Transcript Generative Lexicon

Generative Lexicon- Idea and Practicality
Debasri Chakrabarti
02408601
Guide: Prof.Milind S. Malshe
Co-Guide: Prof. Pushpak Bhattacharyya
1
Overview
• Introduction
• Polysemy and the Logical Problem of Polysemy
• Generative Lexicon Theory
• Lexicon Building
• Applications and Limitations of GLT
• Conclusion
2
Introduction
•
Lexicon— ideally collection of all words of a language
• Information stored in a lexicon Phonetic information
 pronunciation
 Semantic information
 meaning
 Morphological information
 transitivity and intransitivity (verbs) , count vs. mass (noun)
3
Lexicon (contd…)
Example of “eat” in the Oxford Advanced Learner’s Dictionary
eat /i:t/ v (pt ate /et/; pp eaten /i:tn/):1. sth (up) to food into the mouth,chew and swallow it: he was too ill to eat
Pronunciation
Meaning
Morphological information
Lexical entry
4
Mental Lexicon
• Mental Lexicon: information stored in the mind of a native speaker
• Native speakers store information
 Phonetic information
 pronunciation
 Semantic information
 meaning
 Morphological information
 transitivity vs.intransitivity (verbs), count vs. mass (noun)
• Additional information
 use of a word in a new context, syntactic environment of a word, wordformation rules
5
Example of Mental Lexicon
Example of eat in a native speaker’s mind
• Pronunciation: long /i:/ is used in eat
• Grammatical information: past tense is ate /et/
• Word-formation rules: /-s/ is the third person singular present tense marker as in
he eats
• Meaning: 1. Take in solid food: she ate a banana
2. Take a meal: we did not eat until 10 P.M.
3. Worry or cause anxiety in a persistent way: what’s eating you up.
• Syntactic Information: eat needs an agent to perform the action.
the agent role is obligatory.
6
Lexicon in Computational Linguistics
Lexicon meant for Natural Language Processing (NLP) must have the
following properties:
•
Morphological information

Parts of speech information

Rules should be there to deal with both regular and irregular forms
e.g ate (past tense of eat)
men (plural of man)
•
Semantic information

•
Can handle lexical ambiguity
Syntactic information

Action verbs will always have an agent
7
Polysemy and the Logical Problem of
Polysemy
Polysemy
•
An individual word can have indefinite number of subtle meaning
difference
•
Natural Languages are highly polysemous
•
This creates ambiguity
•
Weinreich distinguishes between two types of ambiguity


Contrastive ambiguity
Complementary polysemy
8
Polysemy and the Logical Problem of
Polysemy (contd…)
Contrastive Ambiguity
•
•
A lexical item carries two distinct unrelated meanings
This is a case of homonymy

words spelled or pronounced in the same way but have different
meanings
Example:


bank a financial institution
bank place beside a body of water.
9
Polysemy and the Logical Problem of
Polysemy (contd…)
Complementary Polysemy
•
Manifestation of the same basic sense
• Denotes a relation among different senses
Example,
John crawled through the window.
 The window is closed.

Sense 1. Apparatus
Sense 2. Physical Object
10
Sense Enumeration Lexicon (SEL)
•
Simplest model of lexical design to capture the logical polysemy.
•
Widely accepted in both computational and theoritical linguistics.
•
Direct approach to handle polysemy is to allow the lexicon to have
multiple listing of words, each annotated with a separate meaning
or lexical sense.
11
Sense Enumeration Lexicon (SEL)
•
Example of Contrastive Senses
bank1
CAT= count-noun
GENUS= financial-institution
bank2
CAT= count-noun
GENUS= shore
12
Sense Enumeration Lexicon (SEL)
•
Example of Complementary Polysemy
Window1
CAT= count-noun
GENUS= apparatus
Window2
CAT= count-noun
GENUS= artifact
13
Sense Enumeration Lexicon (SEL)
•
Possible Modification of Complementary Polysemy in
SEL
window
sense1
CAT= count-noun
GENUS= apparatus
sense2
CAT= count-noun
GENUS= artifact
14
Generative Lexicon Theory(GLT)
•
•
Major Problems for Lexical Semantics

to explain the polymorphic nature of language

to characterize the semanticality of natural language utterances

to capture the creative use of words in novel contexts

to develop a richer, co-compositional semantic representation
Generative Lexicon Theory

