Transcript Reference resolution

Pragmatics I:
Reference resolution
Ling 571
Fei Xia
Week 7: 11/8/05
Outline
• Discourse: a related group of sentences
– Ex: articles, dialogue, ….
• Pragmatics: the study of the relation
between language and context-of-use
– Reference resolution
– Discourse structure
Reference resolution
Reference resolution
•
•
•
•
•
Some terms: referents, referring expression
Discourse model
Types of referring expression
Types of referents
Constraints and preference for reference
resolution
• Some algorithms for reference resolution
Some terms
• Ex: John bought a book yesterday. He
thought it was cheap.
• Referring expression: the expression used
to refer to an entity:
– Ex: John, a book, he, it
• Referent: an entity that is referred to.
Some Terms (cont)
• Co-reference: two or more referring
expressions refer to the same entity: e.g.,
“John” and “he”.
– Antecedents: a referring expression that
licenses the use of others. Ex. John
– Anaphora: reference to an entity that has
been previous introduced. Ex: he
Discourse Model
• A discourse model stores the
representations of entities that have been
referred to in the discourse and the
relationships in which they participate.
• Two operations:
– Evoke: first mention
– Access: subsequence mention
Refer (evoke)
Refer (access)
He
John
Corefer
Five types of referring expressions
•
•
•
•
•
Indefinite NPs: a car
Definite NPs: the car
Pronouns: it
Demonstratives: this, that
One-anaphora: one
Indefinite NPs
• Introduce entities that are new to the hearer
• The entity may or may not be identifiable to the
speaker:
– I saw an Acura today. (Specific reading)
– I am going to the dealership to buy an Acura today.
(specific or non-specific)
• I hope that they still have it. (Specific)
• I hope that they have a car I like. (non-specific)
Definite NPs
• Identifiable to the hearer
– I saw an Acura today. The Acura …
(explicitly mentioned before in the context)
– The Eagles ….
(the hearer’s knowledge about the world)
– The largest company in Seattle announced
… (inherently unique)
Pronouns
• Pronouns refer to something that is identifiable
to the hearer.
• The referent must have a high degree of
salience in the discourse model.
• Pronouns can participate in cataphora, in which
they appear before their referents.
– Ex: Before he bought it, John checked over the
Acura very carefully.
Demonstratives
•
•
Demonstratives refer to something that is
identifiable to the hearer.
They are used alone or as a determiner:
– Ex: I want this. I want this car.
•
“this” indicating closeness, “that”
signaling distance: spatial/temporal
distance.
One-anaphora
• “One”  “One of them”
• It selects a member from a set that is identifiable
to the hearer.
• Ex:
–
–
–
–
–
He had a BMW before, now he got another one.
Is he the one?
You like this one, or that one?
John has two BMWs, but I have only one.
One should not pay more than 20K for a Camry.
Five types of referring expressions
•
•
•
•
•
Indefinite NPs: a car
Definite NPs: the car
Pronouns: it
Demonstratives: this, that
One-anaphora: one
Next question: what do a referring
expression refers to?
Types of referents
• Ex: According to John, Bob bought Sue a
BMW, and Sue bought Bob a Honda.
– But that turned out to be a lie. (speech act)
– But that was false. (proposition)
– That caused Bob to become rather poor.
(event)
– That caused them both to become rather
poor. (combination of events)
Inferrables
• Explicitly evoked in the text: John bought a
car.
• Inferrables: inferrentially related to an
evoked entity.
– Whole-part: I almost bought a BMW today, but
a door had a dent and the engine seemed
noisy.
– The results of action: Mix the flour and water,
kneed the dough until smooth.
–…
Discontinuous sets
• Plural references may refer to entities that
have been evoked separately.
• Ex:
– John has an Acura, and Mary has a Mazda.
They drive them all the time. (pairwise
reading)
Generics
• Generic references: individual  generic
• Ex: I saw six BMWs today. They are the
coolest cars.
