
ij and bij , were shown to be effective, but
V
the 43rd Annual Meeting of the Association for
bKij was more effective than bij .
V
Computational Linguistics (ACL’05). Association
for Computational Linguistics, pages 541–548.
6 Conclusion http://aclweb.org/anthology/P05-1067.
In this paper, we proposed a novel Transformer Akiko Eriguchi, Kazuma Hashimoto, and Yoshi-
masa Tsuruoka. 2016. Tree-to-sequence atten-
NMT model that incorporates syntactic distances tional neural machine translation. In Proceed-
between two source words into the relative po- ings of the 54th Annual Meeting of the As-
sitional encoding of an encoder’s self-attention sociation for Computational Linguistics (Volume
mechanism. We demonstrated that our proposed 1: Long Papers). Association for Computational
Linguistics, Berlin, Germany, pages 823–833.
model improved the translation accuracy, in terms https://doi.org/10.18653/v1/P16-1078.
of BLUE score, on the ASPEC Japanese-to-
English translation task. Akiko Eriguchi, Yoshimasa Tsuruoka, and Kyunghyun
Cho. 2017. Learning to parse and translate improves
For future work, we would like to improve our neural machine translation. In Proceedings of the
model by introducing relative positional encoding 55th Annual Meeting of the Association for Compu-
to target dependency structures, i.e., dependency- tational Linguistics (Volume 2: Short Papers). Asso-
based relative positional encoding for decoders. ciation for Computational Linguistics, pages 72–78.
https://doi.org/10.18653/v1/P17-2012.
For example, we would like to integrate our en-
coding into the dependency-based decoder in (Wu Jonas Gehring, Michael Auli, David Grangier, De-
et al., 2018). nis Yarats, and Yann N. Dauphin. 2017. Convolu-
tional sequence to sequence learning. In Doina Pre-
cup and Yee Whye Teh, editors, Proceedings of the
Acknowledgement 34th International Conference on Machine Learn-
ing. PMLR, International Convention Centre, Syd-
The research results have been achieved by “Re- ney, Australia, volume 70 of Proceedings of Ma-
search and Development of Deep Learning Tech- chine Learning Research, pages 1243–1252.
nology for Advanced Multilingual Speech Trans-
lation”, the Commissioned Research of National K. He, X. Zhang, S. Ren, and J. Sun. 2016.
Deep residual learning for image recognition.
Institute of Information and Communications In 2016 IEEE Conference on Computer Vision
Technology (NICT) , JAPAN. This work was par- and Pattern Recognition (CVPR). pages 770–778.
tially supported by JSPS KAKENHI Grant Num- https://doi.org/10.1109/CVPR.2016.90.
ber JP18K18110. Diederik P. Kingma and Jimmy Ba. 2015. Adam:
A method for stochastic optimization. In 3rd
International Conference on Learning Represen-
References tations, ICLR 2015, San Diego, CA, USA,
May 7-9, 2015, Conference Track Proceedings.
Roee Aharoni and Yoav Goldberg. 2017. Towards http://arxiv.org/abs/1412.6980.
string-to-tree neural machine translation. In Pro-
ceedings of the 55th Annual Meeting of the As- Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris
sociation for Computational Linguistics (Volume Callison-Burch, Marcello Federico, Nicola Bertoldi,
2: Short Papers). Association for Computational Brooke Cowan, Wade Shen, Christine Moran,
Linguistics, Vancouver, Canada, pages 132–140. Richard Zens, Chris Dyer, Ondrej Bojar, Alexan-
https://doi.org/10.18653/v1/P17-2021. dra Constantin, and Evan Herbst. 2007. Moses:
Open source toolkit for statistical machine trans-
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin- lation. In Proceedings of the 45th Annual Meet-
ton. 2016. Layer normalization. arXiv preprint ing of the Association for Computational Linguistics
arXiv:1607.06450 . Companion Volume Proceedings of the Demo and
860
Poster Sessions. Association for Computational Lin- Yikang Shen, Zhouhan Lin, Athul Paul Jacob, Alessan-
guistics, Prague, Czech Republic, pages 177–180. dro Sordoni, Aaron Courville, and Yoshua Ben-
https://www.aclweb.org/anthology/P07-2045. gio. 2018. Straight to the tree: Constituency
parsing with neural syntactic distance. In Pro-
Thang Luong, Hieu Pham, and Christopher D. Man- ceedings of the 56th Annual Meeting of the Asso-
ning. 2015. Effective approaches to attention-based ciation for Computational Linguistics (Volume 1:
neural machine translation. In Proceedings of the Long Papers). Association for Computational Lin-
2015 Conference on Empirical Methods in Natu- guistics, Melbourne, Australia, pages 1171–1180.
ral Language Processing. Association for Compu- https://doi.org/10.18653/v1/P18-1108.
tational Linguistics, pages 1412–1421.
Emma Strubell, Patrick Verga, Daniel Andor,
Chunpeng Ma, Akihiro Tamura, Masao Utiyama, Ei- David Weiss, and Andrew McCallum. 2018.
ichiro Sumita, and Tiejun Zhao. 2019. Improv- Linguistically-informed self-attention for se-
ing neural machine translation with neural syntac- mantic role labeling. In Proceedings of the
tic distance. In Proceedings of the 2019 Confer- 2018 Conference on Empirical Methods in
ence of the North American Chapter of the As- Natural Language Processing. Association for
sociation for Computational Linguistics: Human Computational Linguistics, pages 5027–5038.
Language Technologies, Volume 1 (Long and Short http://aclweb.org/anthology/D18-1548.
Papers). Association for Computational Linguis- Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
tics, Minneapolis, Minnesota, pages 2032–2037. Sequence to sequence learning with neural net-
https://www.aclweb.org/anthology/N19-1205. works. In Z. Ghahramani, M. Welling, C. Cortes,
N. D. Lawrence, and K. Q. Weinberger, editors, Ad-
Chunpeng Ma, Akihiro Tamura, Masao Utiyama, vances in Neural Information Processing Systems
Tiejun Zhao, and Eiichiro Sumita. 2018. Forest- 27, Curran Associates, Inc., pages 3104–3112.
based neural machine translation. In Proceed-
ings of the 56th Annual Meeting of the Asso- Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
ciation for Computational Linguistics (Volume 1: Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz
Long Papers). Association for Computational Lin- Kaiser, and Illia Polosukhin. 2017. Attention is all
guistics, Melbourne, Australia, pages 1253–1263. you need. In I. Guyon, U. V. Luxburg, S. Ben-
https://www.aclweb.org/anthology/P18-1116. gio, H. Wallach, R. Fergus, S. Vishwanathan, and
R. Garnett, editors, Advances in Neural Informa-
Toshiaki Nakazawa, Manabu Yaguchi, Kiyotaka Uchi- tion Processing Systems 30, Curran Associates, Inc.,
moto, Masao Utiyama, Eiichiro Sumita, Sadao pages 5998–6008. http://papers.nips.cc/paper/7181-
Kurohashi, and Hitoshi Isahara. 2016. Aspec: Asian attention-is-all-you-need.pdf.
scientific paper excerpt corpus. In Proceedings of
the Ninth International Conference on Language Re- Shuangzhi Wu, Dongdong Zhang, Zhirui
sources and Evaluation (LREC 2016). pages 2204– Zhang, Nan Yang, Mu Li, and Ming Zhou.
2208. 2018. Dependency-to-dependency neural ma-
chine translation. IEEE/ACM Trans. Audio,
Graham Neubig, Yosuke Nakata, and Shinsuke Mori. Speech and Lang. Proc. 26(11):2132–2141.
2011. Pointwise prediction for robust, adaptable https://doi.org/10.1109/TASLP.2018.2855968.
japanese morphological analysis. In Proceedings of
the 49th Annual Meeting of the Association for Com-
putational Linguistics: Human Language Technolo-
gies. Association for Computational Linguistics,
pages 529–533. http://aclweb.org/anthology/P11-
2093.
861
From Image to Text in Sentiment Analysis via Regression and Deep
Learning
862
Proceedings of Recent Advances in Natural Language Processing, pages 862–868,
Varna, Bulgaria, Sep 2–4 2019.
a. b.
They fighting over who more useless
ping performs better for the positive sentiment cat-
This was so kak sad as a child
egory in comparison with the neutral and negative
categories. Furthermore, the experimental results
show that the more complex deep learning features
perform better than the RGB pixel-value features
for all sentiment categories and for larger training
sets.
c. d. The paper is organized as follows. Section 2
NEWS: Rams mailbag: How much has
Case Keenum improved? #SPORTS #LATIMES
Well that's adorable discusses related works. Section 3 describes a
Kernel Ridge Regression model for image to text
mapping. Section 4 shows the experimental evalu-
ation performed on a real-world data set. Section 5
finishes with conclusions and directions for future
research.
2 Related Work
Figure 1: Motivation Figure: Our model treats lan-
guage as a rich label space and generates subjec- Image captioning. The research presented in this
tive descriptions of images. Examples of samples paper is in the direction of image captioning, but
used in the experimental evaluation. Each sam- goes further to map images to text. The texts that
ple consists of a pair made of an image and the we consider are not only descriptions of the im-
subjective text associated to it. Each sample has a ages, which is the task of image captioning, but
sentiment associated to it: a., b. samples convey they contain subjective statements related to the
a negative sentiment; c. sample conveys a neutral images. Mapping images to text is an extension
sentiment; d. sample conveys a positive sentiment. of the image captioning task, and this mapping al-
lows us to build some dictionaries of words and
select from these dictionaries the words which are
sentence represents the output. We employ a Ker- the most relevant to an image. The learning set-
nel Ridge Regression for the task of mapping im- ting that we investigate in this paper is different
ages to text. We considered two types of image to the image captioning setting, because our sys-
features: i) RGB pixel-values features, and ii) fea- tem automatically associates an image to a set of
tures extracted with a deep learning approach. We words from a dictionary, these words being not
used a bag-of-words model to construct the text only descriptors of the content of the image, but
features. In addition, we consider several sen- also subjective opinions of the image. Image cap-
timent categories associated to each image-text tioning has been actively studied in last years, a
sample, and analyze this mapping in the context recent survey on image captioning is given in (Bai
of these sentiment categories. and An, 2018). Several approaches for image cap-
We investigate data from Twitter. These data tioning are making use of the deep learning tech-
contain images and text associated to each image. niques (Bai and An, 2018; P. Singam, 2018).
The text is a subjective description or impression Image description. Several approaches that ad-
of the image, written by a user. Data from so- dress the challenge of generating image descrip-
cial networks, and especially Twitter, is usually tions have been proposed (Kulkarni et al., 2013;
associated to a sentiment, which could be a pos- Karpathy and Fei-Fei, 2015; Park et al., 2017;
itive, neutral or negative sentiment. We designed Ling and Fidler, 2017). However, these models
a system that automatically associates an image to only rely on objective image descriptors, and do
a set of words from a dictionary, these words be- not take into account the subjectivity which ap-
ing not only descriptors of the content of the im- pears when describing an image on social net-
age, but also subjective impressions and opinions works.
of the image. Sentiment analysis. We investigate mapping
One of the interesting findings of our work is images to text in the context of sentiment analysis.
that there is a difference in performance for dif- Most of the previous research in sentiment analy-
ferent sentiment categories, in particular the map- sis is performed on text data. Recent works focus
863
on sentiment analysis in images and videos (Yu IdX in Equation 2 represents the identity matrix
et al., 2016; You et al., 2015; Wang et al., 2016). of dimension dX ).
The research on visual sentiment analysis pro- For high-dimensional data, an explicit compu-
ceeds along two dimensions: i) based on hand- tation of W as presented in Equation 2 without
crafted features and ii) based on features gener- prior dimensionality reduction is computationally
ated automatically. Deep Learning techniques are expensive. Fortunately, Equation 2 can be rewrit-
capable of automatically learning robust features ten as:
from a large number of images (Jindal and Singh,
2015). An interesting direction for sentiment anal- W = FY FXT (FX FXT + αIdx )−1
ysis is related to word representations and capsule = FY (FXT FX + αIn )−1 FXT (3)
networks for NLP applications (Xing et al., 2019;
Zhao et al., 2019). Making use of the kernel trick, the inputs
xi are implicitly mapped to a high-dimensional
3 Kernel Ridge Regression for Mapping Reproducing Kernel Hilbert space (Berlinet and
Images to Text Thomas-Agnan, 2011):
Let X = {x1 , x2 , . . . , xn } and Y =
Φ = [φ(x1 ), . . . , φ(xn )]. (4)
{y1 , y2 , . . . , yn } be the set of inputs and out-
puts, respectively, and n represents the number When predicting a target ynew from a new obser-
of observations. And let FX ∈ RdX ×n and vation xnew , explicit access to Φ is never actually
FY ∈ RdY ×n denote the input and output feature needed:
matrices, where dX , dY represent the dimensions
of the input and output features respectively. ynew = FY (ΦT Φ + αIn )−1 ΦT φ(xnew )
The inputs represent the images, and the input = FY (K + αIn )−1 κ(xnew ) (5)
features can be either simple RGB pixel-values
or something more complex, such as features With Kij = φ(xi )T φ(xj ) and κ(xnew )i =
extracted automatically using convolutional φ(xi )T φ(xnew ), the prediction can be described
neural networks (O’Shea and Nash, 2015). The entirely in terms of inner products in the higher-
outputs represent the texts associated to the dimensional space. Not only does this approach
images and the output features can be extracted work on the original data sets without the need
using Word2Vec (Ma and Zhang, 2015). of dimensionality reduction, but it also opens up
A mapping between the inputs and the out- ways to introduce non-linear mappings into the re-
puts can be formulated as a multi-linear regres- gression by considering different types of kernels,
sion problem (Cortes et al., 2005, 2007). Com- such as a Gaussian or a polynomial kernel.
bined with Tikhonov regularization, this is also
known as Kernel Ridge Regression (KRR). The 4 Experimental Evaluation
KRR method is a regularized least squares method 4.1 Dataset
that is used for classification and regression tasks.
It has the following objective function: We used a data set with images and text that was
introduced in (Vadicamo et al., 2017). The data
1 1 have been collected from Twitter posts over a pe-
argW min( ||W FX − FYT ||2F + α ||W ||2F ) (1)
2 2 riod of 6 months, and using an LSTM-SVM ar-
where || · ||F is the Frobenius norm, α is a regu- chitecture, the tweets have been divided into three
larization term and the superscript T signifies the sentiment categories: positive, neutral, and nega-
transpose of the matrix. tive. For image labelling the authors have selected
The solution of the optimization problem from data with the most confident textual sentiment pre-
Equation 1 involves the Moore-Penrose pseudo- dictions and they used these predictions to au-
inverse and has the following closed-form expres- tomatically assign sentiment labels to the corre-
sion: sponding images. In our experimental evaluation
we selected 10000 images and the corresponding
W = FY FXT (FX FXT + αIdX )−1 ∈ RdY ×dX (2)
10000 tweets from each of the three sentiment cat-
which for low-dimensional feature spaces egories. Figure 1 shows examples of image and
(dX , dY ≤ n) can be calculated explicitly (the text data used in the experimental evaluation.
864
1 Persian cat 1 Plate
1 Aircraft carrier 1 Racket
2 Scoreboard 2 Tabby 2 Cheeseburger
2 Fire boat
3 Drilling platform 3 Ballplayer 3 Pekinese 3 Carbonara
4 Dock 4 Flagpole 4 Egyptian cat 4 Hotdog
5 Submarine 5 Stage 5 Tiger cat 5 Meat loaf
4.2 Image and Text Features jaraman et al., 2018). For this reason, we ex-
4.2.1 Image Features tracted the features from the last layer before the
final classification, so the entire convolutional base
The research on feature extraction from images
was used for this. The features were extracted us-
proceeds along two directions: i) traditional, hand-
ing the pre-trained convolutional base VGG16 net-
crafted features, and ii) automatically generated
work (Simonyan and Zisserman, 2014). For com-
features. With the increasing number of images
putational reasons, the images were resampled to
and videos on the web, traditional methods have
a 3232 pixel resolution. The model was initialized
a hard time handling the scalability and general-
by the ImageNet weights. For understanding what
ization problem. In contrast, automated generated
part of an image was used to extract the features,
feature-based techniques are capable to automat-
visualizing heatmaps of class activation technique
ically learn robust features from a large number
was employed. This is a technique which illus-
of images (Jindal and Singh, 2015). We discuss
trates how intensely the input image activates dif-
below how these two directions for extracting fea-
ferent channels, how important each channel is
tures from images apply in our case, in particular,
with regard to the class and how intensely the input
we use RGB pixel-values features for the first di-
image activates the class. Figure 2 illustrates the
rection and Deep Learning based features for the
heatmaps of class activation for some random im-
second direction.
ages using VGG16 as a pre-trained convolutional
RGB pixel-values. In this approach for extract- base. The VGG16 model makes the final clas-
ing features from images, we simply convert the sification decision based on the highlighted parts
images into arrays. Each image was sliced to get from each image, and furthermore each image is
the RGB data. The 3-channels RGB image for- associated with the five most representative cap-
mat was preferred instead of using 1-channel im- tions.
age format since we wanted to use all the available
information related to an image. Using this ap- 4.2.2 Text Features
proach, each image was described by a 2352 (28 x
We used a Bag-of-Words (BoW) model (Harris,
28 x 3)-dimensional feature vector.
1954) for extracting the features from the text sam-
Deep Learning based features. Deep Learning ples. The first step in building the BoW model
models use a cascade of layers to discover feature consists of pre-processing the text: removing non-
representations from data. Each layer of a con- letter characters, removing the html tag from the
volutional network produces an activation for the Twitter posts, converting words to lower cases, re-
given input. Earlier layers capture low-level fea- moving stop-words and making the split. A vo-
tures of the image like blobs, edges, and colors. cabulary is built from the words that appear in the
This primitive features are abstracted by the high- text samples. The input of the BoW model is a list
level layers. Studies from the literature suggest of strings and the output is a sparse matrix with
that while using pre-trained networks for feature the dimension: number of samples x number of
extraction, the features should be extracted from words in the vocabulary, having 1 if a given word
the layer right before the classification layer (Ra- from the vocabulary is contained in that particular
865
text sample. We initialized the BoW model with
a maximum of 5000 features. We extracted a vo-
cabulary for each sentiment category, and the cor-
responding 0-1 feature vector for each text sample.
1. Could our proposed Kernel Ridge Regression Figure 3: The plots show mean errors and stan-
model map images to natural language de- dard deviation for different sizes of the training
scriptors? set. Comparison between RGB pixel-values fea-
tures and the more complex VGG16 features. The
2. What is the difference between the two types
different rows correspond to different sentiment
of image features that we considered? In par-
categories: top row - positive sentiment category,
ticular, we are interested whether the more
middle row - neutral sentiment category, bottom
complex deep learning features give a bet-
row - negative sentiment category.
ter performance in comparison to the simple
RGB pixel-values features.
866
ing can be interpreted as positive for the majority
of the people, but a neutral or a negative sentiment
can be interpreted as having a different meaning
depending on the people.
Furthermore, analyzing again Figure 3, we see
that the neutral sentiment category has a differ-
ent behaviour in comparison with the positive and
negative sentiment categories, with respect to the
image features used. In the case of neutral sen-
timent, the more complex VGG16 features ap-
pear to have a better performance than the sim-
Figure 4: Comparison of the learning performance pler RGB pixel-values features as the size of the
based on the type of sentiment using the VGG16 data increases. For positive and negative sentiment
image features. categories the simpler RGB pixel-values features
lead to an error which varies a lot, while using the
VGG16 features, the error is more stable.
We designed the following experimental proto-
col. For each of the three sentiment categories, we
randomly split the data 5 times into training and 5 Conclusions and Future Work
testing, taking 70% for training and the rest for
testing. For training the model, we considered dif- In this work, we investigated a method for image
ferent sizes of the training set: from 50 to 7000 to text mapping in the context of sentiment anal-
observations with a step size of 50. For a cor- ysis. The mapping from images to text was per-
rect evaluation, the models built on these different formed using a Kernel Ridge Regression model.
training sets, were evaluated on the same test set. We considered two types of image features: i) the
The error was averaged over the 5 random splits of simple RGB pixel-values features, and ii) a more
the data into training and testing. complex set of image features extracted with a
deep learning approach. Furthermore, in this pa-
Results per we took a step forward form the image cap-
The first two questions raised above can be an- tioning task, which allows us to build some dictio-
swered by analyzing the experimental results naries of words and select from these dictionaries
shown in Figure 3. The plots show the learning the words which are the most relevant to an im-
curve (mean errors and standard deviations) for age. We performed the experimental evaluation on
different sizes of the training set and for different a Twitter data set containing both text and images
sentiment categories. Since the error decreases as and the sentiment associated with these. We found
the training size increases, we can say that there is that there is a difference in performance for differ-
a learning involved, thus our proposed model can ent sentiment categories, in particular the mapping
map images to natural language descriptors. performs better for the positive sentiment category
The plots from Figure 3 also show the compar- in comparison with the neutral and negative ones
ison between the RGB pixel-values and VGG16 for both features extraction techniques.
features for the three categories of sentiments con- We plan to further extend our approach by in-
sidered. Overall, the more complex deep learning vestigating the input-output kernel regression type
features give a better performance in comparison of learning (Brouard et al., 2016). The output ker-
to the simple RGB pixel-values features. nel would allow us to take into account the struc-
To answer the third question, we analyzed the ture in the output space and benefit from the use
experimental results shown in Figure 4. There of kernels. We also plan to integrate in our model
is a significant difference in learning performance textual captions of images obtained using a pre-
for the positive sentiment category in comparison trained network (Simonyan and Zisserman, 2014).
with the other two categories, both using RGB The textual captions could be used as a new type of
pixel-values features and VGG16 features. The features and can be compared and integrated with
positive category is simpler to be learned because the other two types of image features considered.
of the subjective part from images: a positive feel-
867
References Park, C., Kim, B., and Kim, G. (2017). Attend to
you: Personalized image captioning with context se-
Bai, S. and An, S. (2018). A survey on automatic im- quence memory networks. In Proceedings of the
age caption generation. Neurocomputing, 311. IEEE Conference on Computer Vision and Pattern
Recognition, pages 895–903.
Berlinet, A. and Thomas-Agnan, C. (2011). Reproduc-
ing Kernel Hilbert Spaces in Probability and Statis- Rajaraman, S., Antani, S. K., Poostchi, M., Sila-
tics. Springer Science Business Media. mut, K., Hossain, M. A., Maude, R. J., Jaeger, S.,
and Thoma, G. R. (2018). Pre-trained convolu-
Brouard, C., Szafranski, M., and D’Alché-Buc, F.
tional neural networks as feature extractors toward
(2016). Input output kernel regression: super-
improved malaria parasite detection in thin blood
vised and semi-supervised structured output predic-
smear images. PeerJ, 6:e4568.
tion with operator-valued kernels. The Journal of
Machine Learning Research, 17(1):6105–6152. Simonyan, K. and Zisserman, A. (2014). Very deep
convolutional networks for large-scale image recog-
Cortes, C., Mohri, M., and Weston, J. (2005). A gen- nition. arXiv preprint arXiv:1409.1556.
eral regression technique for learning transductions.
In Proceedings of ICML 2005, pages 153–160. ACM Vadicamo, L., Carrara, F., Cimino, A., Cresci, S.,
Press. Dell’Orletta, F., Falchi, F., and Tesconi, M. (2017).
Cross-media learning for image sentiment analysis
Cortes, C., Mohri, M., and Weston, J. (2007). A in the wild. In The IEEE International Conference
general regression framework for learning stringto- on Computer Vision (ICCV) Workshops.
string mappings. In Predicting Structured Data.
MIT Press. Wang, J., Fu, J., Xu, Y., and Mei, T. (2016). Be-
yond object recognition: Visual sentiment analysis
Harris, Z. (1954). Distributional structure. pages 146– with deep coupled adjective and noun neural net-
62. works. In Proceedings of the Twenty-Fifth Interna-
tional Joint Conference on Artificial Intelligence, IJ-
Jindal, S. and Singh, S. (2015). Image sentiment
CAI’16, pages 3484–3490. AAAI Press.
analysis using deep convolutional neural networks
with domain specific fine tuning. In 2015 In- Xing, F. Z., Pallucchini, F., and Cambria, E. (2019).
ternational Conference on Information Processing Cognitive-inspired domain adaptation of sentiment
(ICIP), pages 447–451. IEEE. lexicons. Information Processing & Management,
56(3):554–564.
Karpathy, A. and Fei-Fei, L. (2015). Deep visual-
semantic alignments for generating image descrip- You, Q., Luo, J., Jin, H., and Yang, J. (2015). Joint
tions. In Proceedings of the IEEE conference visual-textual sentiment analysis with deep neural
on computer vision and pattern recognition, pages networks. In Proceedings of the 23rd ACM Inter-
3128–3137. national Conference on Multimedia, MM ’15, pages
1071–1074, New York, NY, USA. ACM.
Kulkarni, G., Premraj, V., Ordonez, V., Dhar, S., Li,
S., Choi, Y., Berg, A. C., and Berg, T. L. (2013). Yu, Y., Lin, H., Meng, J., and Zhao, Z. (2016). Vi-
Babytalk: Understanding and generating simple im- sual and textual sentiment analysis of a microblog
age descriptions. IEEE Transactions on Pattern using deep convolutional neural networks. Algo-
Analysis and Machine Intelligence, 35(12):2891– rithms, 9(2):41.
2903.
Zhao, W., Peng, H., Eger, S., Cambria, E., and Yang,
Ling, H. and Fidler, S. (2017). Teaching machines M. (2019). Towards scalable and reliable capsule
to describe images via natural language feedback. networks for challenging nlp applications. arXiv
In Proceedings of the 31st International Conference preprint arXiv:1906.02829.
on Neural Information Processing Systems, pages
5075–5085. Curran Associates Inc.