developed by James Pustejovsky

crucial aspect of GLT is the representation and treatment of polysemy

it examines the meaning of words to see the range of polysemy
15
Methodology of Generative Lexicon
Theory
Generative lexicon involves the following methodology
•
Argument Structure
 True Arguments
 Default Arguments
 Shadow Arguments
 True Adjuncts
•
Event Structure
•
Qualia Structure
 Formal
 Constitutive
 Telic
 Agentive
16
Argument Structure
•
True Arguments: syntactically realized parameters of the
lexical item
John arrived late
•
Default Arguments: logically present in the expressions
but are not necessarily expressed syntactically.
John built the house out of bricks
•
True Adjuncts:
 modify the logical expression
 part of the situational interpretation
She drove down to New York on Tuesday.
17
Argument Structure (contd…)
•
Shadow Arguments: semantically incorporated in the lexical
item and are expressed by discourse specification and contextual
factors
Mary buttered her toast

hidden argument is the material being spread on the toast

these are not optional arguments but expressible only under specific
conditions

refer to the semantic content that is not necessarily expressed in syntax
Example: Mary buttered her toast with margarine
18
Event Structure
•
•
event type of a lexical item and a phrase
events can be sub-classified into at least three sorts: State, Process
and Transition
Event Structure of build as found in the following expressions
They are building a new house
The house was built by John
build
EVENTSTR=
E1= process
E2= state
19
Qualia Structure
•
gives a relational force for a lexical item
•
composed of four qualia roles




Formal: This qualia role distinguishes a lexical item within a
larger domain.
Constitutive: This is a relation between an object and its constituent
parts.
Telic: This specifies the purpose and function of a lexical item.
Agentive: This indicates the factors involved in the origin of a
lexical item.
20
Qualia Structure (contd…)
Qualia Structure for novel
novel
Qualia
const = narrative
formal = book
telic = reading
agent = writing
21
Lexical Conceptual Paradigm (LCP)
•
The term is used by Pustejovsky and Anick (1988)
•
Refers to the ability of a lexical item to cluster multiple senses
Example,


•
John crawled through the window.
The window is closed.
Resulting LCP

phys-obj.aperture-lcp = [phys-obj]
[aperture]
22
Generative Device
•
Type Coercion

a lexical item or phrase is coerced to a semantic interpretation by a
governing item in the phrase, without changing its syntactic type
Mary wants John to leave
Mary wants to leave
Mary wants the book
•
Function Application with Coercion


different complement type of the verb
different interpretations of the verb that arise for the different
complements
23
Generative Device
•
Selective Binding

•
a lexical item or a phrase operates specifically on the substructure of a phrase,
without changing the overall type in the composition
a good knife: a knife that cuts well
Co-composition

multiple elements within a phrase behave as functors, generating new nonlexicalized senses for the words in composition
John baked the potato
John baked the cake
24
Lexicon Building
• Building of WordNet


lexical database organised in terms of concept
each concept is related to each other in terms of various semantic
relations
• Building of a Universal Word Dictionary


building a lexicon for Universal Networking Language
Universal Networking Language (UNL) is an electronic language
for computers to express and exchange all kinds of information
• Creation of Verb hierarchy Tree

creating a verb knowledge base for the UNL system
25
Building of WordNet
•
Different semantic relations in WordNet





•
Synonymy
Antonymy
Hypernymy and Hyponymy
Meronymy and Holonymy
Entailment and Troponymy
Multiple Hypernymy in Euro WordNet



Disjunctive Hypernym
Conjunctive Hypernym
Nonexclusive Hypernym
26
Building of WordNet
•
Disjunctive Hypernym


these are incompatible types that never apply simultaneously
found among nouns that refer to the participant in an event
but do not restrict for the type of entity participating
threat
- Role- Agent
- Has Hypernym
- Has Hypernym
- Has Hypernym
threaten
person; disjunctive
thing; disjunctive
idea; disjunctive
27
Building of WordNet
•
Conjunctive Hypernym


•
•
these are compatible types that always apply simultaneously
found for verbs in which multiple aspects are combined.
Dutch Example
doodschoppen to kick to death
- Has Hypernym doden (to kill); conjunctive
- Has Hypernym schoppen (to kick); conjunctive
Similar Hindi example
huMkarnaa: Dranao ko ilae jaaor ka Sabd krnaa (to shout to scare somebody)
- Has Hypernym Dranaa
- Has Hypernym icallaanaa
(to scare)
(to shout)
conjunctive
conjunctive
28
Building of WordNet
•
Non-exclusive Hypernym

