Anaphora resolution an introduction

Ayush Kulshrestha 201001219, B.tech 2010, Dhirubhai Ambani institute of Information and Communication Technology 201001219@daiict.ac.in (Clark and Marshall 1981); this type of context Abstract dependence is usually called (visual) deixis Interpreting anaphoric expressions is one of hence giving the idea that the engine is that of the most fundamental aspects of language a train. interpretation. The study of anaphora and Here by the linguistic context we have anaphora resolution (also known in specified the domain of interpretation which Computational Linguistics as coreference affects interpretation of nominal resolution) has brought about many By fixing their domain of quantificationthe fundamental developments in theoretical set of objects of the type specified by the linguistics (e.g., the development of dynamic nominal complex which are included in the models of language interpretation) and domain of interpretation (Parted 1995; computational linguistics (e.g., the Cooper 1996). For instance, what makes the developments of theories of local and global use of definite NP the tanker felicitous is the salience) and has important practical fact that the domain of quantification of Applications, e.g., in work on information nominal tanker consists of a single object (in extraction, summarization, and entity the TRAINS dialogues the domain of disambiguation. interpretation coincides with the visual context) .The domain of quantification can Introduction also be specified by the linguistic context. Let us send engine E2 off with a boxcar to Corning to pick up oranges and while its Pronouns are not the only noun phrases there it should pick up the tanker whose interpretation depends on the entities Anaphoric expressions are those who in the context. indicate expressions that depend on linguistic Proper names as David Mitchell also refers to context i.e. Their meaning has to be derived twenty different individuals from context .But there may be more than (Wikipedia) the actual entity depends only on one candidate for a particular anaphora, thus the domain of interpretation, but not on our task is to identify which parts of a text context as they are the natural language refer to the same discourse entity and thus encoding of constants, and therefore the which is the correct antecedent for the object they are referring to is directly anaphor. encoded in their semantics (as opposed to Here in the example it refers to the engine E2 being recovered from the discourse Also the definite NP engine's interpretation situation).. Hence are unique in their domain depends on of interpretation. The visual context, which in the TRAINS Also expressions such as pro-verbs like did in dialogues is a map of the TRAINS world (a) and ellipsis such as gapping in (b). shared between the participants. The tanker a. Kim is making the same mistakes that I did. has not been mentioned before, but its b. Kim brought the wine, and Robin the On this map, and therefore it is shared and cheese need anaphora resolution for did and has high salience and can be referred to the missing verb.

The Interpretation of anaphoric Expressions: Evidence from Corpora and Psycholinguistics

There is often more than one matching antecedent for an anaphora, thus resolving this ambiguity requires a combination of many different types of information. In these cases any constraints and preferences are applied to resolve these conflicts. Here constraints put a condition to be satisfied satisfying which does not guarantee anything. While preference assigns weight to the most promising one The c-command relation at the heart of Binding theory which says An element binds an element if and only if c-commands , and and corefer. Consider the sentence 'John' likes his mother In this sentence 'John' binds "his" so it satisfies the condition .On the other hand, in the sentence The mother of John likes himself John" have no binding relationship with himself so they cant corefer. Following 3 principal apply the binding theory for the anaphora resolution

Morphological constraints
GenderFor ex. him can only refer to father, not to mother But gender not always used consistently in English Maja arrived to the airport. [Maja a man] He ... b. John brought Maja to the airport. [Maja a small dog] It Plural pronoun They are easy to read following a conjoined noun phrase (Bill and Sue met) than when the antecedents were syntactically divided (Bill met Sue)

Syntactic constraints
Government and Binding theory It is an important theory that allows us to implement syntactic constraints on the candidates for antecedents. An element c-commands node if and only if != ,B don't dominate each other, Every X that dominates also dominates . In the following tree, B c-commands C

Principle A: reflexives and reciprocals must have a c-commanding antecedent in their governing category). Principle B states that pronouns cannot have an antecedent in this governing category. Principle C states that proper names and nominalcannot have c-commanding antecedents.

Preferences
Common Sense knowledge Consider the two sentences The city council refused the women a permit because they feared violence. Here they refers to the council

The city council refused the women a permit because they advocated violence. But in this case the refers to the women Syntactic Preferences Corpuses suggest that in English corpora, 6070% of pronouns occur in subject position, and of these, 70% have an antecedent in subject position. It's been called subject assignment. These features help us to quickly assign the antecedent correctly with high probability Salience It says that the more recently introduced entities are more likely antecedents. Hobbs (1978) reported that in his corpus, 90% of all pronoun antecedents were in the current sentence, 98% in the current or the previous sentence, Given (1992) found that 25% of definite antecedents were in the current clause, 60% in the current or previous 20 clauses but 40% were further apart So it appears that we can just refer to the first antecedent as the correct one with high probability. There is a lot of evidence for a first mention advantagea preference to refer to first mentioned entities in a sentence combined; these results provide support for a search strategy like that proposed by Hobbs (1978): going back one sentence at a time, then leftto-right.