868
Building a Morphological Analyser for Laz
869
Proceedings of Recent Advances in Natural Language Processing, pages 869–877,
Varna, Bulgaria, Sep 2–4 2019.
dangered languages lacks access to Information and Flag diacritics for representing Laz verbal complex
Communication Technology and the representation and for generating complex verbal word forms.
of these languages on digital environment is rather
low. Considering Laz resources online, there are 2 Laz
some online dictionaries and only a couple of web There are eight dialects of the Laz language, none
sites that give information about the Laz language of which is considered to be normative or ‘standard’.
and culture, and many of them are mostly in Turkish Even though underlyingly the structure of these di-
or English. He suggests that regarding ‘digital lan- alects is the same, they show lexical and morpholog-
guage divide’ such small regional languages must be ical, as well as phonological differences. There are
represented more by creating and using resources in two main groups. The Western dialects (Gyulva),
digital format. such as Pazar (Atina), Çamlıhemşin (Furthunaşi
One of the drawbacks while working with these gamayona), Ardeşen (Arthaşeni) and the Eastern
languages is clearly the small amount of data to dialects (Yulva) as Fındıklı (Viʒe), Arhavi (Arkabi),
begin with (written or spoken, annotated or non- Hopa (Xopa), Borçka-İçkale (Çxala).3
annotated) (Riza, 2008). Current dominant compu- For the purposes of this initial study we have cho-
tational methods and tools are mostly used on lan- sen to base the analyser on the Pazar dialect. The
guages with large corpora, following a statistical ap- reasons for this are twofold: Firstly the Pazar dialect
proach to train their systems according to a rele- is less irregular in terms of verbal inflection and sec-
vant task. However, with little data at hand these ondly a separate and well-documented grammar of
methods may not present a good solution. There- Laz written in English is based only on this dialect.4
fore, Gerstenberger et al. (2017) suggests a rule- Unfortunately, there is no study yet that would pro-
based morpho-syntactic modelling for annotating vide an analysis of Laz grammar to be treated as
small language data. On their study of Komi lan- ‘standard’ (Haznedar, 2018).
guage, his results show by-far significant advantages
of rule-based approaches for endangered languages 2.1 Verbs
by providing much more precise results in tagging as In terms of morphosyntactic alignment, Laz is an
well as ‘full- fledged grammatical description based ergative–absolutive language. It marks the subject
on broad empirical evidence’ and a future develop- of unergative predicates and transitives with agen-
ment for computer-assisted language learning sys- tive/causer subjects with ergative case while the
tems. subject of unaccusative predicates and the direct ob-
In this study, the aim is to create a morphologi- ject of transitive and ditransitive verbs are inflected
cal analyzer using the Helsinki Finite State Toolkit with nominative case. These patterns are marked
(HFST) that will help to overcome manual anno- differently on the verbal complex, depending on
tation of a potential Laz corpus. Additionally, as their case markings which also indicate their argu-
Gerstenberger et al. (2017) suggest, these may later ment types. The verb encodes person information
help developing programs to be able to facilitate both preverbally and postverbally as seen in Table 1
learning of Laz, considering the increase of inter- and Table 2 and 3.5 While we can observe verbs
est in Laz courses not only in secondary schools but agreeing with agent-like arguments and sole argu-
also in universities such as Boğaziçi University and 3
We exclude the Sapanca dialect as the region in which it
İstanbul Bilgi University (Haznedar, 2018). From is spoken is further away from the other dialects. Speakers of
spelling and grammar-checkers to machine transla- the Sapanca dialect are considered to be migrated from Batum,
tion systems and language learning materials, this Georgia to Sapanca, Turkey (Bucaklişi and Kojima, 2003)
4
The main grammar book used for this study is Pazar Laz
small study will hopefully lead to further develop- written by Öztürk and Pöchtrager (2011) which is based on
ments on Laz language in the field of Computational courses given by İsmail Bucaklişi in Boğaziçi University and it
Linguistics. is the most recent and only complete study on a dialect of Laz
written in English which would enable us to define grammat-
The remainder of the paper is laid out as follows: ical rules for the morphological analyser. The grammar book
in Section 2, the grammatical structure of Laz is by René Lacroix (2009) was also referred to several times but
discussed and later, in Section 3.1 and 3.2 the pro- since it is mostly based on Arhavi (Arkabi) dialect and writ-
ten in French, we used it when we needed to look for more
cess of preparing a lexicon and corpus for Laz is de- examples for certain structures and specific details, especially
scribed. The Section 4 gives and explains the details valency-related vowels on the verbal complex and verb classes.
5
of the morphological analyzer and the usefulness of It should be noted that post-verbal person markers encode
tense information as well.
870
ments in both positions at the same time, we can or a non-core argument8 respectively. All these
only observe theme-patient (of mono-transitive and operations commonly mark the verb with different
ditransitive verbs) and dative marked recipient-goal valency-related vowel in the same preverbal posi-
or applied non-core argument agreement in pre- tion. Therefore, under conditions where the verb
verbal position. Additionally, Laz applies a hierar- is needed to be inflected both applicativization and
chical selection rule among arguments while mark- causativization at the same time, vowel i/u-
ing person in -2 pre-verbal position seen in Table 1. suppresses vowel o-. This is important for us
This can be represented as in (1), where repre- since when we mark the verb with , the struc-
sents the dative-marked arguments in the structure ture should allow construction as well. We will
and represents theme/ patient argument type while discuss such intersecting constructions and how we
means agent-like argument type.6 deal with them in our lexicon file in Section 4.1 in
detail. An example of this from Öztürk and Pöch-
(1) 1/2 > 1/2 > 1 > 3= 3= 2/3
trager (2011) is given in (4).
The reason why 1/2 arguments comes first but not
3 is that 3 is unmarked; therefore, overt / 1 (4) Him Ayşe-s bere
markings fills the position if they are available in the S/he Ayşe. child-
u-bgar-ap-ap-u-n
structure. Only when we have 2/3 type argument, -cry- - . - - .3. .
the position remains empty. We will also discuss ‘S/he has made Ayşe make the baby cry.’
this topic later in Section 4.1.
871
-4 -3 -2 -1 0
Valency-
Affirmative
Spatial preverb Person marker related Root
preverb
vowel
Table 1: Pre-verbal complex the numbers in the header refers to the pre-verbal position relative to the verbal
root. The spatial preverb is a prefix that indicates the direction or manner of an event. The different forms
of person markers are realised based on the laryngeal properties of the following consonant. This will be
later discussed in Section 4.2.
0 1 2 3 4 5 6
Cusative
Causative Causative
Augment. suffix for Imperfect
suffix for suffix for Thematic
Root stem present stem
intransi- transi- suffix
formant perfect formant
tives tives
cons.
-am,-
-tă x-
-am -in -ap -ap um,-er,- -ť
’break’
ur
7 8 9 10 11
Subjunctive Person Conditional
Plurality Auxiliaries
marker suffixes marker
S-A.1.PST: -i
S-A.2.PST: -i -ǩo -t; -es, an -(e)re,
-a
S-A.3.PST: -u (3rd ) -(e)rťu
S-A.1.PRS: ∅
872
ish alphabet based on Latin letters. Across texts, pronominals were among the hardest tasks for this
they have been encoded with different characters study. We needed to separate these word classes
but these forms will be the standard for our study. semi-automatically because there were words that
should be put in more than one category such as
3 Resources in both noun and adjective or adjective and adverb
3.1 Lexicon (determined only by sentential position) or noun and
adverb. Therefore, it could not be possible for us to
The lexicon composed for this study comes from the
include words (except verbs) that belong to other
Büyük Lazca Sözlük (Didi Lazuri Nenapuna). It is
dialects other than Pazar for this study.
the most extensive dictionary available for Laz pre-
pared by Hasan Uzunhasanoğlu , İsmail Bucaklişi 3.2 Corpus
and İrfan Çağatay Aleksiva in 2007 in Laz and Turk- We have collected different type of written texts
ish. for our Laz corpus. However, differences in terms
The verbs were extracted from the dictionary au- of dialects have forced us to divide texts into their
tomatically whereas other word classes were ex- corresponding dialects for this study since we have
tracted semi-automatically. The words are taken as decided on working Pazar Laz first the reasons of
entries with their dialect labels10 and if available, which are discussed in Section 2. Unfortunately,
dialect-specific forms as seen in (5).11 Pazar Laz has almost no written text known in the
(5) doinu [Atn., Viw, dorinu Gyl., Ark., Xop., literature. The only resource we have is an 800 page
doǩunapa Sap.] document consisting of 111,365 tokens collected by
İsmail Bucaklişi, a native speaker of Pazar Laz, by
We have prepared verb word lists for each dialect himself which contains daily conversations and sto-
separately as well as a complete word list for all. ries shared in his immediate circle. It should be
Considering the possibility that dialects may borrow noted that it also contains Turkish words and sen-
words from one another, we decided to build a lex- tences given as translations throughout the docu-
icon based on not only the Pazar dialect but all di- ment the effects of which on the results can be seen
alects of Laz. This is an important strategy to form in Section 5.2.
a ‘common source lexicon’ (Beesley and Karttunen,
2003). However, for the sake of simplicity, we have 4 Methodology
excluded nominal and verbal compounds from our
lexicon. The purpose of this project is to develop a compu-
The challenging part in preparing the lexicon has tational model for morphological analysis for Laz
been the stemming process for verbs since the verbs by using the Helsinki Finite-State Toolkit (HFST;
in the dictionary are in their infinitival form and (Linden et al., 2011)) which is popular in this
some of them also include preverbs. Even though field of research. A finite-state transducer asso-
the preverbs have been easily separated, the infini- ciates a morphological analysis with the correspond-
tive suffixes were harder to process. For example, ing phonological representation. Xerox lexc and
there are verbs ending with -alu and while some twolc formalism supported by HFST are used to
of these verbs include -al suffix in their bare form, create lexicon files and a two-level grammar file re-
some do not. It means that they are lexically deter- spectively (Beesley and Karttunen, 2003).
mined. 4.1 Lexicon Files
Even though noun declension was easy to de-
The lexc (Lexicon Compiler) formalism is used
fine, extracting substantives from the dictionary
to define lexicons which contain grammatical labels
and carefully separating them into nouns, adverbs,
and morphotactic rules for the morphemes in the
and adjectives as well as categorizing other syn-
language (Beesley and Karttunen, 2003).
tactic elements like interjections, conjunctions and
10
The following dialect codes were found in the dictionary: 4.1.1 lexc File for Substantives
‘Yul’ (Eastern dialects), ‘Gyl’ (Western dialects), ‘Viw’ (Viʒe), The substantive lexc file has 27 tags for mor-
‘Xop’ (Xopa), ‘Ark’ (Arkabi), ‘Çxl’ (Çxala), ‘Atn’ (Atina/Pazar),
‘Fur’ (Furthunaşi gamayona), ‘Arş’ (Arthaşeni), ‘Sap’ (Sapanci).
phemes indicating person, number and case infor-
11
(5) shows that doinu ‘to give birth’ belongs to Atn. and mation and 18 continuation classes for morpho-
Viw. dialects, and it takes the form of dorinu in Gyl., Ark. and tactics or word-formation rules together with the
Xop. dialects, and doǩunapa in Sap. dialect.
873
a b c ç ç̆ d e f g ğ h x i j k ǩ l
[a] [b] [dʒ] [͡tʃh ] [tʃʼ] [d] [e] [f] [g] [ɣ] [h] [x] [i] [ʒ] [kh ] [kʼ] [l]
m n o p p̆ r s ş t t̆ u v y z ž ʒ ʒ̆
[m] [n]/[ŋ] [p] [ph ] [pʼ] [r] [s] [ʃ] [th ] [tʼ] [u] [v]/[w] [j] [z] [dz] [tsh ] [͡tsʼ]
Table 4: The 34 letters of the Laz alphabet. The letters ǩ p̆ ť ç̆ ʒ̆ ʒ and ž represent ejective consonants.
lexemes. We specified pronouns as personal pro- structures with 3rd person suffixes by disallowing
nouns, possessive pronouns, demonstratives, reflex- paths including 1st and 2nd person prefixes. This
ives, interrogative pronouns, indefinite pronouns, is done by setting a flag, @ . - .3@ which means
quantifiers as well as numerals in different contin- Positive setting of the - (dative–patient) type ar-
uation classes. We have continuation classes for gument bearing 3rd person information and later in
case and plural markers to show nominal inflection. person suffix continuation class we reject/Disallow
There are two forms of plural markings and ablative those paths with positively setting of 3rd person in-
markings each, whose differentiation is lexical, not formation by setting @ . - .3@ for the 1st and 2nd
phonological. Therefore, we have encoded this in- person suffixation. Additionally, we can also reject
formation in our lexicon by using flag diacratics that paths that include combinations of 1st and 2nd per-
will be explained in Section 4.1.3. son - with 1st and 2nd person (agent) respec-
tively since the language does not allow such con-
4.1.2 lexc File for Verbs
structions. We use the same patterns as above for
The lexc file for Laz verbal complex has 53 tags for these rules.
the morphemes encoding preverb, valency-related, We have also found flag diacritics useful for
mood, tense, person and number information and overwriting of valency-related vowels. In such
19 continuation classes which correspond to the af- constructions, we engage in two separate op-
fixes in the verbal complex as defined in Öztürk and erations/constructions at the same time such as
Pöchtrager (2011) also seen in Table 1, 2 and 3 causativisation and present perfect construction
with three additions — additive position for suffix both of which mark the verb with their specific
-ti, — question for -i, — participle for -eri. valency-related vowel in the -1 position. How-
We have mostly followed the description in ever, Laz allows overwriting causative o- to be over-
Öztürk and Pöchtrager (2011) when naming the tags written by applicative i- while keeping post ver-
and classes. bal causative markers -in or -ap. We have man-
The final combined lexc file also includes in- aged to form these constructions by also allowing
terjection, conjunction, negation, post-position and applicative i- (as well as causative o-) to have flag
pre-position lexicons with 5 more tags. @ . . @ which will let them through paths
4.1.3 Flag Diacritics defined with @ . . @ (Require the causative
feature to be present). These paths are naturally
Laz verb complex has required substantial use
those causative suffixes which do not allow struc-
of flag diacritics12 to solve problems like depen-
tures with related valency vowel otherwise.13
dent person marking, and causativisation or ap-
Other flag diacritics include which sets the re-
plicativisation processes, which require preverbal
lated feature as Negative. In our study, we use them
valency-related vowel marking as well as postver-
for subjunctive suffix and its special construction
bal causative markers at the same time. The hier-
with thematic suffixes. They do not normally occur
archical selection rule for person marking position
together but we can see that they do in constructions
preverbally among the arguments of the verb is eas-
with the imperfective suffix in between in (6) from
ily applied using flag diacritics. We have allowed
Öztürk and Pöchtrager (2011).
structures with 3rd person prefixes to only occur in
13
12 Additionally, Laz allows only intransitive bases to take the
Flag diacritics are used for feature-setting and feature- -in causative suffix, so we have also tried to use flag diacrit-
unification operations. They represent long-distance con- ics to differentiate between transitive and intransitive bases
straints for dependencies within a word (Beesley and Kart- by encoding the information onto verb itself. However, since
tunen, 2003). As a member of the same language family, Geor- we have automatically extracted verbs from the dictionary, we
gian also shows these kinds of long-distance dependencies in could not label all of them (2240 verb roots) as transitive or in-
verbal complex which are effectively treated again with this de- tansitive for this study, we ignore this differentiation and allow
vice in order to build a computational grammar for Georgian all verb roots to be able to bear both -in and -ap suffixes.
(Meurer, 2009).
874
(6) m-i-ťax-ap-ur-ť-a-s phonological alternation for valency-related vowels.
1- -break- . - - - - .3. . The alternation for valency-related vowels i/u- de-
‘Let say that I have broken it.’ pends on the preverbal person information, i- for 1st
If the subjunctive follows imperfective (sets the- and 2nd person, and -u for 3rd person.
matic suffix information as @ . . @), the The preverbs show a great amount of mor-
path allows subjunctive (normally disallows the- phophonological alternations in their final vowels,
matic suffixes as @ . . @) to follow thematic such as a, e and o. When they combine with overt
suffix after imperfective; therefore, we need to get person prefixes (consonants) together with valency
rid of @ . . @ setting by re-setting the same related vowels, final o and a become e or o and the
feature to as @ . . @ which will allow the change is not always predictable. They can also turn
structure to go through the path by taking non-past into v or can be dropped. Even though they may end
person markers that are set as @ . . @ disal- with the same vowel, the alternations can be differ-
lowing past tense constructions. ent when followed by the same sound; therefore; we
The substantive lexc file includes only one flag need to define different archiphonemes for the same
diacratic case which is to label nouns that take -lepe vowel. For example, the final o sound in exo- drops
plural marker instead of -pe. We label the noun root when it attaches to a verbal complex starting with a
with @ . . @ in order for it to be able to take sound but not the one in moyo-.
the path with @ . . @ label.
5 Results
4.2 twol File We have evaluated the morphological analyser by
The twolc (Two-level Compiler) formalism is used calculating the naïve coverage and doing error anal-
to define phonological and morphophonological al- ysis on randomly selected 100 tokens from the cor-
ternations. The twol file mostly includes person pus.
marking elements differing based on the follow-
ing consonant’s laryngeal property for verbal inflec- 5.1 Coverage
tion. We define rules/environments to account for The coverage is measured by calculating the number
morphophonological changes in the structure with of the tokens that receive at least one morphological
archiphonemes14 as given in Figure 1. analysis by the analyser. It should also be noted that
Laz also exhibits a phonological change in noun the tokens may have other analysis that is correct but
stems starting with n sound when preceded by ejec- not provided by the analyser even though they get at
tive p̆ -, the person prefix for 1. . The two conso- least one analysis.15 We have collected a corpus for
nants are combined and becomes m. We represent Pazar Laz which consists of 111,365 tokens men-
this as ejective p̆ turning to m and dropping the ini- tioned before in Section 3.2. The final morphologi-
tial n of the stem. Additionally, the final i sound cal analyser has 64.9% coverage over this corpus.
of noun stems becomes e when the stem is inflected
with plural marker. Corpus Tokens Coverage
"Assimilation of person prefix to p-" Pazar Laz 111,365 64.9%
{V}:p <=> _ >: Voiceless: ;
"Assimilation of person prefix to b-" Table 5: Naïve coverage of the analyser
{V}:ṕ <=> _ >: Ejectives: ;
Figure 1: Two two-level phonological rules for as- 5.2 Error Analysis
similation. The underspecified prefix archiphoneme We have looked at the tokens that are not covered by
{V} is restricted to surface either as p before voice- the morphological analyser. Randomly selected 100
less consonants or ṕ before ejective consonants. tokens has been examined and separated according
to their error type seen in Table 7. It should also be
We also observe a morphologically-conditioned 15
Unfortunately since there is no annotated corpus which can
14
An archiphoneme is used as a placeholder to be later be used as the ‘gold standard’, we were unable to calculate pre-
replaced with the appropriate sound determined by mor- cision and recall that could show the average accuracy of the
phophonological rules written in twol file. They are given in- analysis provided by the transducer.
side curly brackets.
875
Category Number of Stems Error Type Frequency Percentage
Noun 9417 Missing lexeme 41 37.9%
Verb 2240 Turkish word 35 32.4%
Adjective 745 Missing or erroneous
13 12.0%
Adverb 215 morphotactic rule
Pronoun 92 Typing errors 7 6.4%
Numeral 46 Loanwords 7 6.4%
Interjection 31 Missing or erroneous
5 4.6%
Postposition 29 Phonological rule
Conjunction 8 Total 108
Preposition 3
Negation 4
Table 7: Error analysis for randomly selected 100
Total 12,830 tokens
876
References
Kenneth R. Beesley and Lauri Karttunen. 2003. Finite
state morphology. CSLI Publications.
Steven Bird. 2009. Natural language processing
and linguistic fieldwork. Computational Linguistics
35(3):469–474.
İsmail Bucaklişi and Goichi Kojima. 2003. Laz Gram-
mar (Lazuri Grameri). Chiviyazilari.
Ciprian Gerstenberger, Niko Partanen, and Michael
Rießler. 2017. Instant annotations in ELAN corpora
of spoken and written Komi, an endangered language
of the Barents Sea region. In Proceedings of the 2nd
Workshop on the Use of Computational Methods in
the Study of Endangered Languages. Association for
Computational Linguistics, pages 57–66.
Belma Haznedar. 2018. The living Laz project: The cur-
rent status of the Laz language and Laz-speaking com-
munities in Turkey.
Nurdan Kavaklı. 2015. Novus Ortus: The awakening
of Laz language in Turkey. İdil Journal of Art and
Language 4(16):133–146.
René Lacroix. 2009. Description du dialecte laze
d’Arhavi (caucasique du sud, Turquie) Grammaire et
textes. Ph.D. thesis, Université Lumière Lyon.
Krister Linden, Miikka Silfverberg, Erik Axelson, Sam
Hardwick, and Tommi Pirinen. 2011. HFST–
Framework for Compiling and Applying Morpholo-
gies, volume 100 of Communications in Computer and
Information Science, pages 67–85.
Paul Meurer. 2009. A computational grammar for geor-
gian. In Peter Bosch, David Gabelaia, and Jérôme
Lang, editors, Logic, Language, and Computation.
Springer Berlin Heidelberg, Berlin, Heidelberg, pages
1–15.
Hammam Riza. 2008. Indigenous languages of Indone-
sia: Creating language resources for language preser-
vation. In Proceedings of the IJCNLP-08 Workshop on
NLP for Less Privileged Languages.
Balkız Öztürk and Markus A. Pöchtrager. 2011. Pazar
Laz. LINCOM.
Malgorzata Ćavar, Damir Ćavar, and Hilaria Cruz. 2016.
Endangered language documentation: Bootstrapping
a Chatino speech corpus, forced aligner, asr. In
Proceedings of the Tenth International Conference on
Language Resources and Evaluation (LREC 2016).
European Language Resources Association (ELRA).
877
Term Based Semantic Clusters for Very Short Text Classification
Jasper Paalman Shantanu Mullick
Jheronimus Academy of Data Science School of Industrial Engineering
j.v.paalman@tilburguniversity.edu Eindhoven University of Technology
s.mullick@tue.nl
878
Proceedings of Recent Advances in Natural Language Processing, pages 878–887,
Varna, Bulgaria, Sep 2–4 2019.
Text Class tions and suggestions for further work.
Vervoer Almere-Lille Travel expenses
Transport Almere-Lille
600GB SAS interne harde schijf Automation hardware 2 Related work
600GB SAS internal hard drive
Text or document classification is defined as the as-
Table 1: Example documents and classes signment of text sections or entire documents to a
predefined set of categories (Feldman and Sanger,
2007). For this purpose, algorithms process vari-
semantic variation. The inherent advantage of em- ous types of text representations which are used as
beddings in dealing with out-of-vocabulary words features for describing content. To our knowledge,
presents, at the same time, the disadvantage of invoice text classification has not been investigated
providing a text representation that does not fo- previously1. Work related to text classification of
cus on the importance of individual terms for the short texts has been applied for microblogs (Singh
classification. Conversely, measures of term oc- et al., 2016; Ren et al., 2016; Missier et al., 2016),
currence focus heavily on individual term impor- email subject classification (Alsmadi and Alhami,
tance but are very sensitive to variation. In very 2015) and spam detection (Bahgat et al., 2016).
short text and domain-specific applications, where Initial approaches to document content repre-
occurring terms are both strong indicators of the sentation used counts of term frequency and in-
respective text class, as well as abound in vari- verse document frequency, tf · idf (Salton and
ation, the preferred solution would combine em- Buckley, 1988), whereby frequently occurring
beddings to extracted terms. For example, for in- terms are assumed to represent document content
voices, each occurring term in an invoice descrip- and inverse document term frequency scores se-
tion is highly informative of the respective invoice lect the most distinctive terms for a given doc-
text class. Hence a method is required that not ument within a collection. Various subsequent
only focuses on such terms, but also leverages the approaches use variants of term occurrence mea-
flexibility of embeddings. Our proposed method sures with probabilities, such as χ2-test, log
Term Based Semantic Clusters (TBSeC) attempts likelihood (Dunning, 1993) and mutual informa-
to provide such a solution. The contribution of tion (Church and Hanks, 1990), or attempt to com-
this paper lies in (i) combining the advantages of bine statistical measures with various types of lin-
word embeddings with conventional term extrac- guistic and stop-word filters, so as to refine the
tion techniques (ii) apply our method in an appli- keyword results. Considerations regarding term
cation domain not previously investigated, namely ambiguity and variation also led to rule-based ap-
invoice text, which is characterised by specialised proaches (Jacquemin, 2001) and resource-based
terminology and very short, elliptical and/or un- approaches exploiting existing thesauri and lexica,
grammatical text, in a language that is morpholog- such as UMLS (Hliaoutakis et al., 2009), or Word-
ically richer than English and therefore posing an Net (Aggarwal et al., 2018). Knowledge poor sta-
additional challenge in statistical approaches. tistical approaches, such as Latent Semantic Anal-
TBSeC proposes a two-stage methodology. In ysis (Deerwester et al., 1990) and Latent Dirichlet
the first stage we use class-specific textual informa- Allocation (Blei et al., 2003) attempt to detect doc-
tion to build semantic concept clusters. Concept ument content in an unsupervised manner while
clusters are vector representations of strongly re- reducing the dimensionality of the feature space of
lated terms that are distinctive for a certain class. other bag-of-word approaches, but are also sensi-
In the second stage, we compute cluster similarity tive to sparse data and variation in short texts.
scores on generated concept clusters for a given The advent of large semantic resources in the
description. This serves as a ranking function that form of pre-trained word embeddings (Penning-
can be used in both unsupervised and supervised ton et al., 2014; Grave et al., 2018) gave rise to
learning tasks. a new line of approaches employing word embed-
The remainder of this paper is organized as fol- dings for document content representation, such as
lows: section 2 discusses related work; section Word2Vec (Mikolov et al., 2013), FastText (Bo-
3 elaborates on the TBSeC method; section 4 out- janowski et al., 2017), GloVe (Pennington et al.,
lines the experimental setup and section 5 discusses 1An approach reported by Bartoli et al. (2010) focuses on
the results. We conclude with our main observa- image features of scanned invoices rather than the invoice text.
879
2014), and ELMo (Peters et al., 2018). Within this 3.1 Concept clustering
line of approaches, methods have also been devel-
Concept clustering starts by extracting distinctive
oped using word embeddings specifically for docu-
terms, particular to a class2. Distinctive terms
ment classification, such as task-oriented word em-
are extracted based on normalized term frequency.
beddings (Liu et al., 2018) and word-sense based
For a given class, word embeddings belonging to
embeddings (Jin et al., 2016). Embedding models
found terms are used to form numerous clusters.
encoding words in documents or document sec-
Hence, for each class multiple clusters are created
tions have also been developed, such as Doc2vec
and each cluster can be seen as a group of terms
(Le and Mikolov, 2014), Infersent (Conneau et al.,
that are closely related. Each cluster is transformed
2017), Skip-thought (Kiros et al., 2015) and Fast-
into a concept cluster vector by taking the respec-
Sent (Hill et al., 2016). Such type of embeddings
tive word embeddings of terms and computing the
can be employed to calculate the semantic similar-
mean over each dimension. This process is illus-
ity between texts, but the risk is that intricate word-
trated in Figure 1, with actual examples included
specific semantic relationships are lost. Meth-
and word embeddings indicated as vectors with
ods originating from text similarity research using
dots.
word rather than document embeddings, such as
Kusner et al. (2015); Kenter and De Rijke (2015); Specifically, the method works as shown in Al-
De Boom et al. (2016) attempt to address this issue. gorithm 1 for n distinct classes, k most frequent
Embeddings have been also used for keyphrase ex- terms to be incorporated as cluster and normalized
traction via supervised (Mahata et al., 2018) and term frequency (i.e., Bag-of-words, Bij ) threshold
unsupervised (Bennani-Smires et al., 2018) ap- t. A more detailed description is given underneath
proaches. the algorithm. Algorithm line numbers refer to
steps as denoted in Figure 1.