either both aspects may apply simultaneously or one of both may
apply
knife
- Has Hypernym weapon
- Has Hypernym cutlery
29
Building of a Universal Word
Dictionary
Construction of Universal Word (UW) in Universal Networking
Language (UNL)
• UNL – electronic language for computers to express and exchange all kinds of
information
• UW – character strings representing unique concept
eat (icl>consume) as in he is eating
eat (icl> damage) as in the house was eaten up by the heat
represented by an English word
captures all the meanings conveyed by that word
restrictions are attached to create unique sense
• UNL Knowledge Base (KB)— performs the task of defining all possible
relationships between two UWs.
30
How to create an UW
I. First a category is decided
a. nominal concept (icl> thing) is attached
e.g swallow(icl> thing)
b. verbal concept
(icl>do) concept of an event caused by something or someone
change (icl>do) as in I changed my mind.
(icl>occur) concept of an event that happens of its own accord
change (icl>occur) as in The weather will change.
(icl>be) concept of a state verb
know(icl>be) as in I know you.
31
How to create a UW(contd…)
To handle the ambiguity of a UW
• For a nominal concept, a subordinate category from the uw hierarchy
should be used rather than a thing.
Example: swallow (icl>bird) as in the swallow is singing.
swallow(icl>action) as in he took the drink at [in] one swallow.
swallow(icl>quantity) as in take a swallow of water.
• For a verbal concept possible case relations are attached.
case relations are like obj>thing, obj>person, gol>thing
Example: spring(icl>occur(obj>liquid)): expresses gushing out as in to spring out
spring(icl>do(gol>place)): expresses jumping up as in to spring up
32
Creation of a verb hierarchal tree
Creation of the Verb knowledge base
Following :
1.Beth Levin’s methodology of verb alternation
example, a. Bill sold a car.
b. Bill sold Tom a car.
2. Hypernymy relation of English Wordnet
Hypernym denotes superset of a concept
example,
animal
Hypernym
cat
33
Creation of a verb hierarchal tree
contd…
• Beth Levin gives the syntactic information.
• Hypernymy gives the semantic information.
• The classification is in the following manner:
 "do(agt>thing,obj>thing {,gol>thing,src>thing,icl>do})"
 "argue({icl>do(}agt>thing,obj>thing,ptn>thing{)})"
34
Creation of a verb hierarchal tree
contd…
Format of the entry:
1Tab
"attack({icl>do(}agt>thing,obj>thing{)})"; Most wild animals
won't attack humans unless they are provoked. /Army forces have been
attacking (the town) since dawn with mortar and shell fire. / Napoleon
attacked Russia in 1812 and was defeated and forced to retreat. (to
make an attack on sb/sth)
2Tab
Tab"assault(icl>attack(agt>thing,obj>thing,man>emotionally))"
Nightmares assaulted him regularly.(to attack sb emotionally)
2Tab
Tab"assault(icl>attack(agt>thing,obj>thing,man>physically))"
;He got two year's imprisonment for assaulting a police officer.[Vn](to
attack sb physicaly and violently, esp when this is a crime)
35
Application of GLT
•
•
•
•
•
Formal role is similar with the hypernymy relation
Constitutive role is similar with the meronymy
relation
Telic role is similar with the functional link given
between a Noun and a Verb in the Hindi WordNet
LCP is used in the multi hypernymy process
Event structure is specified by the ontology nodes in
the Hindi WordNet
36
Application of GLT

English Wordnet (1.7.1) gives 63 senses for the verb sense of break
interrupt, break 1-- (end prematurely; break a lucky streak)
break, break off, discontinue, stop 10-- (prevent completion; stop the project; break the
silence)
break, break away 18-- (interrupt a continued activity; She had broken with the traditional
patterns)
break 31-- (stop or interrupt; He broke the engagement; We had to break our plans for a
trip to China)
separate, part, split up, split, break, break up 33-- (discontinue an association or
relation; go different ways; The business partners broke over a tax question; The couple
separated after 25 years of marriage; My friend and I split up)
37
Application of GLT
•
Merging of senses using GLT
Break
EVENTSTR
QUALIA
E: event
FORMAL: interruption
AGENTIVE: break_act
38
Limitations Of GLT
•
Attempts to distinguish between polysemy and
accidental homonymy
Example of bake
 baked a cake (creativity)
 baked a potato (change of state)
•
Pustejovsky’s suggestion
 cake-artifact
 potato-nat obj
Problem: how to deal with artifacts like knife, car?
39
Conclusion
• Generative mechanisms fail to predict polysemy or
generate polysemous sense
• Generative mechanisms along with ontology can be a
powerful device
• This implies the building of a rich ontology
40