Sunteți pe pagina 1din 24

Lexical

Semantics
Lecture 14

So Far
Words are treated as unanalyzed symbols
Franco, Maharani, Restaurant, Closed
without considering their word meaning
(lexical semantics).

What is a Word?
Lemma vs. Wordform
Lemmatization: process of mapping a wordform
to its corresponding lemma uses
morphological parsing.
Not deterministic, need context, e.g. found
(find) or found(found)
Part-of-speech specific, e.g. table (NN) and
table (VB)

Lemma can have many word senses, e.g.


bank1 (financial institution) and bank2
(edge of river) and bank3 (biological
repository)

Word Sense
Lemma can have many word senses, e.g.
bank1 (financial institution) and bank2
(edge of river) and bank3 (repository)
Two word senses that are unrelated (1&2):
homonymy.
Distinct from homophones (wood/would) &
homographs (bass/bass)
Two word senses that are related (1&3):
polysemy.
Metonymy Subtype of polysemy using one
part or aspect of entity to refer to whole entity
or other aspect.
the White House (building administration)

How Many Word Senses?


Dictionaries use finer-grained senses to aid
word learners dont need as many.
Can test if two senses are distinct using
zeugma:
Does this flight serve breakfast?
Does United serve Sacramento?
>> Does United serve breakfast and Sacramento???

How to Define a Word


Sense?
Dictionaries make reference to other things
familiar to humans in defining:

WordNet is large online DB that stores sense


relations, e.g. left and right are similar lemmas but
in opposition
Can perform sophisticated NLP even w/o knowing
left
Create small finite set of semantic primitives and
define word senses relative to primitives.
Difficult and rarely used in NLP systems.

Relations Between Word


Senses

Synonymy
Synonyms if substitutable for each other in
any sentence without changing truth
condition:
couch/sofa
car/automobile
big/large
water/H2O

Same propositional meaning, but not necessarily


same connotation or genre
Applies to word senses, not words
We flew on a big/large plane.
She became a big/large sister.

Big has two senses, only one shared with


large.

Antonymy
Senses are opposite wrt one aspect of
meaning, otherwise very similar:
long/short
fast/slow
cold/hot
dark/light
rise/fall
in/out

Binary opposition
Opposite ends of some scale
Reversives movement in opposite directions
Hard to auto-distinguish from synonyms!

Hyponymy
Hyponym: more specific sense or subclass
Hypernym/Superordinate: more general
sense

Extensional: set B includes all members of set A


Entailment: being an A implies being a B
Transitive
More general than taxonomy: dog hound,
mutt, cur, puppy, poodle, golden retriever

Semantic Fields
Pairwise relations synonymy, hyponymy,
meronymy dont add up to complete
picture of relatedness:
reservation
fare rates

flight travel buy price


meal plane layover

cost

All relate to air travel


Attempts to store background knowledge in
frames, models, or scripts, e.g. FrameNet.

WordNet
Widely used web accessible or
downloadable
Category
Unique Forms
Hierarchically
Noun
117,097
organized
Verb
11,488
Each lemma has a
22,141
gloss (dictionary- Adjective
4,601
style definition) & Adverb
synset (list of
near synonyms like thesaurus) and usage
examples.
No pronunciation representation (unlike
dictionary)

WordNet

nset: set of near synonyms that share a gloss are the mean

Noun Relations

Noun Relations
Hypernym is immediate
superset of synset
Can follow chain to ever
more general sets.
Stops at entity which is
root for all nouns.

Verb Relations

Web Interface

Event Semantics

Thematic Roles
Neo-Davidsonian

form of events separates


each participant/argument of event:

Breaker and opener are both volitional


actors with causal responsibility for event.
Thematic Roles attempt to capture semantic
commonality.

Thematic Roles

Allows for generalization in QA/IE:


Query: Was company B acquired?
In Doc: Company A acquired Company B.

Diathesis Alternations
From different uses of verb:

Infer grammar of roles for a verb:


AGENT: Subject, THEME: Object
AGENT: Subject, THEME: Object, INSTRUMENT: PPwith
THEME: Subject, AGENT: Object

Challenges
Hard to choose set of roles:
Often need to fragment roles, e.g. at least 2
types of instruments:
Intermediary can be subjects
The cook opened the jar with the new gadget.
The new gadget opened the jar.

Enabling cannot be subjects


She ate the cake with a fork.
The fork ate the cake.

Also hard to define them:


Most AGENTS are animate, volitional, sentient &
causal
Make specific to lexeme PropBank and

FrameNet
The price of bananas increased 5%.
The price of bananas rose 5%.
There has been a 5% rise in the price of bananas.

Although different semantic roles and


lexemes, want to recognize commonalities.
FrameNet is semantic-role labelling project.
Defines frames script-like structure with
frame-specific roles
e.g. Change_position_on_a_scale frame:

FrameNet

S-ar putea să vă placă și