Finally, the problem of variation and data spar- Terms that provide a clear distinctive value to
sity with very short texts has been addressed in a class are retrieved using term frequency and
the past with query expansion approaches (Vech- L1 normalization over rows and columns. Terms
tomova, 2009). For text similarity purposes query not appearing in the vocabulary of the embedding
expansion techniques have been used for document model, or having a score below the normalized
term augmentation, exploiting relevant search re- threshold t are filtered out. Word embeddings for
sults and a matrix representation (Sahami and Heil- these selected terms are employed to create concept
man, 2006; Abhishek and Hosanagar, 2007) or a clusters for each class using the DBSCAN cluster-
combination of search results page count differ- ing model (Ester et al., 1996). Terms can either
ence and lexico-syntactic patterns derived from be included in a multi-term cluster through DB-
text snippets (Bollegala et al., 2007). SCAN clustering or can be added as a single-term
cluster when occurring frequently enough based
on k. Ultimately, for each class i, concept clusters
are created as a single vector equal to the averaged
embedding of included terms.
3 The TBSeC methodology
3.2 Semantic cluster similarity
We adapt the similarity measure by Kenter and
Our proposed method, TBSeC, consists of two
De Rijke (2015), which is based on the BM25
stages: In the first stage we use class-specific tex-
framework (Robertson et al., 2009), to propose our
tual information to build semantic concept clus-
semantic cluster similarity measure. We combine
ters. Concept clusters are vector representations of
idf scoring with word by word cosine similarity to
strongly related terms that are distinctive for a cer-
calculate a weighted similarity score.
tain class. In the second stage, we compute cluster
The Kenter and De Rijke (2015) function for
similarity scores on generated concept clusters for
calculating semantic text similarity between two
a given description, thereby forming a semantic
sentences sl and ss is as defined as follows:
feature space. This serves as a ranking function
that can be used in both unsupervised and super- 2In the case of invoice classification, each class refers to a
vised learning tasks. specific expense type.
880
Figure 1: Concept clustering process diagram. Upper panel: Preparing normalized bag of words using
descriptions by class. Lower panel: Employing found values to create concept clusters for a given class
Parameters k1 and b are inherited from the BM25 differentiating is not essential. Third, terms that
framework and serve as smoothing parameters. Pa- don’t hold considerable semantic importance are
rameter k1 influences what semantic text similarity not likely to steer the score towards an incorrect
value is approached asymptotically, thereby limit- class. Each concept cluster is created to serve as a
ing the influence that semantic term similarity can distinctive concept, thus making it unrealistic that
have. Parameter b provides a degree of importance unimportant terms will relate to it well. Fourth,
|ss |
to sentence length ratio avgsl , comparing sentence terms in descriptions are subject to frequent mis-
length to the average sentence length avgsl of sen- spellings and personal abbreviations, making idf
tences being ranked. The function sem returns the scores inherently unreliable in this setting.
semantic term similarity of term w with respect to Based on the changes to Equation 1 discussed
text s, as follows: above, our function for calculating semantic cluster
similarity fscs between text s and concept cluster
c is defined as follows:
sem(w, s) = max
0
fsem (w, w0 ) (2)
w ∈s
881
The semantic threshold semth serves as a way
to add a semantic similarity bound above which
it will be presumed to hold importance. Param-
eter semth achieves that when fsem (w, c) is un-
der the set threshold value, that sem(w, c) equals
Algorithm 1 Concept Clustering 0. Squaring fsem (w, c) through semsq increases
Parameters:
term importance of terms that are a near match and
k - No. of most frequent terms to be incorporated as cluster lowers importance of terms that match to a lesser
t - Normalized bag of words threshold extend. Squaring fsem (w, c) therefore promotes
Input:
Di , i ∈ {1, . . . , n} - Merged descriptions for each class the divergence of semantic similarity scores.
Output: Semantic cluster similarity fscs produces fea-
Ci , i ∈ {1, . . . , n} - Concept clusters for each class tures for use in supervised learning applications,
1 Bij ← Calculate bag of words using Di , but the initial performance of TBSeC is measured
with found vocabulary set V , without a predictive model. For this reason, a sim-
i ∈ {1, . . . , n}, j ∈ {1, . . . |V |} ilarity score for each class is required in order to
Ci ← Instantiate empty concept cluster array,
with i ∈ {1, . . . , n} rank the classes. This is calculated as scs for each
class i, by extracting the maximum score over all
for i ∈ {1, . . . , n} do
2 Bi∗ ← L1 normalize Bi∗ . Normalize by class
concept clusters c ∈ Ci (see Ci in Algorithm 1):
end for
for j ∈ {1, . . . |V |} do scs(s, Ci ) = max fscs (s, c) (5)
3 B∗j ← L1 normalize B∗j . Normalize by term c∈Ci
end for
for i ∈ {1, . . . , n} do . For each class
4 Retrieve terms in vocabulary with a score > t 4 Experimental setup
and occur more than once
terms ← Array of found terms that appear in This section covers data description, data process-
word embedding vocabulary
5 emb ← Respective word embeddings of terms ing and the experimental set-up for our method.
Ci ← Create Clusters(emb, k, terms)
end for 4.1 Data
return C
Our invoice data originate from an auditing com-
function Create Clusters(emb, k, terms) pany. In the data, as illustrated in the examples
Instantiate DBSCAN clustering model with in Table 1, each class refers to a particular type of
eps, min_samples and
metric = cosine similarity expenses. Available data is accessed from a data
Fit emb on DBSCAN model directory, where each file is specific to a client.
6 clusters ← formed clusters as collections of For our purposes, only the invoice description and
word embeddings
for term in k most frequent terms do class assignment are relevant. The volume of the
if term not used in clusters then entire data directory amounts to approximately 1.5
e ← word embedding of term million instances. There is a total of 111 unique
7 cluster.append([e])
end if classes to which assignments are made. The five
end for classes that are least represented have 24, 106, 178,
conceptsc ← empty concept cluster array
with c ∈ {1, . . . , |clusters|}
418 and 452 entries respectively. Invoice descrip-
for c ∈ {1, . . . , |clusters|} do tions on average contain 2.80 terms, with a standard
8 conceptsc ← coordinate mean over each deviation of 1.55.
dimension
end for 4.2 Word embeddings
return concepts
end function Pre-trained Dutch FastText word embeddings3 are
used for sentence embedding construction and
for use in semantic similarity computations (Bo-
janowski et al., 2017). The FastText embedding
model was trained on Dutch Wikipedia.
3https://github.com/facebookresearch/
fastText/blob/master/docs/pretrained-vectors.
md
882
4.3 Data processing Parameter Distinct values
k1 [1.2, 1.6, 2.0]
Descriptions are processed using a procedure sim- semsq [true, false]
semth [0, 0.1, 0.2, 0.3, 0.4]
ilar to the one used in training the FastText model.
Special characters are replaced with a whitespace, Table 2: fscs parameter settings
stopwords in both the English and Dutch language
are dropped, digits are removed and finally terms
are retrieved by splitting on any sequence of uni- incorporated as cluster.
code whitespace characters. When creating valida- Validation is performed on a dataset with ap-
tion sets special care is taken to remove duplicates proximately 5,000 entries. Performance differ-
and to include data from all individual clients and ences for fscs are evaluated using three perfor-
all classes in a randomized manner. As a result, mance measures: (1) accuracy, (2) ranking loss
largely balanced validation sets are formed with and (3) standardized score. Accuracy is calcu-
data from various sources. lated by picking the class with the highest semantic
similarity score as predicted class and finding the
4.4 Learning algorithm
percentage of correctly classified instances. Rank-
During supervised learning a Support Vector Ma- ing loss is calculated by obtaining the rank of the
chine4 is used with regularization parameter C = true class. The standardized score is calculated by
0.1 and a linear kernel. This classifier performed standardizing all scores for a given instance and
best when compared to other feasible classifiers retrieving the score of the true class. By standard-
(e.g. random forest), given a local working mem- izing it can be observed how the score for the true
ory bounded set-up, and allows for the use of sparse class is positioned against all other scores. Ulti-
matrices. Regularization parameter C regulates mately, the objective is to maximize the accuracy
the importance of focusing on correctly classifying and standardized score and to minimize ranking
training samples in favor of realizing a hyperplane loss. Afterwards, we investigate the influence of
with a large minimum margin. A high C can lead parameter k for concept cluster construction on the
to overfitting, a low C can lead to the inability to basis of accuracy and dimension size. Accuracy is
learn meaningful decision boundaries. We set C to calculated as an unsupervised score as well as a 5-
0.1 since it appears to offer a good balance on the fold cross validated supervised score. Both scoring
basis of the main validation set in terms of limited methods use best values for k1 , semth and semsq
running time and general performance. which are found in the previous step. Results are
compared to determine an appropriate value for pa-
4.5 Parameter tuning rameter k for use of features in a predictive model.
Prior to including our framework in a supervised
learning task, we optimize the parameters (see sec- 4.6 Invoice classification
tion 5.1 for results). We construct an initial set of We use the proposed semantic cluster similarity
concept clusters using preset values k = 5 and matching method, to measure performance in a
t = 0.8. Parameters are set such that the model classification task. We compare the performance
offers a well-performing baseline with low chances to existing methods and we test whether combining
of overfitting. Initial concept clusters are used to methods leads to an increase in performance.
tune semantic cluster similarity parameters. This The parameters for generation of semantic clus-
order of parameter tuning is chosen, because it is ter similarity scores are set in accordance with val-
relatively straightforward to pick sensible values ues obtained during parameter tuning.
for k and t, as opposed to fscs hyperparameters. For validation, a dataset containing approxi-
Table 2 lists the attempted combinations of param- mately 20,000 entries is used. The data is used to
eter settings for fscs . perform 5-fold cross validation to determine over-
After parameter tuning for semantic cluster simi- all performance. For all tests, we chose to use ac-
larity, employed concept clusters are reconsidered. curacy as performance quality measurement. Only
Values in the range from 5 to 50 with a step size one out of 111 classes shows significant imbalance,
of 5 are attempted for the k most frequent terms which for testing overall method performance is
4https://scikit-learn.org/stable/modules/ deemed negligible. Consequently, no balancing is
generated/sklearn.svm.LinearSVC.html performed and no alternative quality measure, such
883
Method Parameters Dimension size
tf · idf - 117,766
tf - 117,766
LSA # components: 100 100
LDA # topics: 200 200
FastText - 300
P-mean P-values: {−∞, 1, 2, 3, ∞} 1500
884
Methods test acc test sd train acc train sd techniques leads to surprising results. P-mean per-
Benchmark
tf · idf 43.93 0.89 82.79 0.07 forms well as sole feature, but doesn’t yield better
tf 44.85 0.67 84.84 0.12 results when combined with other techniques. In
LSA 7.89 0.35 9.16 0.20 contrast, FastText does pair well with other tech-
LDA 11.78 0.37 14.15 0.15
TBSeC 36.03 0.60 46.96 0.26 niques and improves performance levels. More-
FastText 32.62 0.79 41.26 0.24 over, when TBSeC is used performance is even bet-
P-mean 39.78 0.63 90.33 0.1 ter. Term frequency in combination with both TB-
Feature com-
binations SeC and FastText does also improve performance,
tf & TBSeC 47.83 0.54 86.17 0.13 although slightly. TBSeC and FastText are more
tf & FastText 47.07 0.48 85.81 0.14 likely to be complementary as soon as TBSeC en-
tf & TBSeC 47.86 0.42 86.38 0.10
& FastText compasses lower number of concept clusters, but
this doesn’t guarantee better overall performance.
Table 5: Invoice classification results In the current configuration it manages to perform
better, but the difference is not major. The tendency
to overfit for tf · idf , term frequency and P-mean
goal of TBSeC is relating input to broader con-
is a likely cause for some of the other unsuccessful
cepts, which is why a relatively moderate value for
feature combinations.
k is preferred.
6 Conclusion
5.2 Invoice classification
A new feature generation framework TBSeC was
In this section, we discuss the results of our method
presented that is suited to the prediction of well
against state-of-the art benchmark methods in a su-
defined classes on the basis of very short texts
pervised invoice classification task. The results,
(2.8 words on average). Generated features were
consisting of cross validated accuracy scores with
proven to be able to function well independently
standard deviation (sd), are shown in Table 5 under
and jointly with traditional feature generation tech-
header ‘Benchmark’. The fact that term frequency
niques. Performance and reliability was improved
has comparable performance to tf · idf reinforces
by pairing multiple disjoint feature generation tech-
the notion that all terms within an invoice descrip-
niques, including TBSeC. A combination of highly
tion are important to take into account and that
specific features with more flexible ones was found
term-specific weighting has limited value. More-
to lead to the best results. Combinations of fea-
over, techniques that are concerned with dimen-
tures were found to reach a bound in effectiveness,
sionality reduction (LSA, LDA) perform worse,
highlighting that methods ultimately start imped-
arguably because they truncate a large amount of
ing each other. Businesses can use our method to
information, most of which should have been re-
derive actionable insights from online user gener-
tained. Sentence embeddings and our method TB-
ated content such as firm-specific tweets, online re-
SeC perform relatively well, with accuracy scores
views and customer chats logs. Future work could
nearing performance levels of tf · idf and term
test TBSeC on larger texts, since it offers more
frequency. Furthermore, a large feature space
room to differentiate from sentence embeddings.
(1500D) that is achieved with P-mean embeddings
appears to have a positive influence on the amount
of information that is contained. It is also found References
that methods tf · idf , term frequency and P-mean
Vibhanshu Abhishek and Kartik Hosanagar. 2007. Key-
have a tendency to overfit on the data, having cross- word generation for search engine advertising using
validation training accuracy scores of over 80%. In semantic similarity between terms. In Proceedings
comparison, TBSeC and FastText have training ac- of the ninth international conference on Electronic
curacy scores closer to test accuracy scores. commerce. ACM, pages 89–94.
Combinations of techniques are attempted next Ayush Aggarwal, Chhavi Sharma, Minni Jain, and
to improve performance. The feature combinations Amita Jain. 2018. Semi supervised graph based key-
are formed by concatenating the feature spaces word extraction using lexical chains and centrality
of each method. The most successful combina- measures. Computación y Sistemas 22(4).
tions are highlighted in Table 5 under header ‘Fea- Izzat Alsmadi and Ikdam Alhami. 2015. Clustering
ture combinations’. Combining feature generation and classification of email contents. Journal of King
885
Saud University-Computer and Information Sciences Scott Deerwester, Susan T Dumais, George W Furnas,
27(1):46–57. Thomas K Landauer, and Richard Harshman. 1990.
Indexing by latent semantic analysis. Journal of the
Isabelle Augenstein, Mrinal Das, Sebastian Riedel, American society for information science 41(6):391–
Lakshmi Vikraman, and Andrew McCallum. 2017. 407.
SemEval 2017 task 10: ScienceIE - extracting
keyphrases and relations from scientific publica- Ted Dunning. 1993. Accurate methods for the statistics
tions. In Proceedings of the 11th International Work- of surprise and coincidence. Computational Linguis-
shop on Semantic Evaluation (SemEval-2017). As- tics 19(1):61–74.
sociation for Computational Linguistics, Vancouver,
Canada, pages 546–555. Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei
Xu, et al. 1996. A density-based algorithm for
Eman M Bahgat, Sherine Rady, and Walaa Gad. discovering clusters in large spatial databases with
2016. An e-mail filtering approach using classifi- noise. In Kdd. volume 96, pages 226–231.
cation techniques. In The 1st International Confer-
ence on Advanced Intelligent System and Informat- Ronen Feldman and James Sanger. 2007. The text min-
ics (AISI2015), November 28-30, 2015, Beni Suef, ing handbook : advanced approaches in analyzing
Egypt. Springer, pages 321–331. unstructured data. Cambridge University Press.
Marco Baroni and Alessandro Lenci. 2010. Dis- Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Ar-
tributional memory: A general framework for mand Joulin, and Tomas Mikolov. 2018. Learning
corpus-based semantics. Computational Linguistics word vectors for 157 languages. In Proceedings of
36(4):673–721. the Eleventh International Conference on Language
Resources and Evaluation, LREC 2018, Miyazaki,
Alberto Bartoli, Giorgio Davanzo, Eric Medvet, and Japan, May 7-12, 2018..
Enrico Sorio. 2010. Improving features extraction
for supervised invoice classification. In Artificial Zellig S Harris. 1954. Distributional structure. Word
Intelligence and Applications. MH Hamza. 10(2-3):146–162.
Kamil Bennani-Smires, Claudiu Musat, Andreea Hoss- Felix Hill, Kyunghyun Cho, and Anna Korhonen.
mann, Michael Baeriswyl, and Martin Jaggi. 2018. 2016. Learning distributed representations of
Simple unsupervised keyphrase extraction using sentences from unlabelled data. arXiv preprint
sentence embeddings. In Proceedings of the arXiv:1602.03483 .
22nd Conference on Computational Natural Lan-
guage Learning. Association for Computational Angelos Hliaoutakis, Kalliopi Zervanou, and Euripides
Linguistics, Brussels, Belgium, pages 221–229. G. M. Petrakis. 2009. The AMTEx approach in the
https://www.aclweb.org/anthology/K18-1022. medical document indexing and retrieval application.
Data and Knowledge Engineering 68(3):380–392.
David M Blei, Andrew Y Ng, and Michael I Jordan.
2003. Latent dirichlet allocation. Journal of ma- Christian Jacquemin. 2001. Spotting and Discovering
chine Learning research 3(Jan):993–1022. Terms Through Natural Language Processing. MIT
Piotr Bojanowski, Edouard Grave, Armand Joulin, and Press.
Tomas Mikolov. 2017. Enriching word vectors with Peng Jin, Yue Zhang, Xingyuan Chen, and Yunqing Xia.
subword information. Transactions of the Associa- 2016. Bag-of-embeddings for text classification. In
tion for Computational Linguistics 5:135–146. IJCAI. volume 16, pages 2824–2830.
Danushka Bollegala, Yutaka Matsuo, and Mitsuru
Prathusha Kameswara Sarma. 2018. Learning word
Ishizuka. 2007. Measuring semantic similarity be-
embeddings for data sparse and sentiment rich data
tween words using web search engines. www 7:757–
sets. In Proceedings of the 2018 Conference of the
766.
North American Chapter of the Association for Com-
Kenneth Ward Church and Patrick Hanks. 1990. Word putational Linguistics: Student Research Workshop.
association norms, mutual information, and lexicog- Association for Computational Linguistics, New Or-
raphy. Computational Linguistics 16(1):22–29. leans, Louisiana, USA, pages 46–53.
Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Tom Kenter and Maarten De Rijke. 2015. Short text
Barrault, and Antoine Bordes. 2017. Supervised similarity with word embeddings. In Proceedings
learning of universal sentence representations from of the 24th ACM international on conference on in-
natural language inference data. arXiv preprint formation and knowledge management. ACM, pages
arXiv:1705.02364 . 1411–1420.
Cedric De Boom, Steven Van Canneyt, Thomas De- Ryan Kiros, Yukun Zhu, Ruslan R Salakhutdinov,
meester, and Bart Dhoedt. 2016. Representation Richard Zemel, Raquel Urtasun, Antonio Torralba,
learning for very short texts using weighted word and Sanja Fidler. 2015. Skip-thought vectors. In
embedding aggregation. Pattern Recognition Letters Advances in neural information processing systems.
80:150–156. pages 3294–3302.
886
Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Mehran Sahami and Timothy D Heilman. 2006. A web-
Weinberger. 2015. From word embeddings to docu- based kernel function for measuring the similarity
ment distances. In International Conference on Ma- of short text snippets. In Proceedings of the 15th
chine Learning. pages 957–966. international conference on World Wide Web. AcM,
pages 377–386.
Quoc Le and Tomas Mikolov. 2014. Distributed repre-
sentations of sentences and documents. In Inter- Gerard Salton and Christopher Buckley. 1988. Term-
national Conference on Machine Learning. pages weighting approaches in automatic text retrieval.
1188–1196. Information processing & management 24(5):513–
523.
Qian Liu, Heyan Huang, Yang Gao, Xiaochi Wei, Yuxin
Monika Singh, Divya Bansal, and Sanjeev Sofat. 2016.
Tian, and Luyang Liu. 2018. Task-oriented word
Behavioral analysis and classification of spammers
embedding for text classification. In Proceedings of
distributing pornographic content in social media.
the 27th International Conference on Computational
Social Network Analysis and Mining 6(1):41.
Linguistics. pages 2023–2032.
Ge Song, Yunming Ye, Xiaolin Du, Xiaohui Huang,
Debanjan Mahata, John Kuriakose, Rajiv Ratn Shah, and Shifu Bie. 2014. Short text classification: A
and Roger Zimmermann. 2018. Key2vec: Auto- survey. Journal of multimedia 9(5):635–644.
matic ranked keyphrase extraction from scientific ar-
ticles using phrase embeddings. In Proceedings of Olga Vechtomova. 2009. Query Expansion for Infor-
the 2018 Conference of the North American Chap- mation Retrieval, Springer US, Boston, MA, pages
ter of the Association for Computational Linguistics: 2254–2257.
Human Language Technologies, Volume 2 (Short Pa-
pers). pages 634–639. Kalliopi Zervanou, Elias Iosif, and Alexandros Potami-
anos. 2014. Word semantic similarity for morpho-
Tomas Mikolov, Kai Chen, Greg Corrado, and Jef- logically rich languages. In Proceedings of the Ninth
frey Dean. 2013. Efficient estimation of word International Conference on Language Resources
representations in vector space. arXiv preprint and Evaluation, LREC 2014, Reykjavik, Iceland,
arXiv:1301.3781 . May 26-31, 2014.. pages 1642–1648.
887
Quotation Detection and Classification with a Corpus-Agnostic Model
888
Proceedings of Recent Advances in Natural Language Processing, pages 888–894,
Varna, Bulgaria, Sep 2–4 2019.
1. The models inherit the corpora’s structural (based on the surface form) between direct quo-
and theoretical assumptions, such as the pres- tations (starting and ending with quotation marks),
ence of a cue assumed by models for the Pareti indirect quotations (without any quotation marks),
et al. (2013) corpus. and mixed quotations (everything else).
Pareti model. Pareti (2015), an extension of
2. The models typically include domain-specific Pareti et al. (2013), presents a pipeline architecture
features and knowledge sources that happened for quotation annotation. It first applies a k-NN
to be available from the corpus, such as lists classifier to identify quotation cues within the cor-
of likely cue verbs or syntactic realizations of pus. Then, a linear-chain conditional random field
quotations. (CRF) is used to identify quotation spans in the
vicinity of each cue. The Pareti model builds on
This corpus dependence amounts to conceptual corpus-specific knowledge, including lists of cue
overfitting: while it leads to better fit for the orig- verbs, and handcrafted features sensitive to punctu-
inal corpus, models are not transferable to new ation conventions in English newswire text.
domains and analysis schemes. In other words, it Scheible model. Scheible et al. (2016) retain the
leads to serious fragmentation. pipeline architecture of Pareti (2015) and its feature
In this paper, we develop and evaluate a corpus- set, but replace the components. Cue annotation is
agnostic neural model architecture for automatic performed with an averaged perceptron. More im-
quotation recognition that makes as few assump- portantly, they replace quotation annotation proper
tions as possible about the corpus to be modelled with a sampling-based procedure: a perceptron
but is still expressive enough to deal with the chal- samples tokens as likely span boundaries, which
lenge of recognizing quotation spans of essentially are then combined into complete quotation spans,
arbitrary length (Scheible et al., 2016). In this re- using a semi-Markov model.
spect, we see our study as a step towards transfer
learning (Pan and Yang, 2009) and task-agnostic 2.2 STOP Dataset
learning (Hashimoto et al., 2017). We find that our Semino and Short (2004) presents a corpus-based
model can perform reasonably well across corpora ontology of quotations in English text. It intro-
differing in genre, language, and structure. duces two dimensions: (a), speech vs. thought
vs. writing; and (b), direct vs. indirect vs. free
2 Related Work: Datasets and Models indirect vs. reported, yielding a Cartesian product
of twelve quotation subclasses. These are used to
We now review the state of the art in automatic annotate the Speech, Thought, and Writing Presen-
quotation annotation, describing the three major tation corpus (STOP). It comprises 120 sections,
quotation corpora for English and German and the split evenly across three genres (fiction, newspaper,
corresponding models. We exclude corpora that and biographies), of about 2,000 words each (To-
focus on one specific quotation subtype such as the tal size: 250,000 tokens; 8,000 quotations). The
Columbia Speech Attribution corpus (Elson and corpus has no linguistic annotation: the only fea-
McKeown, 2010) which only covers direct speech. tures available are words’ surface forms. To our
knowledge, there are no models for this dataset.
2.1 PARC Dataset
Dataset. The Penn Attribution Relation Corpus 2.3 Redewiedergabe Dataset
(Pareti, 2015), version 3 (PARC 3) is a subset of the Dataset The Redewiedergabe (’reported speech’)
Penn Treebank, annotated with quotations and at- corpus (RWG) (Brunner, 2013) is a corpus of Ger-
tribution relations. It consists of English newswire man narrative text, comprising thirteen public-
text from the Wall Street Journal. Each attribu- domain German narratives (1787–1913). The quo-
tion relation consists of a cue, optionally a source tation annotations in RWG adopt the scheme by
(speaker), and content (quotation span), all marked Semino and Short (2004) and distinguish direct, in-
as text spans. As part of the Penn Treebank, PARC 3 direct, free indirect, and reported variants of speech,
provides manually annotated tokenization, POS thought, and writing. The total size of the corpus
tags, lemmas, and constituency parses. is 57.000 tokens, and 17.000 quotation spans.
Quotation spans are not labeled with more spe- Unlike STOP, RWG contains some linguistic in-
cific types, but PARC 3 distinguishes informally formation, namely POS tags, lemmas, and mor-
889
“ Hello , ” she said .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
+ + + + + + +
Figure 1: The NQD architecture. Tokens are represented as a bag of feature embeddings, and each token sequence
is processed by a 2-layer bi-LSTM network, before a max-entropy classifier labels each token.