It incorporates binding theory, and preference for first mentioned entities. It goes up to the dominating NP and finds an antecedent in the left tree & if not found goes up the tree to an S or NP node, finding along the left of the path and then along the right of the path if the node is S node and continue above till it hits an S node where upon continue the same with previous sentences Steps 2, 3 ensure that no NP within the same binding domain as a pronoun will be chosen as antecedent, in that step 3 requires another NP or S node to occur in between the top node (node X) and any candidate: thus for example [NP John] will not be chosen as a candidate antecedent of pronoun him in the example Hobbs Algorithm 1: Begin at the NP node immediately dominating the pronoun. 2: Go up the tree to the first NP or S node encountered. Call this node X, and call the path used to reach it p. 3: Traverse all branches below node X to the left of path p in a left-to-right, breadth first fashion. Propose as the antecedent any NP node that is encountered which has an NP or S node between it and X. 4: if node X is the highest node in the sentence then 5: traverse the surface parse trees of previous sentences in the text in order of recency, the most recent first; each tree is traversed in a leftto-right, breadth-first manner, and when an NP is encountered, it is proposed as antecedent 6: else 7: (X is not the highest node in the sentence) continue to step 9. 8: end if 9: From node X, go up the tree to the first NP or S node encountered. Call this new node X, and call the path traversed to each it p.

Hobbs Algorithm

The earliest, best-known syntax-based algorithm, often used as a baseline, traverses the surface parse tree breadth-first, left-toright, and then going backwards one sentence at a time

10: if X is an NP node and if the path p to X did not pass through the N node that X immediately dominates then 11: propose X as the antecedent 12: end if 13: Traverse all branches below node X to the left of path p in a left-to-right, breadth first manner. Propose any NP node encountered as the antecedent. 14: if X is an S node then 15: traverse all branches of node X to the right of path p in a left-to-right, breadth first manner, but do not go below any NP or S node encountered. 16: Propose any NP node encountered as the antecedent. 17: end if 18: Go to step 4 Because the search is breadth-first, left-to-right, NPs to the left and higher in the node will be preferred over NPs to the right and more deeply embedded ones

participants and common evaluation procedure for the comparative evaluation was Thought which made it possible to train and test anaphora resolution systems on the

same datasets, and therefore to compare their results

Evaluating Coreference Resolution Systems

Link-based Measures

Evaluation of Hobbs Algorithm Lappin and Leass (1994) observed an accuracy of 82% over 360 pronouns from their corpus of computer manuals for their reimplementation of the algorithm. Tetreault (2001) found an accuracy of 76.8% over the 1694 pronouns in the Geet al corpus of news text from the Penn Treebank, and of 80.1% over 511 pronouns from fictional texts

Developing an Experimental Setting There was a shift in focus towards the creation of the first medium-size annotated corpora. These changes primarily brought about by the Message Understanding Conferences (MUC), a DARPA-funded initiative where researchers would compare the quality of their information extraction systems on an annotated corpus provided by funding agencies. MUC hosted two evaluations of coreference resolution systems, MUC-6, and MUC-7 where annotated corpora provided to the

Because most systems are based on the mention-pair model(whether two noun phrases are of the same discourse entity) of anaphora resolution , the simplest way of evaluation is link-based check whether the mention chosen by the system is in fact the mention in the gold standard Since more conservative strategies can leave potentially ambiguous case unresolved and greedier strategies will try to resolve as many potentially anaphoric noun phrases as possible, it is sensible to replace accuracy with two distinct performance measures: Precision is the ratio of the number of correctly resolved anaphoric links to the total number of links that a system resolves; Recall is the ratio of the number of correctly resolved anaphoric links to the total number of anaphoric links in the annotated gold standard. Precision =#correct/#resolved Recall =#correct/#wanted Jointly we get the harmonic mean of precision and recall, called F-measure (F1) F1 =2/(1/Precision + 1/Recall)=2 * #correct/#resolved + #wanted Another measure was adopted for the MUC coreference task definition starting with MUC-6 which uses the concepts of set based measure. requiring a system to reproduce the very same links annotated in the gold-standard (and scoring links to other antecedents