890
This approach to feature representation allows our opposed to entire documents. We use 10-fold cross
model to work with corpora with arbitrary types validation again, randomly partitioning all corpus
of token-level features. In the simplest case, when sentences into 10 subsets. We use GloVe embed-
only raw text is present in the corpus, each token dings pre-trained on the German Wikipedia.1
is given a single feature for that token’s word. If Evaluation. Previous studies on PARC 3 adopted
other token-level features are present, such as POS- an exact span match setting, i.e., only those pre-
tags, lemmas, or even parse tree information, these dicted spans that exactly match a gold standard
can be incorporated as additional feature vectors, span count as true positives. We report precision,
without requiring any changes to the model archi- recall, and F1 in this setting for PARC 3 and STOP.
tecture. Feature vectors can also be initialized to For RWG, we report the sentence-level accuracy
pre-trained representations (e.g. word embeddings) used by Brunner (2013). In this mode, we train and
when these are available, or initialized randomly predict with our model as before, but for evalua-
and learned when they are not. Section 2 describes tion we just record whether the model predicts the
in detail which features are used for the corpora we presence of a quotation type in a sentence.
experiment with.
4.2 Results
4 Experimental Evaluation PARC 3. The results in Table 1 show that NQD can-
We now train and test NQD on the three corpora not beat the performance of Scheible et al. (2016),
and compare against the state-of-the-art. but does almost as well as Pareti et al. (2013).
Given that our model is substantially simpler than
4.1 Experimental Setup either of these two (both include a special cue clas-
sifier, dictionaries, etc.), we see this as a success.
PARC 3. For PARC 3, we train a single classifier
Our model is competitive with the Scheible et al.
on the quote content spans and ignore the cue and
model with regard to recall, but shows subpar preci-
source spans. As features, we use token surface
sion for all quotation types, indicating a remaining
forms, lemmas, POS tags, as well as, for each token,
weakness in the input encoding: for direct quota-
the bags of constituents that start with, end with,
tions, quote characters should provide strong indi-
and contain it. These features are a subset of the
cators for quotation boundaries.
features used by Scheible et al. (2016) and Pareti
(2015), and like these studies, we use gold standard Note that these results, as well as the earlier
annotation. We initialize the features for word sur- studies (Pareti et al., 2013; Scheible et al., 2016),
face forms with the default GloVe Wikipedia word use unrealistic gold standard features. Therefore,
embeddings (Pennington et al., 2014). Our model we ran a second version of NQD using only word
makes predictions on entire documents at a time. features, but no tags or structural information. The
We use performance on the corpus’s development model is clearly worse, but still surprisingly good at
set to guide early stopping during training, and we 61% F1 . Not surprisingly, we see the highest drop
evaluate on the corpus’s test set. for indirect quotations, which are most sensitive
to syntactic structure. This indicates that NQD
STOP. For STOP , we train four classifiers for
does a reasonable job in a setting that is realistic in
the four quote types (direct, indirect, free indirect,
general, and particularly so for non-standard corpus
reported). We train and evaluate our model on a
varieties (e.g., historical and literary corpora) that
per-document basis as for PARC 3. We use word
are often used in Digital Humanities.
surface forms (and their GloVe embeddings) as
features, we used no other features in this model. STOP. To our knowledge, the results in Table 2
As the corpus contains no held-out development are the first modeling results on STOP. In com-
or test sets, we used 10-fold cross validation to parison to PARC 3, the results are noticeably lower.
evaluate our model, using 8 folds for training, 1 for It is still the case that direct quotations are easi-
development, as 1 as for testing in each iteration. est to find, but their F1 is somewhat lower than in
PARC 3. Indirect quotations are much more diffi-
RWG . For RWG , we adopt the same four-
classifier setting as for STOP, using the word, cult, and free indirect quotations essentially impos-
lemma, POS, and morphological features available. sible. This involves multiple factors: (a) STOP is
For the sake of comparability with (Brunner, 2013), 1
Available from deepset at https://deepset.ai/
we train and evaluate on individual sentences, as german-word-embeddings
891
Overall Direct Indirect Mixed
Model Features
Rec Prec F1 Rec Prec F1 Rec Prec F1 Rec Prec F1
Pareti et al. (2013) word, syn, domain 63 80 71 88 94 91 56 78 65 60 67 63
Scheible et al. (2016) word, syn, domain 71 79 75 93 94 94 64 73 69 68 81 74
NQD word, syn 71 67 69 94 82 88 64 64 64 70 59 64
NQD word 61 61 61 90 84 87 53 56 54 60 54 57
892
unannotated quotation in the following sentence: are almost all other cases of quotation spans; verb
phrases, on the other hand, can become arbitrar-
(3) (Britain’s retail price index rose 0.7% in ily complex. In the case above, the segmentation
September from August and was up 7.6% problems are exacerbated by the fact that the noun
for the year), the Central Statistical Office phrase quotation span occurs in a complex syntac-
said. tic environment involving coordination.
Outside of these cases, proper false positives seem
to be rare. Many of the false positives we found 6 Conclusion
were boundary mismatches, discussed separately In this paper, we have argued that existing models
below. of automatic quotation annotation suffer from the
5.2 False Negatives tight relation between corpus annotation and model
properties in particular in terms of model reusabil-
Among the false negatives we analyzed, we found
ity. As an alternative, we have presented a general
that the model is most likely to miss “tricky” quota-
neural architecture, NQD, that can be trained “as
tions that are unusual in their grammatical structure.
is” on various corpora that differ in terms of genre,
In particular, it tends to miss a class of quotations
structure, and language. While the model does not
that are expressed as short noun phrases or adjecti-
reach the state of the art on any particular corpus,
val phrases embedded within a non-quotation sen-
it performs close to it on all of them. Notably,
tence such as
the model is also able to deal relatively graciously
(4) Mandela, considered [the most prominent with the absence of linguistic information. We will
leader of the ANC] remains in prison. But release an implementation with pre-trained models.
[his release within the next few months] is As NQD makes independent predictions for
widely expected. each token, it cannot model correlations and mutual
exclusions between labels, and there is no guaran-
According to the PARC guidelines, these are cases tee for well-formed output class sequences. We in-
of quotations since they are attributable statements, vestigated a number of extensions, including linear-
but they are difficult for the model to retrieve since chain CRF layers that are effective for Named En-
they are hard to distinguish from “non-quotation” tity Recognition (Lample et al., 2016), but did not
nominal phrases – in particular in cases like this obtain improvements. We believe this is due to
one, where there are not even overly realized speak- the unbounded length of quotation spans which is
ers. In STOP and RWG, these cases might arguably challenging for CRFs (Scheible et al., 2016).
not even be annotated as quotations. The overall greatest challenge that NQD faces
5.3 Boundary Mismatches is data scarcity — all existing corpora with manual
annotation are small, and our results show consis-
A large proportion of the errors of NQD are bound- tently bad performance for infrequent quotation
ary errors, where the model identifies the presence types. In this situation, transfer learning seems
of a quotation, but fails to identify its exact bound- like a natural proposition, and our model makes it
aries. This can happen when our model correctly possible for the first time to apply straightforward
predicts one quotation boundary, but not the other. transfer learning to quotation annotation. In future
For example, in the following sentence, our clas- work, we will explore this direction.
sifier identified the first quotation’s beginning, but
not its end (it also failed to identify the second
quotation entirely – a false negative): References
(5) He reiterated ([his opposition to such fund- Apoorv Agarwal, Augusto Corvalan, Jacob Jensen, and
ing], but expressed [hope of a compro- Owen Rambow. 2012. Social network analysis of
mise]. Alice in Wonderland. In Proceedings of the NAACL-
HLT 2012 Workshop on Computational Linguistics
for Literature, pages 88–96, Montréal, Canada. As-
This type of error occurs both for noun phrases
sociation for Computational Linguistics.
and verb phrases and embedded sentences, but for
different reasons: noun phrases are difficult to rec- Mariana S. C. Almeida, Miguel B. Almeida, and André
ognize since they are not marked by punctuation as F. T. Martins. 2014. A joint model for quotation
893
attribution and coreference resolution. In Proceed- Silvia Pareti. 2015. Attribution: A Computational Ap-
ings of the 14th Conference of the European Chap- proach. Ph.D. thesis, University of Edinburgh.
ter of the Association for Computational Linguis-
tics, pages 39–48. Association for Computational Silvia Pareti, Tim O’Keefe, Ioannis Konstas, James R.
Linguistics. Curran, and Irena Koprinska. 2013. Automatically
detecting and attributing indirect quotations. In Pro-
Annelen Brunner. 2013. Automatic recognition of ceedings of the 2013 Conference on Empirical Meth-
speech, thought, and writing representation in Ger- ods in Natural Language Processing, pages 989–
man narrative texts. Literary and Linguistic Com- 999, Seattle, WA.
puting, 28(4):563–575.
Jeffrey Pennington, Richard Socher, and Christopher D.
David Elson, Nicholas Dames, and Kathleen McKe- Manning. 2014. Glove: Global vectors for word rep-
own. 2010. Extracting social networks from liter- resentation. In Proceedings of the Conference on
ary fiction. In Proceedings of the Annual Meet- Empirical Methods in Natural Language Processing,
ing of the Association for Computational Linguistics, pages 1532–1543.
pages 138–147, Uppsala, Sweden.
Bruno Pouliquen, Ralf Steinberger, and Clive Best.
David Elson and Kathleen McKeown. 2010. Auto- 2007. Automatic detection of quotations in multi-
matic attribution of quoted speech in literary nar- lingual news. In Proceedings of Recent Advances
rative. In Proceedings of the Twenty-Fourth AAAI in Natural Language Processing, pages 487–492,
Conference on Artificial Intelligence. Borovets, Bulgaria.
Manaal Faruqui and Sebastian Pado. 2012. Towards Christian Scheible, Roman Klinger, and Sebastian
a model of formal and informal address in English. Padó. 2016. Model architectures for quotation de-
In Proceedings of the 13th Conference of the Euro- tection. In Proceedings of ACL, pages 1736–1745,
pean Chapter of the Association for Computational Berlin, Germany.
Linguistics, pages 623–633, Avignon, France.
Elena Semino and Michael Short. 2004. Corpus Stylis-
Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsu- tics: Speech, Writing And Thought Presentation In
ruoka, and Richard Socher. 2017. A joint many-task A Corpus Of English Writing. Routledge Advances
model: Growing a neural network for multiple nlp In Corpus Linguistics. Routledge, London.
tasks. In Proceedings of the 2017 Conference on
Empirical Methods in Natural Language Processing,
pages 1923–1933, Copenhagen, Denmark.
894
Validation of Facts Against Textual Sources
Vamsi Krishna Pendyala Simran sinha Satya Prakash Shriya Reddy Anupam Jamatia
Department of Computer Science and Engineering
National Institute of Technology
Agartala, Tripura, India
{krishnapendiala, simsinha32}@gmail.com
{satyaprakash30497, shriyavipul90, anupamjamatia}@gmail.com
895
Proceedings of Recent Advances in Natural Language Processing, pages 895–903,
Varna, Bulgaria, Sep 2–4 2019.
textual source related to a given claim. If the tex- 2 Related Work
tual source has no knowledge about the claim, it
There has been a substantial amount of work
would be labelled as ‘Out-of-Context’ for exam-
done in the field of fact verification. Vlachos
ple if we have a source text about sports and we
and Riedel (2014) provided the first dataset re-
need to verify a claim related to politics then the
lated to fact verification containing 211 labelled
claim is Out-of-Context. Although if it has some
claims in the political domain with evidence hy-
information about the claim it would further be
perlinks. An alternative is Wang (2017) which re-
classified as True, False or Inappropriate. The
leased a dataset called LIAR dataset for detecting
result would be ‘True’ if the source supports the
fake news, which contains 12.8K claims labelled
claim, it would be ‘False’ if the source opposes
manually using POLTIFACT.COM2 on different
the claim and it would be labelled as ‘Inappro-
context. Alhindi et al. (2018) extended the LIAR
priate’ if the evidence is not sufficient to con-
dataset and labelled a claim using the speaker re-
clude the classification of the claim, e.g., know-
lated metadata without using the evidence. Basi-
ing that Michael Jackson was a singer, we cannot
cally, they used the justification given by the hu-
infer whether his children will be singers. This
mans at the end of the article in the summary.
is something inappropriate to the information pro-
Modelling the extracted justification along with
vided. Hence the claim ‘Michael Jackson’s chil-
the claim yielded better results rather than using
dren would be singers’ is an inappropriate claim
a machine learning model for binary classifica-
for the evidence ‘Michael Jackson was a Singer’.
tion and a six way classification. Ferreira and
This labelling would not only help in knowing
Vlachos (2016) later presented a new modified
about the claim but also helps us to know about
dataset known as Emergent, where they had 300
our textual source as well. The labels Inappro-
claims and 2,595 related articles and they came
priate and Out-of-Context claim tell us whether
to the conclusion that fact verification can also
the textual source has enough information to con-
be treated as Natural Language Inference Task,
clude about the claim or not, if a claim is Out-of-
as they used textual entailment to predict whether
Context then the domain of our source text can
the article supports the claim or not. The lat-
be expanded to answer the claim. Inappropriate
est large scale dataset is prepared Thorne et al.
means that we do have required knowledge about
(2018a) which was annotated manually and used
the claim and also cannot conclude anything based
for verification against textual sources. It contains
upon this available knowledge. The best perform-
185,441 claims generated from Wikipedia. These
ing version of our system is seen to be giving
claims were classified Supported, Refuted and Not
63.06% accuracy.
Enough Info by annotators. In this, Recogniz-
In Section 2 we have mentioned various works
ing Textual Entailment (RTE) component was pro-
done by different authors related to fact extrac-
ceeded by an evidence retrieval module. The ac-
tion and verification. Section 3 explains the data-
curacy measured found to be 31.87% if evidence
set collection and the system architecture to clas-
was taken into consideration and 50.91% if ev-
sify claims against a textual source. Section 4 de-
idence is ignored. The only drawback of this
scribes the stages of this experiment which are evi-
system is its restriction to the Wikipedia domain.
dence retrieval, similarity measures used to obtain
Thorne and Vlachos (2018) conducted a survey
the relation between claim and evidence, the first
on automated fact checking research stemming us-
level of classification of the claim into in-context
ing natural language processing and other related
and out-context, then the further classification of
fields. According to this survey, the inputs for ver-
in-context claims to True, False and Inappropriate
ification system play a vital role. Evidence re-
claim. In Section 5 we discussed error analysis of
trieval plays a vital role in solving the fact veri-
our system along with the results obtained on us-
fication problem. Fact checking requires the apt
ing different classification algorithms on our data
evidence against which sentences can be predicted
for classification purpose and comparative mea-
to be true or false. Chen et al. (2017a) provides
sure between different models. In Section 6 we
a framework for open domain question answering
conclude about our model with its practical usage
upon Wikipedia and SQuAD data set. This in-
in the real world.
volved machine reading along with the document
2
www.politifact.com
896
retrieval and then identifying the answers. We deal part. Lexical tagging is done by using two lexi-
with a similar retrieval problem like open domain cal tags name such as Parts-of-Speech and Named
Question Answering, which would be succeeded Entity Recognition to enhance the performance.
by verification using textual entailment. Natural
Language Inference is basically a task to find out 3 System Architecture
whether a hypothesis entails, contradicts or is neu- In this section, we discuss the overview of our
tral about the claim. There have been recent devel- system and our approach for classifying a claim
opments in these fields like the SNLI dataset for based upon a particular source text. Our approach
learning natural language inference built by Bow- is divided into two stages: Evidence Retrieval and
man et al. (2015). Different neural NLI models Classification of claim as True, False, Inappropri-
(Nie and Bansal (2017); Parikh et al. (2016); Chen ate or Out-of-Context using a textual entailment
et al. (2017b); Gong et al. (2018)) that achieve module. The reason for using textual entailment
promising performance. Parikh et al. (2016) has is because it precisely gives us a relationship be-
the highest accuracy on the Stanford Natural Lan- tween an evidence and a claim. In the first stage
guage Inference task. We used the similar ap- i.e., evidence retrieval, given a claim, we find its
proach as the FEVER (Thorne et al., 2018c) but, TF-IDF vectors corresponding to the source text
instead of a three class classification we have ex- against which it is being verified.
tended it to a four class classification namely True, Later we find out the cosine similarity between
False, Inappropriate and Out-of-Context. These the TF-IDF vector of a claim to each of the sen-
two modules combined together help us in valida- tences present in the source text. Then, we filter
tion of a fact. In 2018, a shared task known as the out the top four sentences which have the highest
FEVER Shared Task (Thorne et al., 2018b) was cosine similarity values and consider this as the
held which dealt with the fact verification prob- evidence for that particular claim as discussed in
lem. The FEVER shared task is closely related to section 3.2. The reason for considering top 4 sen-
our work as it uses the same two modules. The tences is that they closely correspond to the nearest
following are some of the systems that partici- sentences to the claim. In the next stage, the ex-
pated in the FEVER shared task: Nie et al. (2018) tracted evidence and the present claim are passed
have scored the maximum of 64% accuracy in the into a Textual Entailment module which returns
FEVER shared task, in which they used the neu- the probabilities of two texts entailing, contradict-
ral semantic matching networks. For both the evi- ing or neutral towards each other. These proba-
dence retrieval and RTE model they enhanced the bilities along with other variables discussed later
working using the neural networks. Hidey and are used as a feature vector for our classification
Diab (2018) known as the Team Sweepers, made model. The entire claim classification process is
the evidence retrieval system better using lexical explained in 3.3. Section 3.1 describes the pro-
tagging and syntactic similarity. They used multi- cess of preparation of the dataset.
task learning and trained both the components to-
gether and set the parameters in a way using re- 3.1 Dataset
inforcement learning so that it can first find sen- Due to the uniqueness of our classification, we
tences related to the claim and then find their re- were supposed to either prepare our own dataset or
lation with the claim. DeFactoNLP (Reddy et al., modify an existing standard dataset for serving our
2018) aimed at retrieving the evidence for the val- purpose. Here, we have done both, we modified a
uation of the claim from Wikipedia. The retrieval dataset known as the SICK dataset and prepared a
of documents which is considered as evidence is new dataset called the NITA dataset.
done by TF-IDF vectors of the claim and the sen-
tences in the documents followed by inputting 3.1.1 SICK Dataset
them to a textual entailment recognition module. The main target of our experiment was to classify
Then the Random forest classifier is used for the a claim based upon its evidence, so we required a
classification of the claim. Lee et al. (2018) have dataset consisting of sentence pairs and a correla-
introduced a method by developing a neural ranker tion between these two sentences. Hence we used
using decomposable attention model and lexical a publicly available dataset known as the SICK
tagging instead of TF-IDF for evidence retrieval dataset. The SICK-2014 dataset (Marelli et al.,
2014) was introduced as Task 1 of the SemEval
897
ID Sentence1 Sentence2 Entailment Score
23 A group of kids is playing in a yard and an old man is standing in the background A group of boys in a yard is playing and a man is standing in the background yes 4.5
14 A brown dog is attacking another animal in front of the man in pants. Two dogs are fighting. unknown 3.5
13 Two dogs are wrestling and hugging. There is no dog wrestling and hugging. no 3.3
Table 1: Sample SICK dataset with entailment labels and relatedness scores.
2014 conference and in contrast to SNLI (Bow- entities. The claims were generated based
man et al., 2015), it is geared at specifical bench- upon every source text. For example, con-
marking semantic compositional methods, aiming sider “The rabbit tortoise race” as the source
to capture only similarities on purely language and text, one of the claims related to this source
common knowledge level, without relying on do- text can be “Rabbit won the race”.
main knowledge, and there are no named entities
or multi-word idioms. It consists of total 10,000 • Claim Labelling: Classifying whether a
pairs of sentences. claim is True, False, Inappropriate or Out-
We modified the SICK dataset as per our clas- of-Context based on the evidence from source
sification by adding two more columns to it. We text was done at this stage. We checked ev-
manually labelled a claim and an evidence pair to ery claim manually with respect to its source
be in-context or out-context based upon their re- text and labelled the claim accordingly. The
latedness score as given in the dataset, which indi- labelling is done as per the meaning of each
cates the semantic similarity of these pair of sen- label which was discussed in the introduction
tenced. Firstly, we labelled all the pairs with re- section.
latedness score less than 3 as Out-of-Context and • Dataset Validation: Considering the com-
other claims as True, False or Inappropriate based plexity of labelling of claims, we considered
upon their textual entailment labels provided by validation of the data set generated by us. For
the SICK dataset. this purpose we tried to analyse the labels we
3.1.2 NITA Dataset gave to each claim, where labels generated by
After considering SICK dataset we even wanted to one person were analysed by other to estab-
develop our own dataset consisting of source texts lish an inter annotator agreement. We consid-
and claims along with their labels as follows: ered around 30% that is 240 claims for this
validation process and calculated the Fleiss k
• Source Text Collection: We collected some score (Fleiss, 1971) to be 0.876.
short stories and articles related to sports,
movies, mythology, moral stories, Wikipedia
LABEL No. of Claims
articles in English language and considered
them as source texts. The total number of TRUE 170
source texts collected in this way turned out FALSE 170
to be 53 . OUT-OF-CONTEXT 420
INAPPROPRIATE 168
• Claim Generation: Corresponding to these Total No. of Claims 928
53 stories/articles/textual content, we pre-
pared a total of 928 claims. The purpose
Table 3: NITA Dataset Splitting based upon La-
was to generate claims about a single fact
bels
which could be arbitrarily complex and al-
lowed for a variety of expressions for the
898
3.2 Evidence Retrieval claim, and let ui be the probability of it being un-
We used the concept of Document Retrieval from informative returned by textual entailment. Below
the DrQA system (Chen et al., 2017a). Firstly, we are some variables we considered for our conve-
find out the Term Frequency-Inverse Document nience:
(
Frequency (TFIDF) vectors (Hiemstra, 2000) for 1 if si ≥ ri and si ≥ ui
a claim and the sentences of the source text. We csi =
0 otherwise
then calculate the cosine similarity between the
claim and each sentence. Thereafter, we pick the (1)
(
top four similar sentences based on the cosine 1 if ri ≥ si and ri ≥ ui
similarity between the bigram TF-IDF vectors of cri =
0 otherwise
the sentences and the claim. These sentences are (2)
finally chosen as possible sources of evidence. (
Now, we are left with claim and evidence pairs. 1 if ui ≥ si and ui ≥ ri
cui =
0 otherwise
(3)
3.3 Classification n
C.E
In this final stage, we classify all the claims to be CosineSimilarity = cos(θ) = ||C||||E||
True, False, Inappropriate or Out-of-Context us- (4)
ing machine learning classification models. The
features for this classification are obtained by The similarity variable used here is cosine simi-
passing a claim and an evidence to a textual en- larity between claim and evidence. The value C
tailment module in order to obtain probabilities of and E denote the vector notation of claim c and
entailment, contradiction and neutrality between evidence e based upon their word frequency. Con-
claim and evidence. The RTE is the process of sider the cosine similarity between claim and evi-
determining whether a text fragment (Hypothesis dence i to be Si . Using above variables we form a
H) can be inferred from another fragment (Text feature vector for each claim and evidence pair for
T) (Sammons et al., 2012). The RTE module re- the classification model i as:
ceives the claim and the set of possible evidences feature vector =< si , ri , ui , csi , cri , cui , Si >
from the previous stages. Let there be ’n’ possi-
ble sources of evidence for verifying a claim. For The above feature vector give us an understand-
the ith possible evidence, let si denote the prob- ing of how closely two statements are related i.e.,
ability of it entailing the claim, let ri denote the a relationship between claim and evidence. Some
probability of it contradicting the claim, and let ui statements which are a negation to each other may
be the probability of it being uninformative. The have high cosine similarity but then their contra-
RTE module calculates each of these probabilities. diction probability would be high which would
The SNLI corpus (Bowman et al., 2015) is used help the learning algorithm to classify claims ac-
for training the RTE model. This corpus is com- curately. We used both the datasets i.e., the SICK
posed of sentence pairs T, H where T corresponds dataset and the NITA dataset, along with the above
to the literal description of an image and H is a mentioned feature vector for training and testing
manually created sentence. If H can be inferred purpose of various machine learning classifica-
from T, the “Entailment” label is assigned to the tion models like Naïve Bayes, Logistic Regres-
pair. If H contradicts the information in T, the pair sion, Support Vector Machine, Random Forest and
is labelled as “Contradiction”. Otherwise, the la- Multi-Level Perceptron. These models were used
bel ‘Neutral’ is assigned. We chose to employ the as they are widely used in the industry for practical
state-of-the-art RTE by (Parikh et al., 2016). We applications.
selected this because at the time of development of
this work, it was one of the best performing sys- 4 Experiments
tems on the task with publicly available code. As mentioned in section 3, our model of fact veri-
For a particular claim c and an evidence e let fication consists of two stages:
si denote the probability of it entailing the claim,
let ri denote the probability of it contradicting the 1. Retrieving evidence related to the claim from
the source text.
899
Selected Evidence
Document
Evidence claim
Retrieval
claim evidence
sentence
Filtering & rank-
ing
Figure 1: System Overview: Document Retrieval, Sentence Selection, and Claim Verification.
claim
Relatedness Entailment
[1-2) range 923 (10%) NEUTRAL 5595 (57%)
[2-3) range 1373 (14%) CONTRADICTION 1424 (14%)
[3-4) range 3872 (39%) ENTAILMENT 2821 (29%)
[4-5) range 3672 (37%)
Table 4: Distribution of SICK sentence pairs for each gold relatedness level and entailment label.
900
negation of each other and hence have high word
similarity. Next, upon observing results produced
by the classification model, we saw that most
inappropriate claims were classified as true and
some true claims were classified as inappropri-
ate. This ambiguity is mainly due to evidence
supporting a claim partially, the probability of
entailment for this would be high but due to
Figure 3: Random Forest Confusion Matrix for variance in cosine similarity between claim and
SICK Dataset evidence there can arise an ambiguity. The overall
system performance is at par with other existing
fact verification systems as mentioned in section
in terms of accuracy. Further modifications
to improve the performance of the system are
discussed in the next section .
901
Data-set Model Precision Recall F1-score Accuracy
Naive Bayes 0.53 0.53 0.50 0.535
SVM 0.58 0.58 0.57 0.582
SICK Random Forest 0.63 0.63 0.63 0.630
Logistic Regression 0.57 0.58 0.57 0.577
MLP 0.63 0.62 0.63 0.624
Naive Bayes 0.61 0.63 0.59 0.631
SVM 0.62 0.65 0.61 0.649
NITA Random Forest 0.61 0.65 0.61 0.649
Logistic Regression 0.61 0.65 0.60 0.649
MLP 0.61 0.64 0.61 0.644
References Yichen Gong, Heng Luo, and Jian Zhang. 2018. Nat-
ural language inference over interaction space. In
Tariq Alhindi, Savvas Petridis, and Smaranda Mure- International Conference on Learning Representa-
san. 2018. Where is your evidence: Improving fact- tions.
checking by justification modeling. In Proceedings
of the First Workshop on Fact Extraction and VERi- Christopher Hidey and Mona Diab. 2018. Team
fication (FEVER). pages 85–90. SWEEPer: Joint sentence extraction and fact check-
ing with pointer networks. In Proceedings of the
Samuel R. Bowman, Gabor Angeli, Christopher Potts, First Workshop on Fact Extraction and VERification
and Christopher D. Manning. 2015. A large anno- (FEVER). Brussels, Belgium, pages 150–155.
tated corpus for learning natural language inference.
In Proceedings of the 2015 Conference on Empirical Djoerd Hiemstra. 2000. A probabilistic justification for
Methods in Natural Language Processing. Associa- using tfxidf term weighting in information retrieval.
tion for Computational Linguistics, Lisbon, Portu- International Journal on Digital Libraries 3(2):131–
gal, pages 632–642. 139.
Danqi Chen, Adam Fisch, Jason Weston, and Antoine Nayeon Lee, Chien-Sheng Wu, and Pascale Fung.
Bordes. 2017a. Reading wikipedia to answer open- 2018. Improving large-scale fact-checking using
domain questions. In Proceedings of the 55th An- decomposable attention models and lexical tagging.
nual Meeting of the Association for Computational In Proceedings of the 2018 Conference on Empiri-
Linguistics (Volume 1: Long Papers). Association cal Methods in Natural Language Processing. pages
for Computational Linguistics, Vancouver, Canada, 1133–1138.
pages 1870–1879.
Marco Marelli, Luisa Bentivogli, Marco Baroni, Raf-
Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui faella Bernardi, Stefano Menini, and Roberto Zam-
Jiang, and Diana Inkpen. 2017b. Enhanced LSTM parelli. 2014. Semeval-2014 task 1: Evaluation of
for natural language inference. In Proceedings of compositional distributional semantic models on full
the 55th Annual Meeting of the Association for Com- sentences through semantic relatedness and textual
putational Linguistics (Volume 1: Long Papers). As- entailment. In Proceedings of the 8th International
sociation for Computational Linguistics, Vancouver, Workshop on Semantic Evaluation (SemEval 2014).
Canada, pages 1657–1668. Association for Computational Linguistics, Dublin,
Ireland.
William Ferreira and Andreas Vlachos. 2016. Emer-
gent: a novel data-set for stance classification. In Yixin Nie and Mohit Bansal. 2017. Shortcut-stacked
Proceedings of the 2016 Conference of the North sentence encoders for multi-domain inference. In
American Chapter of the Association for Compu- Proceedings of the 2nd Workshop on Evaluating
tational Linguistics: Human Language Technolo- Vector Space Representations for NLP. Association
gies. Association for Computational Linguistics, for Computational Linguistics, Copenhagen, Den-
San Diego, California, pages 1163–1168. mark, pages 41–45.
J. L. Fleiss. 1971. Measuring nominal scale agree- Yixin Nie, Haonan Chen, and Mohit Bansal. 2018.
ment among many raters. Psychological Bulletin Combining fact extraction and verification with neu-
76(5):378–382. ral semantic matching networks. arXiv preprint
arXiv:1811.07039 .
902
Ankur Parikh, Oscar Täckström, Dipanjan Das, and FEVER: a large-scale dataset for fact extraction
Jakob Uszkoreit. 2016. A decomposable attention and VERification. In Proceedings of the 2018
model for natural language inference. In Proceed- Conference of the North American Chapter of the
ings of the 2016 Conference on Empirical Meth- Association for Computational Linguistics: Human
ods in Natural Language Processing. Association Language Technologies, Volume 1 (Long Papers).
for Computational Linguistics, Austin, Texas, pages Association for Computational Linguistics, New
2249–2255. Orleans, Louisiana, pages 809–819.
Aniketh Janardhan Reddy, Gil Rocha, and Diego Es- James Thorne, Andreas Vlachos, Oana Cocarascu,
teves. 2018. Defactonlp: Fact verification using en- Christos Christodoulopoulos, and Arpit Mittal.
tity recognition, tfidf vector comparison and decom- 2018b. The Fact Extraction and VERification
posable attention. In Proceedings of the First Work- (FEVER) shared task. In Proceedings of the
shop on Fact Extraction and VERification (FEVER). First Workshop on Fact Extraction and VERification
Association for Computational Linguistics, pages (FEVER).
132–137.
James Thorne, Andreas Vlachos, Oana Cocarascu,
Benjamin Riedel, Isabelle Augenstein, Georgios P Sp- Christos Christodoulopoulos, and Arpit Mittal.
ithourakis, and Sebastian Riedel. 2017. A sim- 2018c. Proceedings of the first workshop on fact
ple but tough-to-beat baseline for the fake news extraction and verification (fever). In Proceedings
challenge stance detection task. arXiv preprint of the First Workshop on Fact Extraction and VERi-
arXiv:1707.03264 . fication (FEVER).
Mark Sammons, Vinod Vydiswaran, and Dan Roth.
Andreas Vlachos and Sebastian Riedel. 2014. Fact
2012. Recognizing textual entailment. Multilin-
checking: Task definition and dataset construction.
gual Natural Language Applications: From Theory
In Proceedings of the ACL 2014 Workshop on Lan-
to Practice pages 209–258.
guage Technologies and Computational Social Sci-
James Thorne and Andreas Vlachos. 2018. Automated ence. pages 18–22.
fact checking: Task formulations, methods and fu- William Yang Wang. 2017. “liar, liar pants on fire”: A
ture directions. In Proceedings of the 27th Inter- new benchmark dataset for fake news detection. In
national Conference on Computational Linguistics. Proceedings of the 55th Annual Meeting of the As-
Association for Computational Linguistics, pages sociation for Computational Linguistics (Volume 2:
3346–3359. Short Papers). Association for Computational Lin-
James Thorne, Andreas Vlachos, Christos guistics, Vancouver, Canada, pages 422–426.
Christodoulopoulos, and Arpit Mittal. 2018a.
903
A Neural Network Component for Knowledge-Based Semantic
Representations of Text
Alejandro Piad-Morffis1 , Rafael Muñoz2 , Yudivián Almeida-Cruz1 ,
Yoan Gutiérrez3 , Suilan Estevez-Velarde1 , and Andrés Montoyo2
1
School of Math and Computer Science, University of Havana, Cuba
{sestevez,yudy,apiad}@matcom.uh.cu
2
Department of Languages and Computing Systems, University of Alicante, Spain
3
U.I. for Computer Research (IUII), University of Alicante, Spain
{montoyo,ygutierrez,rafael}@dlsi.ua.es
904
Proceedings of Recent Advances in Natural Language Processing, pages 904–911,
Varna, Bulgaria, Sep 2–4 2019.
In this work, SNNs have been designed to repre- common task for many machine learning prob-
sent the relevant concepts of a knowledge base as lems. In this section we present different ap-
part of the artificial network architecture. Hence, proaches to the design of semantic representations
during the whole process, the structures that map for natural text.
to specific concepts and relations are clearly iden-
tifiable inside the network. This process attaches Network-Based Approaches for semantic rep-
a semantic meaning to the network architecture, resentations usually consist of defining some sim-
which is useful for debugging and understanding ilarity metric based on the relations of terms in
the learning dynamics. Furthermore, the SNN is some knowledge base, interpreted as a graph. It
trained to learn the specific instances in the knowl- is based in the assumption that words which are
edge base and their relations. This way, the SNN connected by short paths in a knowledge base
not only encodes abstract concepts, but also true should have similar semantics. WordNet is com-
facts about instances of those concepts. Prelim- monly used as a knowledge base where differ-
inary source code is available online for the re- ent semantic relations among words can be ex-
search community.1 ploited for defining similarity metrics. Using
WordNet (Miller, 1995), several semantic simi-
When used as a component of a larger deep
larity metrics are defined by exploring the graph
learning architecture, a SNN that is trained for a
structure of the knowledge base, mostly depend-
specific knowledge domain can be seen as a se-
ing on the graph structure of words, such as Hirst-
mantic representational component. Its input con-
St-Onge (Hirst and St-Onge, 1998) and Leacock-
sists of a low-level representation of data, (for ex-
Chodorow (Leacock and Chodorow, 1998). In this
ample, words or entities), and its output consists of
direction, other researchers include information
an implicit representation of this data expressed in
content formulae to measure appearances of terms
terms of the learned domain. This representation
in a corpus, such as the Resnik metric (Resnik,
can be seen as a type of embedding that maps raw
1999).
input to a semantic space defined by the concepts
and relations of the learned knowledge base. The Corpus-Based Similarity Metrics are defined
SNN is used as a representational layer in a larger by some measure of the co-occurrence of terms
neural network, for a natural language processing in a corpus of natural text. The intuitive idea is
problem in which the semantic representation in- that words which co-occur within the same con-
duced by the learned knowledge base is expected text must have similar semantics. One such metric
to be a good representation. is PMI-IR (point mutual information - information
The paper is organized as follows. Section 2 retrieval) (Turney, 2001), which considers the in-
presents a review of related works and relevant formation content of each pair of words (wi , wj ),
concepts in the domain of representational learn- measured as the relative number of co-occurrences
ing, with an emphasis on different semantic rep- of wi and wj in a document, with respect to
resentations of natural language. In section 3 the the individual count of occurrences of each word.
architecture and training procedure for the SNN Another corpus-based similarity metric based is
is formalized and explained. Section 4 presents ESA (explicit semantic analysis) (Gabrilovich and
a brief experimental setup designed to illustrate Markovitch, 2007), which considers Wikipedia as
the behavior and performance of a SNN in a spe- a corpus for building a co-occurrence matrix of
cific natural language problem. Finally, section 5 words. A similar approach is HAL (hyperspace
discusses the main contributions of this proposal, analogue to language) (Lund and Burgess, 1996),
whereas section 6 presents the final conclusions of which also builds a co-occurrence matrix, but only
the research and highlights possible future lines of considering words within a small window.
development.
Dimensionality Reduction Techniques such as
2 Related Works PCA (principal component analysis) (Martinsson
et al., 2011) can be interpreted as a projection
Building semantic representations of raw input from the BOW (bag of words) or TF-IDF (term
data, specifically for natural language text, is a frequency - inverse document frequency) space
1
https://github.com/ to a semantic space. An interesting recent ap-
knowledge-learning/snn proach, that mixes ideas from the previous tech-
905
niques, is the family of word embedding algo- crucial problems (e.g.medical diagnosis or legal
rithms (Mikolov et al., 2013). In word embed- advice). By carefully designing the learning crite-
dings, similar to PCA, each word is mapped to a ria and structure of embeddings, it is conceivable
vector which encodes the semantics of the word. that a semantic representation can be interpreted
The embedding is chosen such that a word’s vec- in terms of a formal conceptualization defined a
tor contains an implicit representation of the prob- priori.
abilistic distribution of the word’s context in a
given corpus. To achieve this, a neural network 3 Semantic Neural Networks
is trained to predict, given a word wi ’s embed-
ding, the probability that some other word wj ap- We define a Semantic Neural Network (SNN) as
pears in a small window centered around wi . In an artificial neural network architecture that en-
this sense, word embeddings can be seen as a codes knowledge. Two main semantic elements
generalization of corpus-based metrics, whereby are encoded from the Knowledge Base. First, the
the best representation is learned from the data, graph structure of the Knowledge Base (i.e. en-
rather than handcrafted. Even though word em- tity classes and their relations) is directly repre-
beddings don’t explicitly model specific semantic sented in the architecture of a SNN. Second from
relations (such as hypernymy, or synonymy), it has a Knowledge Base KB, the information about
been shown that several interesting semantic rela- which instances belong to which entity classes
tions get encoded in specific directions in the em- and their specific relations are encoded into the
bedding space, enabling the solution of analogue weights of the SNN. By design, the architecture of
inference queries (Schnabel et al., 2015). the SNN is built to represent each specific entity
class and relation in a clearly recognizable struc-
Entity Embeddings are a specific type of em- ture (a set of neurons with a pre-designed connec-
bedding technique that encodes the context of en- tion pattern). This allows a semantic meaning to
tities in a knowledge base. Several metrics can be attached to an activation of the SNN in terms of
be used to define the notion of “context similar- the classes and relations defined in the Knowledge
ity” when using a knowledge base for entity em- Base.
bedding. For example, embeddings can be de- The purpose of the SNN is to provide a semantic
signed such that a particular direction dr is asso- representation of the input data that can be used as
ciated with each particular relation of the knowl- component inside a larger deep learning architec-
edge base, such that ei + dr ≈ ej whenever ei ture, to solve a related learning problem L. If the
and ej are related by r. These formulations allow knowledge represented in KB is useful for solv-
a semantic meaning to be attached to a particular ing the problem L, then the representation pro-
algebraic operation and properties, and enable a vided by the SNN should be richer than plain bag-
whole new field of study that finds the “meaning” of-words or general purpose embeddings. With
of, say, other directions d which are orthogonal to the same computational power (same number of
or linearly dependent on a specific relation. Entity parameters), a deep learning architecture using a
embeddings have been extended to encode also SNN is expected to achieve equal or better perfor-
the hierarchical structure of knowledge bases (Hu mance than using other representations not specif-
et al., 2015) and mixed with word embeddings for ically designed to exploit the knowledge in KB.
tasks such as entity disambiguation (Yamada et al., Furthermore, the SNN architecture provides a se-
2016). mantic explanation for the model’s predictions.
A SNN is built based on a specific Knowl-
Word and entity embeddings in general are edge Base KB which is of interest for solving
promising approaches to deal with learning se- a related learning problem L. Figure 1 shows a
mantic representations of data. Moreover, recent schematic representation of the process for con-
research deals with finding ways to exploit the structing, training and using an SNN in an arbi-
structure of these representations to explain why a trary learning problem L. First, given the prob-
specific answer is output by a neural network. Be- lem L to solve, a relevant Knowledge Base KB
ing able to explain neural networks is a first step is chosen. The entity classes and relations of
towards designing accountable machine learning KB are used to define the architecture of the
systems that humans can trust for solving the most SNN (Section 3.1). Then, training instances are
906
Problem (5)
(L) Instances Training Neural Network representation. The output dimension is a fixed
size dense vector of small dimension (e.g., 10).
The input and output are connected by a linear
(1) (4)
Select Insert in
model matrix operator. Additionally, a one-dimensional
Knowledge
indicator neuron is connected to the output layer
(2)
Define
Base Class
architecture
and
(KB) Relations SNN
through a dot product operator with a sigmoid ac-
tivation function. The purpose of the indicator is
(3)
Instances Training
to signal when the activation of the output is large
in absolute value. This is interpreted as the impor-
Figure 1: The process for constructing and using tance of the corresponding concept in a particular
a SNN for a specific learning problem (L) and a input text.
suitable Knowledge Base (KB). The SNN is first A relation block is a similar computational
defined and pre-trained based on KB, and then graph, but with an input variable whose size is
used in a larger neural network trained specifically twice the entity output size. Thus, the relation
for L. input shape corresponds with the output shape of
the two entity blocks that will be connected. The
outputs from each incoming entity block are con-
extracted from KB and used to pre-train the SNN
catenated, forming a single vector, which is then
weights (Section 3.2.1). Afterwards, the SNN is
connected through a linear matrix operator to the
included in a deep learning model (which can con-
output variable. An identical indicator neuron is
tain other components such as extra layers). Fi-
connected to the output variable.
nally, this larger model is trained on instances
from L using standard optimization techniques The overall architecture of an SNN consists of
and loss functions suitable for the problem L (Sec- several entity blocks, one for each class in the
tion 3.2.2). knowledge base, and several relation blocks, one
Different Semantic Neural Networks can be for each relation defined. The entity blocks are all
trained on different Knowledge Bases ahead of connected to a single input (e.g., a bag of words
time and reused for many related problems. In representation). Every relation block is connected
this sense, SNNs are similar to pre-trained word to the respective outputs of the entity blocks that
embeddings, since a SNN trained for a commonly represent the classes for which the relation is de-
used Knowledge Base (i.e., DBPedia or Word- fined. Figure 2 shows a schematic representation
Net) could be useful in solving different problems. of an example SNN built from a knowledge base
However, pre-trained SNNs can (and should) be in the cinematic domain (specifically the Internet
fine-tuned on a specific problem L after decid- Movie Database, IMDB).
ing a convenient deep learning architecture for this The input size of the SNN is the size of some
problem. vocabulary chosen before hand. This vocabulary
should include the common terms in the knowl-
3.1 Architecture of the Semantic Neural edge domain(s) of interest. An additional input
Network dimension can be added to account for unknown
words (those not present in the vocabulary at the
The architecture of a Semantic Neural Net-
moment of training).
work (SNN) is composed of several instances of
two simple structures: entity blocks and relation
3.2 Training the Semantic Neural Network
blocks. For each class of the knowledge base, an
entity block is created, and for each relation, a cor- The training procedure for a SNN is divided in
responding relation block. two phases, each of which solves a different learn-
An entity block is a computational graph with a ing problem. In the first phase, which we call
single input variable and a single output variable. “structured pre-training”, the parameters of the en-
The input dimension and shape is determined by tity and relation blocks are adjusted. The learning
the specific application, a sensible default consists objective for this phase consists of predicting to
of a single one-hot encoding layer when using a which instance of the knowledge base the sample
bag-of-words representation, but an alternative in- of natural text refers. In the second phase, in order
put could be a general-purpose word embedding to deal with the original natural language process-
907
that hold in each relation must be encoded in the
Indicators Wr weights of the corresponding relation block.
Furthermore, in order to differentiate the relations,
the SNN needs to learn to differentiate the entities
implicitly because there is no requisite to define
the specific learning objective, namely represent-
ing different entities and their corresponding en-
codings.
908
Classes Instances
Person 2 854 359 Indicators
909
Indicators
Model Parameters Accuracy
Person
SNN + 3 ReLU 5,491,981 67.82 Ben
910
architectures, such as LSTMs, CNNs and differ- Kevin Lund and Curt Burgess. 1996. Pro-
ent embedding strategies. Finally, the SNN inter- ducing high-dimensional semantic spaces
from lexical co-occurrence. Behavior Re-
nal structure has a semantic meaning attached to
search Methods, Instruments, and Computers
each neuron block. This opens the door for the https://doi.org/10.3354/ame01683.
design of interpretation models that can automat-
ically output an explanation of a neural networks Per Gunnar Martinsson, Vladimir Rokhlin, and
Mark Tygert. 2011. A randomized algo-
response in terms of human-defined concepts. rithm for the decomposition of matrices. Ap-
plied and Computational Harmonic Analysis
Acknowledgments https://doi.org/10.1016/j.acha.2010.02.003.
This research has been supported by a Carolina Tomas Mikolov, Greg Corrado, Kai Chen, and
Foundation grant in agreement with University Jeffrey Dean. 2013. Efficient Estimation of
of Alicante and University of Havana. More- Word Representations in Vector Space. Proceed-
ings of the International Conference on Learn-
over, it has also been partially funded by both ing Representations (ICLR 2013) pages 1–12.
aforementioned universities, the Generalitat Va- https://doi.org/10.1162/153244303322533223.
lenciana and the Spanish Government through the
George A Miller. 1995. Wordnet: a lexical database for
projects SIIA (PROMETEU/2018/089), LIVING- english. Communications of the ACM 38(11):39–
LANG (RTI2018-094653-B-C22) and INTEGER 41.
(RTI2018-094649-B-I00).
Grégoire Montavon, Wojciech Samek, and Klaus-
Robert Müller. 2017. Methods for interpreting and
understanding deep neural networks. Digital Signal
References Processing .
Yoshua Bengio. 2012. Deep Learning of Representa-
tions for Unsupervised and Transfer Learning. In Bo Pang and Lillian Lee. 2004. A sentimental educa-
JMLR: Workshop and Conference Proceedings. vol- tion: Sentiment analysis using subjectivity summa-
ume 27, page 17–37. rization based on minimum cuts. In Proceedings of
the ACL.
Evgeniy Gabrilovich and Shaul Markovitch. 2007.
Philip Resnik. 1999. Semantic Similarity in a Taxon-
Computing semantic relatedness using wikipedia-
omy: An Information-Based Measure and its Ap-
based explicit semantic analysis. In IJCAI Inter-
plication to Problems of Ambiguity in Natural Lan-
national Joint Conference on Artificial Intelligence.
guage. Journal of Artificial Intelligence Research
https://doi.org/10.1145/2063576.2063865.
11:95–130. https://doi.org/10.1613/jair.514.
Graeme Hirst and David St-Onge. 1998. Lexical
Tobias Schnabel, Igor Labutov, David Mimno, and
chains as representations of context for the detec-
Thorsten Joachims. 2015. Evaluation methods for
tion and correction of malapropisms. WordNet:
unsupervised word embeddings. In Proceedings of
An electronic lexical database 305(April):305–332.
the 2015 Conference on Empirical Methods in Nat-
https://doi.org/citeulike-article-id:4893262.
ural Language Processing. pages 298–307.
Zhiting Hu, Poyao Huang, Yuntian Deng, Yingkai Gao, Peter D Turney. 2001. Mining the Web for Synonyms:
and Eric P Xing. 2015. Entity hierarchy embed- PMI-IR versus LSA on TOEFL. Lecture Notes
ding. In Proceedings of the 53rd Annual Meet- in Computer Science https://doi.org/10.1007/3-540-
ing of the Association for Computational Linguis- 44795-4 42.
tics and the 7th International Joint Conference on
Natural Language Processing (Volume 1: Long Pa- Ikuya Yamada, Hiroyuki Shindo, Hideaki Takeda, and
pers). Association for Computational Linguistics, Yoshiyasu Takefuji. 2016. Joint Learning of the
Stroudsburg, PA, USA, volume 1, pages 1292–1300. Embedding of Words and Entities for Named Entity
https://doi.org/10.3115/v1/P15-1125. Disambiguation http://arxiv.org/abs/1601.01343.
Quoc V Le. 2013. Building high-level features us-
ing large scale unsupervised learning. In Acoustics,
Speech and Signal Processing (ICASSP), 2013 IEEE
International Conference on. IEEE, pages 8595–
8598.
911
Toponym Detection in the Bio-Medical Domain:
A Hybrid Approach with Deep Learning
912
Proceedings of Recent Advances in Natural Language Processing, pages 912–921,
Varna, Bulgaria, Sep 2–4 2019.
2005). A more specialised area included the cording as in SemEval (Section 4.2). Finally, in
automatic tracking of biological specimens from Section 5 we come to general conclusions about
different places around the world (Beaman and the project.
Conn, 2003). Today, the possible uses range
from analysing social media texts (Ireson and 2 Related Work
Ciravegna, 2010) to news streams (Lieberman and
Samet, 2012), where locations of large-scale ac- Detecting and resolving locations or toponyms has
tivity and problematic regions are mapped, re- undergone some changes in its approach. The
spectively. A further example is the bio-medical topic was first dealt with more extensively to-
domain, where the spreading of viruses can be wards the late 1990s and early 2000s (Larson,
tracked by analysing texts that mention locations 1996; Hill et al., 1999; Smith and Crane, 2001;
(Weissenbacher et al., 2019). However, as the ar- Leidner et al., 2004). These earliest approaches
eas of application are varied, the effectiveness of were aimed mainly at using geographical infor-
toponym resolvers is also said to vary among dif- mation for information retrieval, as well as cata-
ferent types of text (Gritta et al., 2018). logue searches in digital libraries. While the first
of these approaches used named entity tagging, as
This paper presents further research into TR, well as specially constructed gazetteers to detect
or to be more precise the detection of toponyms. (Larson, 1996; Hill et al., 1999), others went be-
Moreover, this research was carried out in the con- yond this and used a combination of methods to
text of SemEval-2019 Task 12: Toponym Res- disambiguate. Smith and Crane (2001) used NE
olution (Weissenbacher et al., 2019) and aimed tagging and a gazetteer to detect locations, and
specifically at subtask 2, which deals only with disambiguated these using information gathered
the detection of toponyms. To the best of our beforehand. This information, ”local” and ”doc-
knowledge, prior work on exploring how a ma- ument” context, is said to include co-occurring
chine learning classifier can be used together with words and other locations mentioned throughout
relatively simple string matching to detect loca- the text, respectively. The authors also use ”world
tions in texts has been limited. Previously, we knowledge” gathered from other sources, which
explored the use of machine learning classifiers mainly includes meta information such as coordi-
to predict a location within short word windows nates, size, corresponding political entities and so
in the context of SemEval-2019 Task 12 (Plum on (Smith and Crane, 2001). Similarly, Leidner
et al., 2019).We employ our system submitted to et al. (2004) used a combination of simple heuris-
the SemEval-2019 task as a baseline and make use tics, linguistic cues, co-occurrence statistics and
of the same dataset, consisting of texts from the discourse information to detect locations and as-
bio-medical domain Plum et al. (2019). While the sign coordinates.
system serving as a baseline was reasonably com-
TR has shifted from the methods of earlier ap-
petitive in terms of precision, it did not achieve
proaches and followed the trend of using machine
a high recall. The neural network architectures
learning techniques. Whereas in the past learning
that are used for this research are novel to this
techniques lacked data (Smith and Crane, 2001),
type of task even though some of the architectures
this is no longer the case. Approaches using ma-
have been used for sentence classification tasks
chine learning with (indirect) supervision include
like sentiment analysis, spam detection and so on.