which are correct but not the closest as wrong) results inspurious errors as this would still assign different scores to link structures that result in the same equivalence classes, as in the following example: [A1 Peter] likes to have milk and cereal for breakfast, but was out of milk. Fortunately, [B1 the milkman] came by and [A2 he] was able to buy some milk. Resolving both B1 (The milkman) and A2 (he) to A1(Peter) would then get a better score than resolving B1(the milkman) to A1 (Peter) and A2 (he) to B1(the milkman), even though they both result in the same partition {{A1, A2 , B1}} and both deviate highly from the gold standard (cf. Figure 3). This can be solved by using the set based measures. Here the gold-standard partition of figure 3 is the set G of equivalence classes G = {{A1, A2}, {B1}} |G| = 2. m(A)= the number of class members(the markables) m(G)={A1 , A2 , B1} = 3 l(S) =m(S) |S| =m (|S|) 1 which would give 3 2 = 1 for the case of G, which is incidentally the number of links in the gold standard S1 S2:= {A1A2 |A1 S1 A2 S2 A1 A2=} Here we define precision and recall asP =l (GS)/l(S) R = l (GS)/l (G)

BART - a Beautiful Anaphora Resolution Toolkit

BART, the Beautiful Anaphora Resolution Toolkit, is a product of the project Exploiting Lexical and Encyclopedic sources For Entity Disambiguation at the Johns Hopkins Summer Workshop 2007. BART performs automatic coreference resolution, including all necessary preprocessing steps. BART incorporates a variety of machine learning approaches and can use several machine learning toolkits, including WEKA and an included MaxEnt implementation. It uses Berkeley parser and Stanford NER, as well as a model trained on MUC6 .Available as java software at www.sfs.unituebingen.de/~versley/BART/ with installation details at www.sfs.unituebingen.de/~versley/BART/BARTintro.pdf

BART -how does it work?

View long coreference resolution as a binary classication problem. Each classication instance consists of two markables, i.e. an anaphor and potential antecedent. Instances are modelled as feature vectors and are handed over to a binary classier that decides, given the features, whether the anaphor and the candidate are coreferent or not English system is based on a novel model of coreference. The key concept of model is a Semantic Tree a lecard associated with each discourse entity. Semantic trees are used for both computing feature

In the example in Figure 3, we have G = {{A1, A2}, {B1}} and S1 = S2 = {{A1, A2, B1}}. This would give us l (G) = 1 and l(S) = 2 for both systems, and, with GS = {{A1, A2}, {B1}} l (GS) = 1, P = 1 and R = 1/2

values and guiding the resolution process. It typically contains the following elds: Types: the list of types for mentions of a given entity. For example, if an entity contains the mention software from India, the shallow predicate software is added to the types. Attributes: this eld collects the premodiers. For instance, if one of the mentions is the expensive software the shallow attribute expensive is added to the list of attributes. Relations: this eld collects the prepositional postmodiers. If an entity contains the mention software from India, the shallow relation "from (India) is added to the list of relations. For each mention BART creates such a lecard using syntactic information. It starts by creating a Semantic Tree for each mention. It processes the document from left to right, trying to nd an antecedent for each mention (candidate anaphor). When the antecedent is found, it extends its Semantic Tree with the types, attributes and relations of the anaphor, provided they are mutually compatible. Consider, for example, Software from India, the software and software from China. Initially, BART creates the following semantic trees: (type: software) (relation: from (India)), (type: software) and (type: software) (relation: from (China)). When the second mention gets resolved to the rst one, their semantic trees are merged to (type: software) (relation: from (India). Therefore, when we attempt to resolve the third mention, both candidate antecedents are rejected,

as their relation attributes are incompatible with from (China)

Analysis of Bart
Does not recognizes expletives, never resolves pronouns to inanimate nouns

References Wikipedia.com Anaphora Resolution: A Centering Approach -Aravind K. Joshi Introduction to BART BART: A Multilingual Anaphora Resolution System Samuel Broscheit Computational Models of Anaphora Resolution A Survey Massimo Poesio (University of Essex) Simone Paolo Ponzetto (University of Heidelberg) BART: A Modular Toolkit for Coreference Resolution. Proceedings of the 6th International Conference on Language Resources and Evaluation (LREC 2008). Language Technology for Cultural Heritage: Selected Papers from the LaTeCH ANAPHORA RESOLUTION: THE STATE OF THE ART- Ruslan Mitkov