Hu and Ge (2009) and Speriosu and Baldridge
The rest of the paper is structured as follows. (2013). Hu and Ge (2009) make use of hierar-
We present related work in the field first (Section chical structures ensuing from geographical re-
2), followed by the methodology employed (Sec- lations, an approach said to perform in an ac-
tion 3). This section includes a description of the curacy range of 73.55 to 85.38 percent on an
dataset, system and network architectures. Sec- Australian news corpus. Speriosu and Baldridge
tion 4 presents the overall results, which are split (2013) on the other hand, present their text-driven
into classifier results at the sentence level, i.e. dis- approach, which uses context information to re-
regarding the indexes and therefore not compara- solve toponyms. The classifiers themselves are
ble to the SemEval evaluation (Section 4.1) and trained mainly on semi-automatically generated
the overall results at the word level, where explicit data, obtained primarily from locations tagged in
indexes are retrieved in order to be evaluated ac- Wikipedia. While the aforementioned approach
913
relies on the availability of gazetteers, a gazetteer- enough to be used for other applications of NLP.
independent approach has been brought forward Our System described in Plum et al. (2019)
by DeLozier et al. (2015). This approach, which that was submitted to SemEval-2019 Task 12 will
also relies on a machine learning classifier to dis- serve as the baseline. The approach uses GATE
ambiguate toponyms, solely uses NER techniques with ANNIE to detect all the occurrences of loca-
to detect locations, and is said to perform on a tions in a text, using custom gazetteers based on
state-of-the-art level on the TR-CoNLL (Leidner, GeoNames. Several gazetteers were tested and the
2006) and Civil War corpora (DeLozier et al., best results achieved with a gazetteer of locations
2015). with a population of 15, 000 people or more. Fol-
Most recently, Gritta et al. (2018) have pre- lowing the string matching, two neural network
sented a survey of the current state of TR, which models are used to classify five-word windows
they refer to as geoparsing. While the two terms around the matched location. For each window a
are essentially synonymous, the authors use geop- prediction is made whether a real location is con-
arsing (or geoparsers) to refer to fully fledged end- tained or not. The method is reported to have a
to-end systems. These include CLAVIN3 , the Ed- significant drawback, since a five-word context is
inburgh Parser (Tobin et al., 2010; Grover et al., not enough to carry out proper classification, as the
2010), Topocluster (DeLozier et al., 2015) and location itself could, for instance, be a multi-word
GeoTxt (Karimzadeh et al., 2019). These sys- expression. Furthermore, the gazetteer matching
tems are said to perform at state-of-the-art lev- carried out beforehand severely limits the overall
els, and are tested on the Local Global Corpus, as recall of the approach, as it is not able to detect
well as a corpus compiled by the authors, which locations that are written across line-breaks, or are
is based on Wikipedia and GeoNames data (Gritta simply not contained in the gazetteer. In contrast
et al., 2018). However, it should be mentioned to this system, the approach proposed in this paper
that Topocluster and CLAVIN apply learning tech- predicts locations on a sentence-by-sentence basis,
niques. The Edinburgh Parser and GeoTxt rely on then attempts to retrieve the correct index of each
NER and heuristics to rank possible candidates. location by using a gazetteer lookup.
The evaluation of toponym resolvers is carried
out on specifically created datasets or corpora. 3 System Description
Leidner (2004) was the first to raise awareness This section describes the system we developed
for the need of a gold standard for these pur- for detecting toponyms in bio-medical texts. Our
poses. To this end, the paper describes the ongo- approach is based on our system submitted to
ing effort to create such a corpus, including a cus- SemEval-2019 Task 12, described in Plum et al.
tom markup language (TRML) and editor (TAME) (2019). It differs mainly in the order of the pro-
(Leidner, 2004). Later, Leidner (2006) describes cessing stages, as well as in the architectures that
the resulting corpus of the previous efforts. It were used. The previous system matches location
is based on news articles, with 6, 980 human- names using a gazetteer, followed by a machine
annotated instances of toponyms. The corpus, util- learning classifier to predict whether the matched
ising the CoNLL format, is still used today (De- location is a proper location or not (Plum et al.,
Lozier et al., 2015; Gritta et al., 2018). However, 2019). In the present approach, we use a machine
recently concerns have been raised again concern- learning classifier to predict whether a sentence
ing the availability of datasets for toponym resolu- contains a relevant location first, and on this pre-
tion or geoparsing by both Gritta et al. (2018) and selection we perform a gazetteer lookup to identify
Karimzadeh and MacEachren (2019). Gritta et al. the specific index range of each location. We also
(2018) have contributed their own dataset com- use spaCy NER4 to compare the effectiveness of
piled from Wikipedia. In addition, Karimzadeh the gazetteer.
and MacEachren (2019) present their tool GeoAn- The approach for the system is split into three
notator which has been developed to aid the com- steps, which are explained in the following three
pilation of such corpora. The tool is said to not sections. The first step deals with the preparation
only be useful for the creation of large-scale cor- of the texts from the dataset. This involves clean-
pora on a collaborative basis, but also versatile ing noisy sections of text and outputting an input
3 4
https://clavin.bericotechnologies.com/about-clavin/. Version 2.1.3, available at https://spacy.io/
914
file for the machine learning classifier containing 3.1.1 Text Cleaning
all texts split into sentences. Next, each classifier The cleaning of the texts is performed in two steps.
is trained and run in order to obtain predictions First, all line breaks are removed and replaced
at the sentence level. Finally, the output from the with spaces. This is mainly to deal with sentence
classifier is used in conjunction with a gazetteer splitting and string matching problems that could
lookup algorithm (and later spaCy NER) in order occur over line breaks. For instance, a line break
to determine the exact indexes of the detected lo- character between New and York would lead to this
cations. location to be detected as York, not as New York.
The line breaks have to be replaced by one space,
3.1 Dataset as the annotations take line breaks into account
To ensure comparability with the baseline system, and add these to the index range of a location. It
we work with the same dataset from SemEval- should be mentioned that this requires the texts to
2019 Task 12. This dataset is made up of 150 use LF-type line breaks, as any other type would
journal articles from PubMed Central and are from require a different number of replacement charac-
the domain of epidemiology (Weissenbacher et al., ters. For this dataset, it was ensured that all files
2019). As mentioned previously, the main idea be- conform to this standard.
hind detecting locations in texts from this domain Next, we carry out more methods to clean the
is to track the spreading of viruses (Weissenbacher texts. Using the guidelines set out by the task or-
et al., 2019). The articles were downloaded as ganisers and a brief analysis carried out by Plum
PDFs and converted to plain text using a pdf-to-txt et al. (2019), we determined that certain parts of
tool by the organisers of the task. The toponyms the texts are not relevant for detection. This in-
were manually disambiguated by the organisers cludes references and certain character and word
and subsequently annotated using the Brat anno- strings that describe biological genome sequences.
tator (Stenetorp et al., 2012). Two texts of the As these often include toponyms that were ex-
training set were removed, as they were unread- cluded from annotation in the SemEval-2019 task,
able, probably caused by PDF to text conversion these could also be disregarded. In order to re-
problems. Apart from this, we work with the same move all irrelevant parts but retain indexing con-
training and test splits as supplied by the organ- sistency, we use regular expressions to find and
isers, which are 73 and 45 texts, respectively. It replace certain text sequences with an equivalent
should be pointed out that we had to adjust the an- number of spaces. As before, we replace each
notations of some of the training texts, as these character with a space in order to ensure that the
were carried out on texts using CRLF-type line indexes match up with the annotations.
breaks5 and did not match the indexes read by our We wanted to test the effectiveness of our clean-
system, as it used LF-type line breaks, which lead ing methods. Therefore, we tested our methods
to the indexes being offset. with both types of text: texts where only the line
The texts had to be prepared for the classifica- breaks have been removed and texts that have been
tion task. As some parts of the texts had been completely cleaned. During testing it was clear
deemed irrelevant by the organisers of SemEval- that the performance was better on fully cleaned
2019 Task 12 (Weissenbacher et al., 2019), we texts, due to the reduction in seemingly random
had to remove these. This includes the references strings that could be detected (i.e. the string [..]
and acknowledgement sections of each article. We ACG GGG MA AUA UGC [..] could produce the
also had to remove certain character strings which match MA, as in the U.S. state Massachusetts).
are specific to texts from the bio-medical domain. 3.1.2 Sentence Splitting
Finally, the texts had to be split into sentences and
As the identification of locations primarily hap-
stored with further information in order to be clas-
pens at the sentence level, each individual text
sified in the next step.
needs to be split into sentences. We use spaCy
5
CRLF-type line breaks are commonly used in text files in order to complete this task. The sentences are
created with Microsoft DOS/Windows operating systems. then output to a CSV file, containing information
This type of line break uses two characters to denote the on each sentence’s text id, as well as its own spe-
end of a line. LF-type line breaks are commonly used
on UNIX/Linux-based operating systems, and only use one cific index range. This information is necessary at
character to denote the end of a line. the stage of identifying the exact index range of
915
a location and then producing the annotation files 3.2.1 Pooled GRU
for the SemEval evaluation script. For the machine This model takes pre-trained fasttext embeddings
learning training file, a separate column indicating (Mikolov et al., 2018) as a matrix for the in-
whether or not a location is contained in the sen- put which is comprised of the vertically stacked
tence is also included. embedding vectors corresponding to the words
present in the sentence. The matrix can be
thought of as a sequence of embedded words.
3.2 Sentence Level Location Identification Each of these embedding vectors is fed to the bi-
directional GRU (Chung et al., 2014) at their re-
Once the texts have been pre-processed, we run spective timestep. The final timestep output is fed
a binary classification on the sentences, predict- into a max pooling layer and an average pooling
ing whether a sentence contains a location, or not. layer in parallel (Scherer et al., 2010). Following
Table 1 shows some examples from the training this, the outputs of the two pooling layers are con-
dataset which is used for this classification task. catenated and connected to a dense layer (Huang
Out of the 17, 535 sentences in the training et al., 2017) activated with a sigmoid function.
set, only 2, 117 sentences contain locations. This Additionally, there is a spatial dropout (Tompson
means that the dataset is highly unbalanced, thus et al., 2015) between the embedding layer and the
making the classification task quite difficult. The bi-directional GRU layer to avoid over-fitting.
increased difficulty was also caused by the lan- The network was trained using adam optimiser
guage used. As an example, in the row with the (Kingma and Ba, 2015), with a reduced learning
ID PMC2857219, shown in Table 1, the sentence rate once learning stagnates. This model has been
contains Korea and is not annotated as a location discussed in Kowsari et al. (2019) as a common
since it is a part of an organisation name. Also, in model to perform text classification tasks.
the row with the ID PMC2857219, US is not anno-
tated as a location, as it functions as an adjective 3.2.2 Stacked LSTM with Attention
to the word soldier. General named entity recog- As with the previous model, this model takes pre-
nisers such as spaCy or gazetteer matching could trained fasttext embeddings (Mikolov et al., 2018)
possibly falsely detect that these sentences contain as an input. Each of these embedding vectors
a proper location as defined for this task. Fur- are then fed into a bi-directional LSTM (Schus-
thermore, sentences similar to ID PMC5837706, ter and Paliwal, 1997). The output of this layer
shown in 1, caused difficulties, as the text in the is again fed into a bi-directional LSTM (Schus-
sentence was not clear. Considering all of these is- ter and Paliwal, 1997) with self attention (Vaswani
sues, we need an intelligent classifier that detects et al., 2017). Finally, the output is connected to
whether a sentence contains a location or not, by two dense layers that are (Huang et al., 2017) ac-
considering the words that appear in the sentence. tivated first with a relu function, and then with a
sigmoid function.
We use five different recurrent network archi-
Again, this network was trained using adam op-
tectures to perform the binary classification task
timiser (Kingma and Ba, 2015), with a reduced
on the sentences: pooled Gated Recurrent Unit
learning rate once learning stagnates. We adopted
(GRU) (3.2.1), stacked Long Short-Term Memory
this model from the Toxic Comment Classification
(LSTM) with attention (3.2.2), LSTM and GRU
Challenge in Kaggle6 .
with attention (3.2.3), 2D convolution with pool-
ing (3.2.4) and GRU with capsule (3.2.5). Each 3.2.3 LSTM and GRU with Attention
classifier was run on prepared and cleaned text (as
explained in Section 3.1.1). These models were This architecture applies a spatial dropout to
successfully applied to a number of classification the embedding layer (Tompson et al., 2015).
tasks such as GRU for sequence labeling Chung The output is then fed in parallel to a bi-
et al. (2014), LSTM for semantic similarity and directional LSTM-layer (Schuster and Paliwal,
word analogy Coates and Bollegala (2018), and 1997) with self attention and a bidirectional GRU-
GRU with capsule for toponym detection Plum layer (Chung et al., 2014) with self attention
et al. (2019). Their success in these tasks inspired 6
https://www.kaggle.com/c/jigsaw-toxic-comment-
us to try them for our problem. classification-challenge
916
id start end sentence loc? location
Dr Jin-Won Song is a professor of
PMC2857219 15686 15753 0 NA
microbiology at Korea University.
kConFab recruited multi-generational,
PMC5005937 9817 9928 multiple-case families through cancer 1 Australia
family clinics in Australia.
These data showed the epidemiologic
link between US soldier patients and
PMC2857219 14913 14947 1 South Korea
rodent hosts at the training sites near
the Demilitarized Zone in South Korea.
PMC5837706 14489 14531 (46.9) 395 (39.4) 75 (7.5) 245 (24.4) 329 0 NA
Table 1: Example rows in the training set. Sentences containing a location are represented with 1 in the
loc? column and 0 otherwise. The location column contains the explicit location names contained in the
sentence.
(Vaswani et al., 2017). The output from the bi- 3.2.5 GRU with Capsule
directional GRU-layer is fed into an average pool- Most of the previous architectures rely on a pool-
ing layer and a max pooling layer. The output from ing layer. However, this architecture uses a cap-
these layers and the output of the bi-directional sule layer (Hinton et al., 2018) rather than pooling
LSTM-layer are concatenated and connected to layers. After applying a spatial dropout (Tompson
a dense layer with relu activation. After that, a et al., 2015) the output of the embedding layer is
dropout (Srivastava et al., 2014) is applied to the fed into a bi-directional GRU-layer (Chung et al.,
output and connected to a dense layer activated 2014). The output is then connected to a capsule
with a sigmoid function. layer (Hinton et al., 2018). The output of the cap-
While this network was also trained using adam sule layer is flattened and connected to a dense
optimiser (Kingma and Ba, 2015), it was trained layer with relu activation, a dropout (Srivastava
with a cyclical learning rate (Smith, 2017) this et al., 2014) and batch normalisation applied, and
time. Plum et al. (2019) has used this model to re-connected to a dense layer with sigmoid activa-
predict whether a word window contains a loca- tion.
tion or not. The capsule network was trained using adam
optimiser (Kingma and Ba, 2015), with a reduced
3.2.4 2D Convolution with Pooling learning rate once learning stagnates. This model
has been used in Plum et al. (2019) to predict
The fourth architecture takes a different approach
whether a word window contains a location or not.
than the previous architectures by using 2D convo-
lution layers (Wu et al., 2018), rather than LSTM- 3.3 Word Level Location Identification
or GRU-layers. The outputs of the embedding lay-
ers are connected to four 2D convolution layers The location predictions made by the ML-
(Wu et al., 2018), each with max pooling layers. classifier at the sentence level are passed as a
The outputs of these are concatenated and con- CSV file to the string matching script. It runs
nected to a dense layer activated with a sigmoid through each sentence, matching locations from
function after applying a dropout (Srivastava et al., a gazetteer. For matching the strings we use a
2014). fast and efficient Aho-Corasick algorithm (Aho
and Corasick, 1975). The implementation used is
This network also uses adam optimiser
available for the Python programming language8
(Kingma and Ba, 2015) and a reduced learning
We use a large gazetteer that is comprised of
rate once learning stagnates. This model has
the full list of all locations from the GeoNames
been used in the Quora Insincere Questions
database9 . The main idea behind this is that we
Classification Kaggle competition7 .
want to achieve the highest chance of detecting
7 8
https://www.kaggle.com/c/quora-insincere-questions- https://github.com/WojciechMula/pyahocorasick/
9
classification http://www.geonames.org/, last accessed 21.05.2019
917
every location. However, as the gazetteer is so which are described in the next section. It should
large, is causes a lot of ”noise” during the string be mentioned that we did run some word level pre-
matching, including partial matches and numbers dictions on the output of the other classifiers, but
that have no meaning. In order to combat this, as expected the results were always much lower,
we apply several filters after finding matches, in due to the decreased starting point.
order to exclude certain results. We consciously
avoided removing anything from the gazetteer it- 4.2 Toponym Identification
self, as this would be either time-consuming (man-
ual) or could falsely remove desired locations (au- As mentioned previously, we regard the word level
tomatic). Therefore, we use filters to remove all predictions as the overall result of the system. The
matches with numbers (i.e. 717, 7Palms, 50da), evaluation was carried out in accordance with pa-
strings shorter than three characters which are not rameters set out for SemEval-2019 Task 12, fea-
both uppercase (i.e. Bl, b1, al but not AL, CA, NY), turing strict and overlap categories on macro and
sub-string matches (i.e. London in Londonderry) micro levels. For the strict measure, predicted lo-
and all lowercase strings (which are usually frag- cations are only considered as correct if the text
ments of location names left in the database, as span matches the gold standard exactly. For the
in paseo caribe). The resulting tables are sorted overlap measure, predictions are considered to
by index, and duplicates removed. Where matches be correct if they share a common span of text
overlap in terms of indexing, we give preference to with the gold annotations. The python script was
the longest match. This ensures that in sentences made available on Bitbucket10 by the SemEval-
such as I live in New York, we detect only New 2019 Task 12 organizers.
York and not York (as these are separate entries in Table 3 shows the results for both the string
the gazetteer). matching method using gazetteers to extract the
For comparison purposes, we also employ locations with a custom script for indexes, as well
spaCy to detect locations in the sentences at this as the spaCy NER algorithm (only considering
stage. We used the standard English web corpus locations) and the baseline system submitted to
and the spaCy NER algorithm. SemEval-2019. Our best results were achieved in
the overlap macro classes, and are highlighted in
4 Results bold. Overall, while we were not able to beat the
best precision score of Plum et al. (2019), we came
As our system operates at two levels, we first quite close. Nonetheless, we were able to improve
present the results of location prediction at the sen- the recall significantly, as well as the overall f-
tence level using the five different recurrent net- score.
work architectures. Next, we present the results
The trade-offs that each approach brings with
of the prediction at the word level. These results
it should become clear when regarding the re-
are also regarded as our final results, as these pre-
sults.The approach using the GeoNames gazetteer
dictions yielded the index range of each location,
detects a higher number of locations overall. This
which were evaluated with an evaluation script.
is due to the simplistic string matching method
The evaluation script that was utilised was the one
backed by such a large gazetteer, and comes at the
provided for SemEval-2019 Task 12.
cost of overall precision. The spaCy NER algo-
4.1 Sentence Level Prediction rithm is much more precise, but is more limited
in terms of recall. We find it most likely that this
Results of the sentence level predictions for the approach does not tag many locations as such, be-
test set are shown in Table 2. We use precision cause the texts are still quite noisy, and because we
and recall to evaluate the results. The third model did not train it on our dataset. Due to the small size
described, LSTM and GRU with Attention, pro- and unbalanced nature of our dataset, we did not
vided the best results for the cleaned text. Despite consider training spaCy any further. In the future,
the dataset being quite unbalanced, the model re- given an appropriate dataset from the bio-medical
ported good precision and recall scores of 0.852 domain, this could perhaps lead to better results.
and 0.853, which provided a high F1 score, too.
As this is the best model, we use these predic- 10
https://bitbucket.org/dweissen/semevaltask
tions as the basis for our word level predictions, 12evaluator/src/master/
918
Uncleaned Cleaned
Model
P R F1 P R F1
Pooled GRU .755 .792 .772 .789 .816 .793
Stacked LSTM with Attention .795 .784 .789 .826 .796 .813
LSTM and GRU with Attention .811 .840 .825 .852 .853 .852
2D Convolution with Pooling .802 .743 .769 .842 .758 .792
GRU with Capsule .843 .816 .829 .865 .823 .842
All sentences predicted 1 .042 .500 .078 .060 .500 .107
All sentences predicted 0 .457 .500 .478 .439 .500 .468
Table 2: Results for the sentence level classification. We report Precision (P), Recall (R), and F1 for each
model (bold indicates the best set of results). Two baseline predictions with all sentences predicted 1 and
all sentences predicted 0 are also reported for comparison.
Table 3: Results for the word level classification. We report the same measures as previously, for the
categories described in Section 4. Results are shown for both approaches (bold indicates the best set of
results).
919
References Morteza Karimzadeh and Alan M. MacEachren. 2019.
GeoAnnotator: A Collaborative Semi-Automatic
Alfred V. Aho and Margaret J. Corasick. 1975. Ef- Platform for Constructing Geo-Annotated Text Cor-
ficient string matching: An aid to bibliographic pora. ISPRS International Journal of Geo-
search. Commun. ACM 18(6):333–340. Information 8(4).
Reed S. Beaman and Barry J. Conn. 2003. Automated Morteza Karimzadeh, Scott Pezanowski, Alan M.
geoparsing and georeferencing of Malesian collec- MacEachren, and Jan O. Wallgrn. 2019. Geotxt: A
tion locality data. Telopea 10(1):43–52. scalable geoparsing system for unstructured text ge-
olocation. Transactions in GIS 23(1):118–136.
Junyoung Chung, aglar Gülehre, Kyunghyun Cho,
and Yoshua Bengio. 2014. Empirical Evaluation Diederik P. Kingma and Jimmy Ba. 2015. Adam: A
of Gated Recurrent Neural Networks on Sequence Method for Stochastic Optimization. Proceedings
Modeling. CoRR . of CoRR 2015 .
Paul Clough. 2005. Extracting metadata for spatially- Kamran Kowsari, Kiana Jafari Meimandi, Mojtaba
aware information retrieval on the internet. In Pro- Heidarysafa, Sanjana Mendu, Laura E. Barnes, and
ceedings of GIR’05. Donald E. Brown. 2019. Text Classification Algo-
rithms: A Survey. Information 10(4).
Joshua Coates and Danushka Bollegala. 2018. Frus-
Ray R Larson. 1996. Geographic information retrieval
tratingly easy meta-embedding - computing meta-
and spatial browsing. Geographic information sys-
embeddings by averaging source word embeddings.
tems and libraries: patrons, maps, and spatial infor-
In Proceedings of NAACL-HLT 2018.
mation .
Grant DeLozier, Jason Baldridge, and Loretta London. Jinhyuk Lee, Wonjin Yoon, Sungdong Kim,
2015. Gazetteer-independent toponym resolution Donghyeon Kim, Sunkyu Kim, Chan Ho So, and
using geographic word profiles. In Proceedings of Jaewoo Kang. 2019. Biobert: a pre-trained biomed-
AAAI 2015. ical language representation model for biomedical
text mining. arXiv preprint arXiv:1901.08746 .
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2018. Bert: Pre-training of deep Jochen L Leidner. 2004. Towards a reference cor-
bidirectional transformers for language understand- pus for automatic toponym resolution evaluation. In
ing. ArXiv abs/1810.04805. Proceedings of GIR’04.
Milan Gritta, Mohammad Taher Pilehvar, Nut Lim- Jochen L. Leidner. 2006. An evaluation dataset for the
sopatham, and Nigel Collier. 2018. Whats missing toponym resolution task. Computers, Environment
in geographical parsing? Language Resources and and Urban Systems 30(4):400 – 417. Geographic
Evaluation 52(2):603–623. Information Retrieval (GIR).
Claire Grover, Richard Tobin, Kate Byrne, Matthew Jochen L Leidner et al. 2004. Toponym resolution
Woollard, James Reid, Stuart Dunn, and Julian in text:Which Sheffield is it?. In Proceedings of
Ball. 2010. Use of the Edinburgh geoparser GIR’04.
for georeferencing digitized historical collections.
Philosophical Transactions of the Royal Society A: Michael D. Lieberman and Hanan Samet. 2012. Adap-
Mathematical, Physical and Engineering Sciences tive Context Features for Toponym Resolution in
368(1925):3875–3889. Streaming News. In Proceedings of ACM SIGIR.
Tomas Mikolov, Edouard Grave, Piotr Bojanowski,
Linda L Hill, James Frew, and Qi Zheng. 1999. Geo- Christian Puhrsch, and Armand Joulin. 2018. Ad-
graphic names. D-lib Magazine 5(1):17. vances in Pre-Training Distributed Word Represen-
tations. In Proceedings of LREC 2018.
Geoffrey E. Hinton, Sara Sabour, and Nicholas Frosst.
2018. Matrix capsules with EM routing. In Pro- Jakub Piskorski and Roman Yangarber. 2013. Infor-
ceedings of ICLR 2018. mation extraction: Past, present and future. In
Multi-source, multilingual information extraction
You-Heng Hu and Linlin Ge. 2009. A Supervised Ma- and summarization, Springer, pages 23–49.
chine Learning Approach to Toponym Disambigua-
tion, Springer London, pages 117–128. Alistair Plum, Tharindu Ranasinghe, Pablo Calleja,
Constantin Orasan, and Ruslan Mitkov. 2019.
Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. RGCL-WLV at SemEval-2019 Task 12: Toponym
2017. Densely Connected Convolutional Networks. Detection. In Proceedings of SemEval-2019.
Proceedings of IEEE CVPR 2017 .
Dominik Scherer, Andreas C. Müller, and Sven
Neil Ireson and Fabio Ciravegna. 2010. Toponym Res- Behnke. 2010. Evaluation of Pooling Operations in
olution in Social Media. In Proceedings of ISWC Convolutional Architectures for Object Recognition.
2010. In Proceedings of ICANN 2010.
920
Mike Schuster and Kuldip K. Paliwal. 1997. Bidirec-
tional recurrent neural networks. IEEE Trans. Sig-
nal Processing 45:2673–2681.
David A. Smith and Gregory Crane. 2001. Disam-
biguating Geographic Names in a Historical Digital
Library. In Proceedings of ECDL 01.
Leslie N. Smith. 2017. Cyclical Learning Rates for
Training Neural Networks. In Proceedings of IEEE
WACV 2017.
921
Combining SMT and NMT Back-Translated Data for Efficient NMT
Abstract lel corpus from L1 to L2, with the source (L1) side
being the translated text and the target side being
Neural Machine Translation (NMT) mod- the monolingual data. Back-translation has been
els achieve their best performance when shown to be beneficial not only for MT but also
large sets of parallel data are used for train- for other NLP tasks where data is scarce, e.g. auto-
ing. Consequently, techniques for aug- matic post-editing (APE) (Junczys-Dowmunt and
menting the training set have become pop- Grundkiewicz, 2016; Negri et al., 2018). How-
ular recently. One of these methods is ever, the effects of various parameters for creating
back-translation (Sennrich et al., 2016a), back-translated (BT) data have not been investi-
which consists on generating synthetic gated enough as to indicate what are the optimal
sentences by translating a set of mono- conditions in not only creating but also employ-
lingual, target-language sentences using a ing such data to train high-quality neural machine
Machine Translation (MT) model. translation (NMT) systems.
Generally, NMT models are used for back-
translation. In this work, we analyze the
performance of models when the training The work presented in Poncelas et al. (2018)
data is extended with synthetic data using draws an early-stage empirical roadmap to investi-
different MT approaches. In particular we gating the effects of BT data. In particular, it looks
investigate back-translated data generated at how the amount of BT data impacts the perfor-
not only by NMT but also by Statistical mance of the final NMT system. In Sennrich et al.
Machine Translation (SMT) models and (2016a) and Poncelas et al. (2018), the systems
combinations of both. The results reveal used to generate the BT data are neural. However,
that the models achieve the best perfor- it has been noted that often different paradigms
mances when the training set is augmented can contribute differently to a given task. For ex-
with back-translated data created by merg- ample, it has been shown that applying an APE
ing different MT approaches. system based on NMT technology improves sta-
tistical machine translation (SMT) output, but has
1 Introduction lower impact on NMT output (Bojar et al., 2017;
Chatterjee et al., 2018).
Machine translation (MT) nowadays is heavily
dependent on the quantity and quality of train-
ing data. The amount of available good-quality In this work we assess the impact of different
parallel data for the desired domain and/or lan- amounts of BT data generated by two different
guage pair is often insufficient to reach the re- types of MT systems – NMT and SMT. Our con-
quired translation performance. In such cases, it tribution is two-fold: (i) we provide a systematic
has become the norm to resort to back-translating comparison of the BT data by building NMT sys-
freely available monolingual data as proposed in tems with a combination of SMT and NMT BT
(Sennrich et al., 2016a). That is, one can trans- data and (ii) we identify the effects of BT data that
late a set of sentences from language L2 into L1 originates from SMT or NMT on the end-quality
with an already trained MT system for the lan- of the trained NMT system. We aim to answer the
guage pair L2→L1. Then create a synthetic paral- question: "What is the best choice for BT data?"
922
Proceedings of Recent Advances in Natural Language Processing, pages 922–931,
Varna, Bulgaria, Sep 2–4 2019.
2 Preparatory Study: the Effect of the extra training effort is not required as no sig-
Back-Translation when Controlling for nificant improvement has been observed. Based
the Amount of Training Effort on the outcome of these experiments we chose the
rest of our experiments.
A typical assumption made when training NMT
models, is that when more training data is used, 3 Using Back-Translation from Different
more training effort is warranted. Based on this Sources
assumption when training NMT systems what is
normally kept constant is the amount of training The work of (Sennrich et al., 2016a) showed
epochs rather than the amount of training effort that adding BT data is beneficial to achieve bet-
in the form of steps/mini-batches. Nevertheless, ter translation performances. In this work we
when adding back-translated data to the training compare the details related to the translation hy-
set, while keeping the amount of epochs the same, potheses originating from SMT and NMT back-
the effective amount of training increases. It could translated training data as well as combine the data
then be questioned whether the extra training ef- from those two different sources. To the best of
fort in itself does not partly explain the positive our knowledge, this has not been investigated yet.
effect of back-translation. For this reason, we seek We compare German-to-English translation hy-
to answer the question: “Does the effect of back- potheses generated by systems trained (i) only on
translation change when we control for the amount authentic data, (ii) only on synthetic data, and (iii)
of training effort, by keeping the total amount on authentic data enhanced with different types
of steps/mini-batches constant?". To answer this of BT data: SMT, NMT. We exploit two types
question we compare the performance of systems of synthetic and authentic data combinations: (a)
trained on purely authentic data to those trained on randomly selected half of target sentences back-
authentic plus synthetic data, while keeping either translated by SMT and another half by NMT sys-
the number of steps/mini-batches or the number of tem, and (b) joining all BT data (thus repeating
epochs constant in both settings: each target segment).
The translation hypotheses are compared in
1. Models trained with 1M auth + 2M synth sen- terms of four automatic evaluation metrics:
tences using the default settings, including 13 BLEU (Papineni et al., 2002), TER (Snover et al.,
training epochs. 2006), METEOR (Banerjee and Lavie, 2005) and
2. Models trained on 1M auth data only, trained CHRF (Popović, 2015). These metrics give an
either: overall estimate of the quality of the translations
with respect to the reference (human translation of
(a) using the default settings, including 13 the test set). In addition, the translation hypothe-
training epochs. ses are analyzed in terms of five error categories,
(b) Trained for 39 epochs, to obtain a same lexical variety and syntactic variety.
amount of training effort as for the 1M
auth + 2M synth sentences model. 4 Related Work
When increasing the epochs to 39, we take ap- A comparison between MT models trained with
propriate measures to keep the starting point and synthetic and with authentic data that originate
speed of decay of the learning rate constant for the from the same source has been presented in Pon-
amount of training steps/epochs.1 celas et al. (2018). They show that while the per-
The results of these experiments indicate that formances of models trained with both synthetic
training a model on authentic data with 1/3 of the and authentic data are better than those of models
amount of the total parallel data (authentic + syn- trained with only authentic data, there is a satura-
thetic) for an additional 26 epochs to account for tion point beyond which the quality does not im-
1 prove by adding more synthetic data. Nonetheless,
This is implemented by changing the start of the learning
rate decay from epoch 8 to epoch √22 (= 7 ∗ 3 + 1) and chang- models trained only with synthetic (BT) data per-
ing the decay factor from 0.5 to 3 0.5 = 0.7936. This way, form very reasonably, with evaluation scores be-
the learning rate decay starts after the same amount of data ing close to those of models trained with only au-
when using the 1M auth dataset (7 × 3M ) and the decay rate
is maintained at 0.5 for each 3M sentences from this point thentic parallel data. In fact, when appropriately
onwards. selected, BT data can be used to enhance NMT
923
models (Poncelas et al., 2019). tic gradient descent (SGD), in combination with
Edunov et al. (2018) confirmed that synthetic learning rate decay, halving the learning rate start-
data can sometimes match the performance of au- ing from the 8th epoch.
thentic data. In addition, a comprehensive analysis In order to build the models, all data sets are
of different methods to generate synthetic source tokenized and truecased and segmented with Byte-
sentences was carried out. This analysis revealed Pair Encoding (BPE) (Sennrich et al., 2016b) built
that sampling from the model distribution or nois- on the joint vocabulary using 89500 merge op-
ing beam outputs out-performs pure beam search, erations. For testing the models we use the test
which is typically used in NMT. Their analysis set provided in the WMT 2015 News Translation
shows that synthetic data based on sampling and Task (Bojar et al., 2015). As development set, we
noised beam search provides a stronger training use 5K randomly sampled sentences from devel-
signal than synthetic data based on argmax infer- opment sets provided in previous years of WMT.
ence.
One of the experiments reported in Burlot and 6 Data
Yvon (2018) is comparing performance between The parallel data used for the experiments has
models trained with NMT and SMT BT data. been obtained from WMT 2015 (Bojar et al.,
The best Moses system (Koehn et al., 2007) is al- 2015). We build two parallel sets with these sen-
most as good as the NMT system trained with the tences: base (1M sentences) and auth (3M sen-
same (authentic) data, and much faster to train. tences). We use the target side of auth to create
Improvements obtained with the Moses system the following datasets:
trained with a small training corpus are much
smaller; this system even decreases the perfor- • SMTsynth: Created by translating the target-
mance for the out-of-domain test. The authors also side sentences of auth. The model used
investigated some properties of BT data and found to generate the sentences is an SMT model
out that the back-translated sources are on average trained with base set in the English to Ger-
shorter than authentic ones, syntactically simpler man direction. It has been built using the
than authentic ones, and contain smaller number Moses toolkit with default settings, using
of rare events. Furthermore, automatic word align- GIZA++ for word alignment and tuned using
ments tend to be more monotonic between arti- MERT (Och, 2003)). The language model
ficial sources and authentic targets than between (of order 8) is built with the KenLM toolkit
authentic sources and authentic targets. (Heafield, 2011) using the German side of
Burlot and Yvon (2018) also compared training base.
BT data with authentic data in terms of lexical and • NMTsynth: Created by translating the target-
syntactic variety, segment length and alignment side sentences of auth. The model used to
monotony, however they did not analyze the ob- generate the sentences is an NMT model
tained translation hypotheses. In (Vanmassenhove (with the same configuration as described in
et al., 2019) it is shown that MT systems trained on Section 5 but in the English to German direc-
authentic and on backtranslated data lead to gen- tion) trained with the base set.
eral loss of linguistic richness in their translation
hypotheses. • hybrNMTSMT: Synthetic parallel corpus
combining NMTsynth and SMTsynth sets. It
5 Experimental Settings has been built by maintaining the same target
side of auth, and as source side we alternate
For the experiments we have built German-to- between NMTsynth and SMTsynth each 500K
English NMT models using the Pytorch port of sentences.
OpenNMT (Klein et al., 2017). We use the de-
fault parameters: 2-layer LSTM with 500 hidden • fullhybrNMTSMT: Synthetic parallel corpus
units. The models are trained for the same number combining all segments from NMTsynth and
of epochs. As the model trained with all authentic SMTsynth sets (double size, each original
data converges after 13 epochs, we use that many target sentence repeated twice with both an
iterations to train the models (we use the same NMT and SMT back-translation-generated
amount of epochs). As optimizer we use stochas- translation).
924
7 Experiments The results show that adding synthetic data has
a positive impact on the performance of the mod-
In our experiments, we build models on differ- els as all of them achieve improvements when
ent portions of the datasets described in Section compared to that built only with authentic data 1M
6. First, we train an initial NMT model using the base. These improvements are statistically signif-
base data set. Then, in order to investigate how icant at p=0.01 (computed with multeval (Clark
much the models benefit from using synthetic data et al., 2011) using Bootstrap Resampling (Koehn,
generated by different approaches, we build mod- 2004)). However, the increases of quality are dif-
els with increasing sizes of data (from the data sets ferent depending on the approach followed to cre-
described in Section 6). ate the BT data.
The models explored are built with data that
First, we observe that models in which SMT-
ranges from 1M sentences (built with only authen-
generated data is added do not outperform the
tic data from base data set) to 4M sentences (con-
models built with the same size of authentic data.
sisting on 1M sentences from base and 3M sen-
For example, the models built with 4M sentences
tences generated artificially with different mod-
(1M authentic and 3M SMT-produced sentences,
els). We also include the models built with the
in cell + 3M SMTsynth) achieve a performance
fullhybrNMTSMT set. As this set contains dupli-
comparable to the model trained with smaller
cated target-side sentences, the largest model we
number of sentences of authentic data (such as +
build contains 7M sentences in total but only 4M
1M auth cell, 2M sentences).
distinct target-side sentences.
Models built by using NMT-created data have
8 Results a better performance than those built with data
generated by SMT. When performing a pair-
8.1 Controlling the Amount of Training wise comparison between models using an equal
Effort amount of either SMT or NMT-created data, we
Table 1 shows the effect of controlling the amount observe that the latter models outperform the for-
of training effort when using back-translation. It mer by around one BLEU point. In fact, the per-
can be observed that increasing the number of formance of models using NMT-translated sen-
epochs from 13 to 39 when using just the 1M base tences is closer to those built with authentic data,
training set does not increase the performance over and some NMTsynth models produce better trans-
using just 13 epochs (i.e. not compensating the lation qualities. This is the case of +1M NMT-
relatively smaller training set with more epochs), synth model (according to all evaluation metrics)
rather it deteriorates it. From these results we con- or +3M NMTsynth (according to BLEU).
clude that there is no reason to believe that the Our experiments also include the performance
positive effects of using back-translation is caused of models augmented with a combination of SMT-
by an effectively larger training effort, rather than and NMT- generated data. We see that adding hy-
by the advantage of the larger training set itself. brNMTSMT data, with one half of the data orig-
We therefore also conclude that it is reasonable to inating from SMT and the other half from NMT
keep the number of epochs constant across exper- models, have performances similar to those mod-
iments, rather than fixing the amount of training els built on authentic data only. According to some
effort as measured by steps/mini-batches, and we evaluation metrics, such as METEOR, the perfor-
do the former throughout the rest of the paper. mance is better than auth models when adding
1M or 2M artificial sentences (although none of
8.2 Addition of Synthetic Data from SMT these improvements are statistically significant at
and NMT Models p=0.01). For these amount of sentences, it also
Table 2 shows the results of the performance of the outperforms those models in which only SMT or
different NMT models we have built. The sub- only NMT BT data have been included.
tables indicate the size of the data used for build- The models extended with synthetic data
ing the models (from 1M to 4M lines). In each that perform best are fullhybrNMTSMT mod-
column it is indicated whether base has been aug- els. Furthermore, they also outperform authen-
mented with the auth, SMTsynth, NMTsynth, hy- tic models when built with less than 4M distinct
brNMTSMT, or fullhybrNMTSMT data set. target-sentences according to BLEU, METEOR
925
1M base.- 13 1M base.- 39 1M base + 2M
Epochs Epochs- NMTsynth
BLEU↑ 23.40 23.22 25.44
TER↓ 57.23 58.21 55.62
METEOR↑ 28.09 27.75 29.47
CHRF1↑ 50.66 50.18 52.5
Table 1: Results for experimental procedure validation: checking that it is reasonable to use constant
number of epochs, not constant amount of training effort, in the experiments.
1M base. - - - -
BLEU↑ 23.40 - - - -
1M lines
TER↓ 57.23 - - - -
METEOR↑ 28.09 - - - -
CHRF1↑ 50.66 - - - -
+ 1M auth + 1M SMT- +1M NMT- + 1M hybrN- + 2M fullhy-
synth synth MTSMT brNMTSMT
BLEU↑ 24.87 24.38 25.32 25.21 25.34
2M lines
Table 2: Performance of models built with increasing sizes of authentic set (first column) and different
synthetic datasets (last four columns). +1M, +2M and +3M indicate the amount of sentences added to
the base set (1M authentic sentences).
(showing statistically significant improvements at sion, addition and lexical errors, and we compared
p=0.01) and CHRF1. Despite that, when using lexical and syntactic variety of different outputs
large sizes of data (i.e. adding 3M synthetic sen- in terms of vocabulary size and number of dis-
tences) the models built with SMT-generated arti- tinct POS n-grams. We also analyzed the sentence
ficial data have the lowest performances whereas lengths in different translation hypotheses, how-
the performance of the other three tends to be sim- ever no differences were observed, neither in the
ilar. average sentence length nor in the distribution of
different lengths.
8.3 Further Analysis Automatic Error Analysis
In order to better understand the described sys- For automatic error analysis results, we used Hjer-
tems, we carried out more detailed analysis of all son (Popović, 2011), an open-source tool based on
translation outputs. We analyzed five error cate- Levenshtein distance, precision and recall. The re-
gories: morphological errors, word order, omis- sults are presented in Table 3.
926
error class rates↓
training morph order omission addition mistranslation
1M base 2.8 9.8 12.0 4.8 29.1
1M base + 1M auth 2.7 9.5 11.4 4.9 28.2
1M base + 1M SMTsynth 2.8 10.0 11.6 4.8 28.1
1M base + 1M NMTsynth 2.7 9.8 10.9 5.0 28.1
1M base + 1M hybrNMTSMT 2.7 9.6 11.4 5.2 27.7
1M base + 1M fullhybrNMTSMT 2.6 9.5 11.0 5.2 27.8
1M base + 2M auth 2.6 9.6 11.2 4.8 27.7
1M base + 2M SMTsynth 2.7 10.0 11.9 4.5 28.0
1M base + 2M NMTsynth 2.6 9.7 11.1 5.1 27.9
1M base + 2M hybrNMTSMT 2.6 9.6 11.0 5.2 27.6
1M base + 2M fullhybrNMTSMT 2.6 9.6 10.7 5.3 27.4
1M base + 3M auth 2.7 9.8 11.2 4.6 27.6
1M base + 3M SMTsynth 2.7 9.8 11.9 4.6 27.9
1M base + 3M NMTsynth 2.5 9.6 11.3 5.3 27.4
1M base + 3M hybrNMTSMT 2.6 9.5 11.0 5.1 27.6
1M base + 3M fullhybrNMTSMT 2.5 9.7 10.8 4.8 27.7
Table 3: Results of automatic error classification into five error categories: morphological error (morph),
word order error (order), omission, addition and mistranslation.
It can be seen that morphological errors are adding authentic data. As for the 3M corpus, the
slightly improved by any additional data, but it is improvement in this error category is similar to the
hard to draw any conclusions. This is not surpris- one by authentic data, but the best option is to use
ing given that our target language, English, is not NMT BT data alone.
particularly morphologically rich. Nevertheless, In total, the clear advantage of using hybrid sys-
for all three corpus sizes, the numbers are smallest tems can be noted for mistranslations, omissions
for the full hybrid system, being comparable to the and word order which is the most interesting cate-
results with adding authentic data. gory. This error category is augmented by adding
As for word order, adding SMT data is not par- BT SMT data or not affected by adding BT NMT
ticularly beneficial since it either increases (1M data, but combining two types of data creates ben-
and 2M) or does not change (3M) this error type. eficial signals in the source text.
NMT systems alone do not help much either, ex-
cept a little bit for the 3M corpus. Hybrid systems Lexical and Syntactic Variety
yield the best results for this error category for all Lexical and syntactic variety is estimated for each
corpus sizes, reaching or even slightly surpassing translation hypothesis as well as for the human ref-
the result with authentic data. erence translation. The motivation for this is the
Furthermore, all BT data are beneficial for re- observation that machine-translated data is gener-
ducing omissions, especially hybrid which can be ally lexically poorer and syntactically simpler than
even better than the authentic data result. human translations or texts written in the original
As for additions, no systematic changes can be language (Vanmassenhove et al., 2019). We want
observed, except an increase for all types of BT to see how different or similar our translation hy-
data. However, it should be noted that this er- potheses are in this sense, and also how they relate
ror category is reported not to be very reliable for to the reference.
comparing different MT outputs (see for example Lexical variety is measured by vocabulary size
(Popović and Burchardt, 2011)). (number of distinct words) in the given text, and
The mostly affected error category is mistrans- syntactic variety by number of distinct POS n-
lations. All types of additional data are reduc- grams where n ranges from 1 to 4. The results
ing this type of errors, especially the hybrid BT are shown in Figure 1.
data for 1M and 2M, even surpassing the effect of First of all, it can be seen that none of the
927
Figure 1: Lexical variety and syntactic variety for all translation hypotheses and for human reference
translations.
translation hypotheses reaches the variety of the hypotheses are getting closer to the reference. We
reference translation (the black line on the top). take word 4-grams instead of single words because
The difference is even more notable for the syn- it is not only important that a word makes sense in
tax, where the differences between translation hy- isolation, but also in a context. Of course, it is still
potheses are smaller and the difference between possible that some of the new instances are valid
them and the reference is larger than for vocabu- despite being different from the given single refer-
lary. ence.
Furthermore, it can be seen that for authen- The results of precision and recall for word/POS
tic data (thin gray line on the bottom and thick 4-grams are are shown in Figure 2. Several ten-
gray line) the variety increases monotonically with dencies can be observed:
adding more text.
Lexical variety is increased by all synthetic • hybrid BT data is especially beneficial for the
data, too, even more than by authentic data, how- 1M and 2M additional corpora, for 1M even
ever, for the NMT and hybrid synthetic data the outperforming the authentic additional data,
increase for the 3M corpus is smaller than for especially regarding word 4-grams;
smaller corpora.
• NMT BT is the best synthetic option for
The increase of syntactic variety is lower both
the 3M additional corpus, however not better
for authentic and for synthetic data than the in-
than adding 3M of authentic data. This ten-
crease of lexical variety. For 1M and 2M cor-
dency is largest for POS 4-gram precision.
pus, syntactic variety is barely increased by SMT
synthetic data whereas NMT and hybrid data are • SMT BT data achieves the lowest scores, es-
adding more new instances. For the 3M cor- pecially for POS 4-grams; this is probably re-
pus, however, all synthetic methods yield similar lated to the fact that it produces less gram-
syntactic variety, larger than the one obtained by matical BT sources, which are then propa-
adding authentic data. gated to the translation hypotheses. The dif-
ferences are largest for the 3M additional cor-
Word/POS 4-gram Precision and Recall
pus, which is probably the reason of dimin-
Whereas the increase of lexical and syntactic va- ished effect of the hybrid BT data for this
rieties is a positive trend in general, there is no setup.
guarantee that the MT systems are not introduc-
ing noise thereby. To estimate how many of added Overall tendencies are that the hybrid BT data
words and POS sequences are sensible, we calcu- is capable even of outperforming the same amount
late precision and recall of word and POS 4-grams of authentic data if the amount of added data does
when compared to the given reference translation. not exceed the double size of the baseline authen-
The idea is to estimate how much the translation tic data. For larger data, a deterioration can be ob-
928
Figure 2: Word/POS 4-gram precision and recall for all translation hypotheses.
served for the SMT BT data, leading to saturation data, mostly by reducing the number of mistrans-
of hybrid models. lations, and increasing the lexical and syntactic va-
Further work dealing with mixing data tech- riety in a positive way (introducing useful new in-
niques is necessary, in order to investigate refined stances).
selection methods (for example, removing SMT However, if the amount of synthetic data be-
segments which introduce noise). comes too large (three times larger than the au-
thentic baseline data), the benefits of hybrid sys-
9 Conclusion and Future Work tem start to diminish. The most probable reason is
the decrease in grammaticality introduced by SMT
In this work we have presented a comparison of BT data which becomes dominant for the larger
the performance of models trained with increasing synthetic corpora.
size of back-translated data. The artificial data sets The presented findings offer several directions
explored include sentences generated by using an for the future work, such as exploring efficient
SMT model, and NMT model and a combination strategies for mixing SMT and NMT data for dif-
of both. Two mixing strategies are explored: ran- ferent authentic/synthetic ratios and investigating
domly selecting one half of the source segments morphologically richer target languages.
from the SMT BT data and the other half from the
NMT BT data, and using all BT source segments Acknowledgements
thus repeating each target segment.
This research has been supported by the ADAPT
Some findings from previous work (Burlot and
Centre for Digital Content Technology which
Yvon, 2018) are confirmed, namely that in terms
is funded under the SFI Research Centres Pro-
of overall automatic evaluation scores, SMT BT
gramme (Grant 13/RC/2106) and is co-funded un-
data reaches slightly worse performance than
der the European Regional Development Fund.
NMT BT data. Our main findings are that mix-
ing SMT and NMT BT data further improves over This work has also received funding
each data used alone, especially if full hybridis- from the European Union’s Horizon 2020 re-
ation is used (using two sources for each target search and innovation programme under the Marie
side). These data can even reach better perfor- Skłodowska-Curie grant agreement No 713567.
mance than adding the same amount of authentic
929
References colocated with ACL. Berlin, Germany, pages 751–
758.
Satanjeev Banerjee and Alon Lavie. 2005. Meteor:
An automatic metric for MT evaluation with im- Guillaume Klein, Yoon Kim, Yuntian Deng, Jean
proved correlation with human judgments. In Pro- Senellart, and Alexander M. Rush. 2017. Open-
ceedings of the ACL workshop on intrinsic and ex- nmt: Open-source toolkit for neural machine trans-
trinsic evaluation measures for machine translation lation. In Proceedings of the 55th Annual Meeting
and/or summarization. Ann Arbor, Michigan, pages of the Association for Computational Linguistics-
65–72. System Demonstrations. Vancouver, Canada, pages
67–72.
Ondřej Bojar, Rajen Chatterjee, Christian Federmann,
Yvette Graham, Barry Haddow, Shujian Huang, Philipp Koehn. 2004. Statistical significance tests for
Matthias Huck, Philipp Koehn, Qun Liu, Varvara machine translation evaluation. In Proceedings of
Logacheva, et al. 2017. Findings of the 2017 confer- the 2004 Conference on Empirical Methods in Nat-
ence on machine translation (wmt17). In Proceed- ural Language Processing. Barcelona, Spain, pages
ings of the Second Conference on Machine Transla- 388–395.
tion. Copenhagen, Denmark, pages 169–214.
Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris
Ondřej Bojar, Rajen Chatterjee, Christian Federmann, Callison-Burch, Marcello Federico, Nicola Bertoldi,
Barry Haddow, Matthias Huck, Chris Hokamp, Brooke Cowan, Wade Shen, Christine Moran,
Philipp Koehn, Varvara Logacheva, Christof Monz, Richard Zens, Chris Dyer, Ondřej Bojar, Alexan-
Matteo Negri, Matt Post, Carolina Scarton, Lucia dra Constantin, and Evan Herbst. 2007. Moses:
Specia, and Marco Turchi. 2015. Findings of the Open source toolkit for SMT. In Proceedings of
2015 Workshop on Statistical Machine Translation. 45th annual meeting of the ACL on interactive poster
In Proceedings of the Tenth Workshop on Statistical & demonstration sessions. Prague, Czech Republic,
Machine Translation. Lisboa, Portugal, pages 1–46. pages 177–180.
Franck Burlot and François Yvon. 2018. Using mono- Matteo Negri, Marco Turchi, Rajen Chatterjee, and
lingual data in neural machine translation: a system- Nicola Bertoldi. 2018. ESCAPE: a large-scale syn-
atic study. In Proceedings of the Third Conference thetic corpus for automatic post-editing. In Pro-
on Machine Translation: Research Papers. Associa- ceedings of the Eleventh International Conference
tion for Computational Linguistics, pages 144–155. on Language Resources and Evaluation, LREC.
http://aclweb.org/anthology/W18-6315. Miyazaki, Japan.
Rajen Chatterjee, Matteo Negri, Raphael Rubino, and Franz Och. 2003. Minimum error rate training in sta-
Marco Turchi. 2018. Findings of the WMT 2018 tistical machine translation. In ACL-2003: 41st An-
shared task on automatic post-editing. In WMT nual Meeting of the Association for Computational
(shared task). Association for Computational Lin- Linguistics, Proceedings. Sapporo, Japan, pages
guistics, pages 710–725. 160–167.
Jonathan H. Clark, Chris Dyer, Alon Lavie, and Kishore Papineni, Salim Roukos, Todd Ward, and Wei-
Noah A. Smith. 2011. Better hypothesis testing for Jing Zhu. 2002. Bleu: a method for automatic eval-
statistical machine translation: Controlling for opti- uation of machine translation. In Proceedings of
mizer instability. In Proceedings of the 49th Annual 40th Annual Meeting of the Association for Com-
Meeting of the Association for Computational Lin- putational Linguistics. Philadelphia, Pennsylvania,
guistics: Human Language Technologies (Volume 2: USA, pages 311–318.
Short Papers). Portland, Oregon, page 176–181.
Alberto Poncelas, Gideon Maillette de Buy Wenniger,
Sergey Edunov, Myle Ott, Michael Auli, and David and Andy Way. 2019. Adaptation of machine trans-
Grangier. 2018. Understanding back-translation lation models with back-translated data using trans-
at scale. In Proceedings of the 2018 Conference ductive data selection methods. In 20th Interna-
on Empirical Methods in Natural Language Pro- tional Conference on Computational Linguistics and
cessing. Association for Computational Linguistics, Intelligent Text Processing. La Rochelle, France.
pages 489–500. http://aclweb.org/anthology/D18-
1045. Alberto Poncelas, Dimitar Shterionov, Andy Way,
Gideon Maillette de Buy Wenniger, and Peyman
Kenneth Heafield. 2011. KenLM: Faster and smaller Passban. 2018. Investigating Back translation in
language model queries. In Proceedings of the Sixth Neural Machine Translation. In Proceedings of the
Workshop on Statistical Machine Translation. Edin- 21st Annual Conference of the European Associ-
burgh, Scotland, pages 187–197. ation for Machine Translation (EAMT 2018). Ali-
cante, Spain.
Marcin Junczys-Dowmunt and Roman Grundkiewicz.
2016. Log-linear combinations of monolingual and Maja Popović. 2011. Hjerson: An Open Source
bilingual neural machine translation models for au- Tool for Automatic Error Classification of Machine
tomatic post-editing. In Proceedings of the First Translation Output. Prague Bulletin of Mathemati-
Conference on Machine Translation, WMT 2016, cal Linguistics 96:59–68.
930
Maja Popović. 2015. chrF: character n-gram F-score
for automatic MT evaluation. In Proceedings of the
Tenth Workshop on Statistical Machine Translation.
Lisbon, Portugal, pages 392–395.
Maja Popović and Aljoscha Burchardt. 2011. From
human to automatic error classification for machine
translation output. In Proceedings of the 15th Inter-
national Conference of the European Association for
Machine Translation (EAMT 2011). Leuven, Bel-
gium.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2016a. Improving neural machine translation mod-
els with monolingual data. In Proceedings of the
54th Annual Meeting of the Association for Com-
putational Linguistics (Volume 1: Long Papers).
Berlin, Germany, pages 86–96.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2016b. Neural machine translation of rare words
with subword units. In Proceedings of the 54th An-
nual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers). Berlin, Ger-
many, volume 1, pages 1715–1725.
931
Unsupervised dialogue intent detection via hierarchical topic model
932
Proceedings of Recent Advances in Natural Language Processing, pages 932–938,
Varna, Bulgaria, Sep 2–4 2019.
introduce a custom evaluation metric which mea- ample, Wikipedia. Further improvement in qual-
sures the quality of hierarchical relations between ity of clustering models with embeddings can be
topics and intent detection. achieved through fine-tuning. Similar to the tf-
The hierarchy structure helps to make a provi- idf approach dimensionality reduction is often em-
sional clustering more interpretative. Namely, we ployed for the clustering problem (Park et al.,
require first level topics to describe the dialogue 2019). Several averaging schemes can be used
subject and the second level topics to describe the to aggregate word embeddings: mean, where all
action user is interested in. We accomplish this words contribute equally to the document, or idf-
by incorporating information about part-of-speech weighted, where rare words have a greater contri-
(PoS) tags into the model. bution than frequent words.
This paper is organized as follows. Section
2.2 Topic modeling
two describes popular approaches to an unsuper-
vised text classification. Section three describes Another approach to text clustering problem is
our reasoning behind our choices of model archi- topic modeling. The topic model simultaneously
tecture. Section four briefly reviews our prepro- computes words and document embeddings and
cessing pipeline and introduces several enhance- perform clusterization. It should be noted that
ments to the existing NLP techniques. We demon- in some cases topic model-based embeddings out-
strate the results of our model in section five. We perform traditional word embeddings, (Potapenko
conclude our work in section six. et al., 2017). The probability of the word w in the
document d is represented by formula below:
2 Text clustering approaches
X X
2.1 Embeddings approaches p(w | d) = p(w | t)p(t | d) = φwt θtd
The simplest way to build a clustering model on t∈T t∈T
a collection of text documents includes two steps. where matrix Φ contains probabilities φwt of word
On the first step, each document is mapped to a w in topic t, matrix Θ contains probabilities θtd of
real-valued vector. On the second step, one of the topic t in document d.
standard clustering algorithms is applied to the re- Probabilistic Latent Semantic Analysis (pLSA)
sulting vectors. (Hofmann, 2000) is the simplest topic model
There are many methods to build an embedding which describes words in documents by a mix-
of a document. The simplest way is the tf-idf ture of hidden topics. The Φ and Θ distribu-
representation. Logistic regression on the tf-idf tions are obtained via maximization of the like-
representation is quite a strong algorithm for the lihood given probabilistic normalization and non-
text classification problem. This algorithm is re- negativity constraints:
spectable baseline even in deep neural networks
research (Park et al., 2019). However, the di- X X
rect use of the tf-idf representation leads to poor L(Φ, Θ) = ndw log p(w|d) → max
Φ,Θ
results in the clustering problem because of the d∈D w∈W
curse of dimensionality. Dimensionality reduction X
methods could be used to improve clustering qual- φwt = 1, φwt ≥ 0
w∈W
ity: PCA or Uniform Manifold Approximation X
and Projection (UMAP, McInnes et al. (2018)). θtd = 1, θtd ≥ 0
Another popular approach makes use of differ- t∈T
ent word embeddings (Esposito et al., 2016). First This optimization problem can be effectively
of all, each word is mapped to a real-valued vector. solved via EM-algorithm or its online modifica-
Then the document representation is derived from tions (Kochedykov et al., 2017).
the embeddings of its words. The most popular Latend Dirichlet Allocation (LDA) (Blei et al.,
embedding models belong to the word2vec family 2003) model is an extension of pLSA with a prior
(Mikolov et al., 2013b): CBOW, Skip-gram and estimation of the Φ and Θ, widely used in topic
their modifications (Mikolov et al. (2013a)). For modelling. However, as a solution for both pLSA
correct representation word2vec models should be and LDA optimization problem is not unique, each
trained on a large collection of documents, for ex- solution may have different characteristics.
933
Additive Regularization of Topic Models p(t|d, w) more stable for w belonging to a same
(ARTM) (Vorontsov and Potapenko, 2015) is local segment. In a way, p(t|d, w) distribution
non-bayesian extension of likelihood optimization could be interpreted as the analogue for context
task, providing robustness of the solution by ap- embeddings in topic modeling world. p(t|d, w)
plying different regularizers. Each regularizer distribution isn’t used directly for topic represen-
is used to pursue different solution characteris- tation, but it is used on the E-step of EM-algorithm
tics. For example, many varieties of LDA can for φwt and θtd recalculation.
be obtained from ARTM model by using certain In order to obtain more control over intent ro-
smoothing regularizer; pLSA model is an ARTM bustness we propose to use a two-level hierarchi-
model without regularizers. Furthermore, docu- cal topic model. The first level is responsible for
ments can contain not only words but also terms of coarse-grained similarity, while the second one
other modalities (e.g. authors, classes, n-grams), could take into account less obvious but important
which allow us to select specific for our task lan- differences.
guage features. In this case, instead of a sin- The hierarchical ARTM model consists of two
gle Φ matrix, we have several Φm matrices for different ARTM models for each level, which are
each modality m. Resulting functional to be opti- linked to each other. The first level of the hierar-
mized is the sum of weighted with αm coefficients chical model can be any ARTM model. The sec-
modalities likelihoods with regularization terms: ond level is built using regularizer from (Chirkova
X and Vorontsov, 2016) which ensures that each
αm L(Φm , Θm ) + R(∪m Φm , Θ) → max
m
Φ,Θ first-level topic is a convex sum of second-level
topics. Various methods could be employed to en-
3 Multilevel clustering
sure that each parent topic is connected to only a
Our goal is to build a topic model with topics cor- handful of relevant children topics: one can use ei-
responding to the user’s intents. We use the fol- ther interlevel sparsing regularizer (Chirkova and
lowing operational definition of intent: two di- Vorontsov, 2016) or remove “bad“ edges accord-
alogues (as represented by user’s utterances) are ing to EmbedSim metric (Belyy, 2018).
said to have the same intent if both users would
be satisfied with the essentially same reaction by 3.1 Distinct hierarchy levels
the call centre operator. This definition, while in-
Building a two-level clustering model is a diffi-
herently problematic, allows us to highlight sev-
cult task due to the inaccuracy of clustering al-
eral important practical problems:
gorithms. Provided that documents in the model
• Simple bag-of-words (BoW) approach isn’t first-level clusters are already similar to each other
sufficient. Compare: “I want my credit card (as they should be), further separation could be
to be blocked. What should I do¿‘ and “My complicated (especially if we attempt to subdivide
credit card is blocked, what should I do¿‘. each cluster by the same algorithm). In practice,
the second-level clusters tend to repeat first-level
• In some cases, the intent of conversation is
clusters at smaller scale instead of demonstrating
not robust to a single word change. “I want to
some meaningful differences. In order to make
make an appointment with cardiologist“ and
our model able to distinguish new dissimilarities
“I want to make an appointment with neurol-
in clusters on the second level, we adjust algorithm
ogist“ are considered to have the same intent
at the second level: in broad strokes, we base the
since they require the user to perform a virtu-
second level of model on different features.
ally identical set of actions. However, “Pay-
In the context of our problem, separation based
ment of state duty for a passport“ and “Pay-
on the functional purpose of the model tokens is
ment of state duty for vehicle“ are vastly dif-
proposed. We divide all words and n-grams into
ferent.
two groups based on the PoS analysis: “thematic”
To account for the BoW problem we add an and “functional”. The “functional” group consists
n-gram modality and ptdw smoothing regularizer of the verb words and n-grams that contain at least
(Skachkov and Vorontsov, 2018) for all tokens. one verb. The “thematic” group consists of the
The ptdw smoothing regularizer respects the se- nouns and adjectives and n-grams that contain at
quential nature of text, making the distributions least one noun and have no verbs. Inspired by
934
multi–level (Tang et al., 2018) and multi–syntactic modification will extract every high-scoring collo-
(Gupta et al., 2018) phrases annotation, among cation at the cost of increased memory usage.
with hierarchical partition, our approach is essen-
tial for client goal and subgoals extraction. 4.2 Named entity recognition
The purpose of the first hierarchy level is to de- There are a lot of references to the speakers’
termine the conversation subject (the entities the names, company/product names, streets, cities in
dialogue is about). Hence, at the first level of the the dialogue collection. It makes sense to take into
hierarchy thematic tokens should have a notice- account some entities in a special way.
ably higher weight than functional tokens. The For the named entity recognition problem
purpose of the second level of hierarchy, by con- (NER) different methods are commonly used:
trast, is to determine client intent concerning par- rule-based, machine-learning-based or neural-
ticular objects (e.g. what action the client is trying networks-based. We used neural network from
to perform). Functional tokens should have higher Arkhipov et al. (2017) pretrained on a PERSONS-
impact over thematic ones. The tokens unrelated 1000 (Vlasova et al. (2014)) for our experi-
to these two groups are used on both levels and ments. We replace all person related tokens by the
serve as a connection between the layers. hPERSONi tag.
935
ity. The usual procedure involves evaluating the 5.2 Baselines
list of most frequent topic words by human ex- As one of the baselines, we use the following
perts. However, this approach suffers from several procedure. First, we convert raw texts into real-
fundamental limitations (Alekseev, 2018). There- valued vectors using pretrained embeddings or tf-
fore we choose to employ a different method. idf scores in a way described in 2.1. Second, we
For each dataset, we collect a set of dialogue cluster this dataset via K-Means algorithm. Third,
pairs to score our model. Following the reasoning we treat each cluster as a separate collection and
outlined in the section 3, we generated a number perform K-Means algorithm again. As a result, we
of (d1 , d2 ) pairs (where di is a dialogue) and asked obtain both first-level and second-level clusters.
three human experts to label them. To measure the Another baseline models are hierarchical topic
quality of the model, we compare these labels to model without any additional regularizers and hi-
the labels predicted by model. erarchical topic model with Φ and Θ smoothing
The following list summarizes our approach for for both levels. For K-Means based algorithms
model estimation and labeling guidelines for hu- we tune embeddings dimensionality and both level
man experts: cluster number. For topic modeling based al-
• 0: d1 and d2 have nothing in common. Such gorithms we tune both level topics number. As
objects should correspond to the different shown in table 1 regularized topic model outper-
first-level topics. forms K-Means approaches at two out of the three
pair sets.
• 1: both d1 and d2 are related to the same
subject, but there are significant differences. 1-big 1-small 2-small
Such dialogues should correspond to the hKmeans (tf-idf) 0.568 0.593 0.649
same first-level topic, but to the different hKmeans (emb.) 0.615 0.638 0.641
second-level topics. hPLSA 0.603 0.675 0.633
hARTM 0.636 0.683 0.631
• 2: d1 shares an intent with d2 . Such dia-
logues should correspond to the same first- Table 1: Baselines accuracy
level and second-level topics.
• ?: it is impossible to determine the intent for 5.3 Proposed model perfomance
at least one of the dialogues. We use several NLP-based techniques described
We select the best model according to the accu- in 4 to improve main model quality. We start with
racy metric on a given labeled pairs. Three sets the hPLSA model. For each problem we test a few
of pairs are used for the estimation (∼ 12K and approaches and choose the best one. We add all
∼ 1.5K for the first dataset, ∼ 1.5K for the sec- main features one by one, e.g. we choose the best
ond dataset). All model hyperparameters are tuned method for extracting n-grams and use it on the
according to the accuracy on a 12K dataset (“1- next step. We conduct all the experiments in the
big“). Two other sets are used to control over- following order:
fitting (“1-small“ and “2-small“). Notably, the 1. including additional n-gram modality, choos-
good performance on 2-small dataset implies that ing between the based and modified n-grams
the model generalizes beyond the initial training extracting methods, tuning modality weights
dataset. and topics number;
The same preprocessing procedures are used for
both datasets. All tokens are lemmatized, stop- 2. adding ptdw smoothing at the first model level
tokens are deleted, simple entities (e-mails, web- for all tokens, tuning regularizer coefficient
sites e.t.c) are replaced by their tags. Operator ut- and topics number;
terances are deleted from the dialogue document 3. replacing person related named entities,
(they are not informative in our datasets; for ex- choosing between the dictionary-based and
ample, there are many cases where operator fails rnn-based methods;
to reply at all). Finally, each document is a con-
catenation of one dialogue user utterances from a 4. typo correction, choosing between the base
single dialogue. and modified algorithm
936
1-big 1-small 2-small Tariff plan
hPLSA 0.603 0.675 0.633 How to change the tariff plan?
+ n-grams base 0.612 0.634 0.633 When did the tariff change happen?
+ n-grams mod. 0.635 0.674 0.655 How often can I change my tariff plan?
+ ptdw smooth. 0.64 0.678 0.66 When will the changes take effect when the
tariff is changed?
+ NER dict. 0.634 0.661 0.635
Why can’t I change the tariff?
+ NER NN 0.64 0.68 0.662
Why was the tariff plan changed without my
+ Jamspell 0.635 0.674 0.655 knowledge?
+ mod. Jamspell 0.657 0.686 0.663 Why there are no available tariff plans for the
Table 2: NLP techniques quality improvement transition?
937
References Denis Kochedykov, Murat Apishev, Lev Golitsyn, and
Konstantin Vorontsov. 2017. Fast and modular reg-
V.A. Bulatov V.G. Vorontsov K.V. Alekseev. 2018. ularized topic modelling. In 2017 21st Conference
Intra-text coherence as a measure of topic models’ of Open Innovations Association (FRUCT). IEEE,
interpretability. In Computational Linguistics and pages 182–193.
Intellectual Technologies: Papers from the Annual
International Conference Dialogue. pages 1–13. Leland McInnes, John Healy, and James Melville.
2018. Umap: Uniform manifold approximation and
Mikhail Y Arkhipov, Mikhail S Burtsev, et al. 2017.
projection for dimension reduction. arXiv preprint
Application of a hybrid bi-lstm-crf model to the
arXiv:1802.03426 .
task of russian named entity recognition. In Con-
ference on Artificial Intelligence and Natural Lan- Tomas Mikolov, Kai Chen, Greg Corrado, and Jef-
guage. Springer, pages 91–103. frey Dean. 2013a. Efficient estimation of word
representations in vector space. arXiv preprint
A.V. Seleznova M.S. Sholokhov A.K. Vorontsov K.V.
arXiv:1301.3781 .
Belyy. 2018. Quality evaluation and improvement
for hierarchical topic modeling. In Computational Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor-
Linguistics and Intellectual Technologies: Papers rado, and Jeff Dean. 2013b. Distributed representa-
from the Annual International Conference Dialogue. tions of words and phrases and their compositional-
pages 110–123. ity. In Advances in neural information processing
David M Blei, Andrew Y Ng, and Michael I Jordan. systems. pages 3111–3119.
2003. Latent dirichlet allocation. Journal of ma- Jinuk Park, Chanhee Park, Jeongwoo Kim, Minsoo
chine Learning research 3(Jan):993–1022. Cho, and Sanghyun Park. 2019. Adc: Advanced
Yun-Nung Chen, William Yang Wang, and Alexander I document clustering using contextualized represen-
Rudnicky. 2015. Learning semantic hierarchy with tations. Expert Systems with Applications .
distributed representations for unsupervised spoken
Anna Potapenko, Artem Popov, and Konstantin
language understanding. In Sixteenth Annual Con-
Vorontsov. 2017. Interpretable probabilistic embed-
ference of the International Speech Communication
dings: bridging the gap between topic models and
Association.
neural networks. In Conference on Artificial Intelli-
Zheqian Chen, Rongqin Yang, Zhou Zhao, Deng gence and Natural Language. Springer, pages 167–
Cai, and Xiaofei He. 2017. Dialogue act recog- 180.
nition via crf-attentive structured network. CoRR
Andriy Shepitsen, Jonathan Gemmell, Bamshad
abs/1711.05568.
Mobasher, and Robin Burke. 2008. Personalized
Nadezhda Chirkova and Konstantin Vorontsov. 2016. recommendation in social tagging systems using hi-
Additive regularization for hierarchical multimodal erarchical clustering. In Proceedings of the 2008
topic modeling. Journal Machine Learning and ACM conference on Recommender systems. ACM,
Data Analysis 2(2):187–200. pages 259–266.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Nikolay Skachkov and Konstantin Vorontsov. 2018.
Kristina Toutanova. 2018. BERT: pre-training of Improving topic models with segmental structure of
deep bidirectional transformers for language under- texts. In Computational Linguistics and Intellectual
standing. CoRR abs/1810.04805. Technologies: Papers from the Annual International
Conference Dialogue. pages 652–661.
Ahmed El-Kishky, Yanglei Song, Chi Wang, Clare R
Voss, and Jiawei Han. 2014. Scalable topical phrase Da Tang, Xiujun Li, Jianfeng Gao, Chong Wang, Li-
mining from text corpora. Proceedings of the VLDB hong Li, and Tony Jebara. 2018. Subgoal discov-
Endowment 8(3):305–316. ery for hierarchical dialogue policy learning. In
EMNLP.
Fabrizio Esposito, Anna Corazza, and Francesco Cu-
tugno. 2016. Topic modelling with word embed- Nataliya Vlasova, Elena Syleymanova, and Igor Trofi-
dings. In Proceedings of the Third Italian Confer- mov. 2014. The russian language collection for the
ence on Computational Linguistics CLiC-it 2016). named-entity recognition task. Language seman-
pages 129–134. tics: models and technologies pages 36–40.
Sonal Gupta, Rushin Shah, Mrinal Mohit, Anuj Ku- Konstantin Vorontsov and Anna Potapenko. 2015. Ad-
mar, and Mike Lewis. 2018. Semantic parsing for ditive regularization of topic models. Machine
task oriented dialog using hierarchical representa- Learning 101(1-3):303–323.
tions. In EMNLP.
Zhao Yan, Nan Duan, Peng Chen, Ming Zhou, Jianshe
Thomas Hofmann. 2000. Learning the similarity of Zhou, and Zhoujun Li. 2017. Building task-oriented
documents: An information-geometric approach to dialogue systems for online shopping. In Thirty-
document retrieval and categorization pages 914– First AAAI Conference on Artificial Intelligence.
920.
938
Graph Embeddings for Frame Identification
939
Proceedings of Recent Advances in Natural Language Processing, pages 939–948,
Varna, Bulgaria, Sep 2–4 2019.
Section 4 focuses on the FrameNet and WordNet Subsequent FrameID models followed, including
graphs, while Sections 5 and 6 explain how we a system that constructed frame embeddings us-
use the FrameNet knowledge base to build graph ing the Word2Vec model (Botschen et al., 2017).
embeddings. Section 7 outlines our experiments, More recent state-of-the-art models use contextu-
and Section 8 gives results and further analysis of alized embeddings of frames in the BERT frame-
the model performance. work (Sikos and Padó, 2019) or joint models with
semantic dependencies and frames (Peng et al.,
2 Background 2018). All of these prior neural architectures con-
struct frame embeddings directly from annotated
2.1 Graph Embeddings
corpora, meaning the frame embeddings that are
Graph algorithms, especially random walk al- used to make predictions in the model are limited
gorithms, have been applied to prediction tasks to frames that are seen in the corpora.
such as word sense disambiguation (Agirre et al.,
2014), measuring semantic similarity and relat- 3 Frame Identification Model
edness between words (Agirre et al., 2010), and
This section describes the overall architecture of
entity linking (Guo et al., 2011). These algo-
our FrameID model. We adopt a bidirectional re-
rithms traverse over the relations and nodes in a
current neural network (Bi-LSTM) that accepts as
large knowledge base such as WordNet (Miller,
input different embeddings that represent frames
1995) or taxonomies built from Wikipedia links
and predicates. This allows us to measure the ef-
(Cucerzan, 2007) to uncover relationships be-
fectiveness of the pre-trained embeddings (corpus-
tween the nodes in the knowledge base. How-
based, graph-based and combined) for an apples-
ever, the effectiveness of shallow neural networks
to-apples comparison within the same setting. The
in learning word similarity, as shown by the pop-
components of our neural network include:
ular Word2Vec and GloVe models (Mikolov et al.,
2013; Pennington et al., 2014), rapidly replaced • an input embedding layer; the network
traditional graph-based methods for word simi- can allow two embedding inputs per word
larity and prediction tasks. Word representations – the embeddings from separate sources
learned in these models, called embeddings, pro- (corpus/graph-based) are concatenated;
vide the latent features of a word in context as • bidirectional recurrent layers; the forward
low-dimensional vectors. Recent work has sought and backward states are concatenated;
to incorporate the best of graph algorithms and • an output vector with the same dimensional-
embeddings by learning representations for words ity as the frame embeddings;
over a neural network while using the structural in- • an objective function that compares the fi-
formation found in knowledge bases (Goikoetxea nal output and the gold frame embeddings;
et al., 2015; Faruqui et al., 2015). at those steps that do not introduce frame-
Our method for constructing graph embeddings evoking predicates the network attempts to
follows the work of Goikoetxea et al. (2015) who reconstruct the embedding for the input word
use random walks to build a synthetic corpus itself, so that “infect” should result in the em-
based on entries in WordNet. In this corpus, a bedding of I NFECTING, but “the” should just
word’s contexts are the other words in the knowl- be reconstructed to the embedding of “the”.
edge base that it is related to. They use the corpus
to generate embeddings for words with CBOW At evaluation time, output vectors are compared
and Skipgram models. via cosine similarity to the frame embeddings,
where the label for the closest frame is selected
2.2 Neural Architectures for Frame as the model’s prediction. Figure 1 shows the ar-
Identification chitecture of the network.
The first work to use embeddings for FrameID
4 FrameNet as a Graph
used the WSABIE algorithm to project frames
and predicate contexts into the same shared space As described above in Section 1, FrameNet has
(Hermann et al., 2014). The authors then ap- a structure which connects frames via different
ply a pairwise loss to minimize the distance be- semantic relations. Relations in the knowledge
tween the frames and their predicate instances. base include Inheritance, an is-a relation akin
940
nected to their abstract counterparts with an is-a
relation (R EVENGE ::AVENGER is-a AVENGER),
which preserves the connectedness of the FE in
FrameNet and also provides some connectivity be-
tween frames that share the same FE.
4.1 Extending the FrameNet graph
We empirically find that, on its own, the graph de-
scribed above does not form a dense enough graph
structure. This is due to the fact that few frames
can be connected through short paths since some
frames contain little to no relations to others and
FrameNet FEs are often very specific to particular
types of scenarios.
To overcome this limitation, we mapped
FrameNet to the WordNet lexical resource. Word-
Net groups synonymous words of the same part-
of-speech category into concept clusters called
Figure 1: Bi-LSTM for FrameID. The diagram synsets and provides lexical relations between
shows a model with one embedding layer, but word senses and the relations between synsets. Its
combining with an additional embedding source is dense semantic network provides a rich represen-
done trivially via concatenation. The embeddings tation of the ecology of lexical meaning, espe-
used for the input and the gold frames can either cially when associated resources are brought into
come from the same or a different vector space. play. This allows for connecting concepts that
The size of the word/frame embeddings is n; k is might not be linked via lexico-semantic relations
the number of hidden layers. but nevertheless are mutually dependent due to
contextual co-occurrence.
to hypernymy in WordNet (e.g., R EVENGE in- Mapping FrameNet predicates with the Pred-
herits from R EWARDS AND PUNISHMENTS), Us- icate Matrix The mapping process is done in
ing, which connects a concrete frame to a related, two steps. First, the verbal predicates in FrameNet
but more abstract frame (e.g. S PEED uses the are connected to their corresponding synsets in
M OTION frame), and Subframe, where the par- WordNet. Since one word can be mapped to
ent frame is a more complex event that subsumes several senses in WordNet, this mapping is far
the child frame(s) (e.g. C RIMINAL PROCESS sub- from straightforward. We map FrameNet pred-
sumes A RREST). See Ruppenhofer et al. (2006) icates and WordNet synsets through an auxil-
for a more detailed description of all possible rela- iary resource called the Predicate Matrix (De La-
tion types. calle et al., 2014), which aimed to automati-
cally extend the predicate mappings from Sem-
This network of relations is used as the basis of
Link (Palmer, 2009). Within the Predicate Matrix,
our graph. We construct a graph where each frame
predicates from a FrameNet frame are linked to
is a node, and we build edges between frames ac-
concept nodes, which are also connected to a cor-
cording to the available relational information in
responding WordNet synset. Via these synsets, we
FrameNet. Frame elements (FEs) are also nodes
now have access to WordNet’s semantic network,
that are connected to their frames. Some FE nodes,
where words are interconnected through lexico-
such as AGENT, are widely connected in the graph
semantic relations such as hypernymy, antonymy,
since they are linked to a diverse set of frames. In
meronymy, as well as through other relations ex-
order to have a single FE node be distinctive to a
pressing relatedness.
single frame, the FE nodes are given unique iden-
tifi