Refer (evoke)
Refer (access)
He
John
Corefer
Constraints and preferences for
reference resolution
• Constraints (filters):
– Agreement: number, person, gender
– Syntax: reflexives
– Semantics: selectional restrictions
• Preferences:
– Salience
– Parallelism
– Verb semantics
Agreement
• Number:
–
–
–
–
(1) John bought a BMW.
(2a) It is great.
(2b) They are great.
(2c) ??They are red.
• Person:
– (1) John and I have BMWs.
– (2a) We love them.
– (2b) They love them.
Agreement (cont)
• Gender: she, he, it.
– (1) John looked at the new ship.
– (2) She was beautiful.
– (1’) Mary looked at the new ship.
– (2) She was beautiful.
Syntactic constraints
• Reflexives and personal pronouns.
– John bought himself a car.
– John bought him a car.
– John wrapped a blanket around himself.
– John wrapped a blanket around him.
Semantic constraints
• Selectional restrictions
– (1) John parked his car in the garage.
– (2a) He had driven it around for hours.
– (2b) It is very messy, with old bike and car
parts lying around everywhere.
– (1) John parked his Acura in downtown
Beverly Hills.
– (2) It is very messy, with old bikes and car
parts lying around everywhere.
Preferences in pronoun
interpretation
• Saliency:
– Recency
– Grammatical role
– Repeated Mention
• Parallelism
• Verb semantics
Saliency
• Recency:
– John has an Integra. …Bill has a BMW. Mary likes to
drive it.
• Grammatical role:
– John went the dealership with Bill. He bought a car.
• Repeated mention:
– John needed a car. He decided to get a BMW. Bill
went to the dealership with him. He bought one.
Parallelism
• Mary went with Sue to the Acura dealership.
Sally went with her to the Mazda dealership.
Verb semantics
• John telephoned
criticized Bill. He lost the
pamphlet on BMWs.
passed the pamphlet to Bill. He
• John seized
loves reading about cars.
• The car dealer impressed
admired John. He
knows Acuras inside and out.
Thematic roles or world knowledge?
Constraints and preferences for
reference resolution
• Hard-and-fast constraints (filters):
– Agreement: number, person, case, gender
– Syntax: reflexives
– Semantics: selectional restrictions
• Preferences:
– Saliency: recency, thematic roles, repeated
mention
– Parallelism
– Verb semantics: thematic roles or world knowledge
Algorithms for pronoun resolution
• Heuristics approaches:
– Lappin & Leass (1994)
– Hobbs (1978)
– Centering Theory (Grosz, Joshi, Weinstein
1995, and various)
• Machine learning approaches
Lappin & Leass 1994
• A heuristic approach.
• Use agreement and syntactic constraints.
• Represent preferences (saliency, parallelism)
with weights.
• Not using: selectional restrictions, verb
semantics, world knowledge.
Salience factors and weights
• Sentence recency:
100
• Subject:
• Existential position:
80
70
– There is a car ….
• Direct object:
• Indirect object:
50
40
• Non-adv:
50
– Inside his car, John …..
• Head noun of max NP:
– The manual for the car is …
80
The algorithm
• Start with an empty set of referents.
• Process each sentence
– For each referring expression
• Calculate the salience value of the expression.
• If it could be merged with existing referents
then choose the referent with the highest saliency value
else add it as a new referent.
• Update the value of the corresponding referent.
– Cut the values of all the referents by half.
An example
• John saw a beautiful Acura at the dealership.
John
Rec
Subj
100
80
Acura 100
dealer 100
ship
Obj
Nonadv
50
Head Total
noun
80
310
50
50
80
280
50
80
230
Before moving on to
the 2nd sentence
Referent
Value
John
Referring
expressions
{John}
Acura
{Acura}
140
dealership {dealership}
155
115
Handling “He”
• He showed it to Bob.
• The value of “He” is 310
Referent
Value
John
Referring
expressions
{John}
Acura
{Acura}
140
dealership {dealership}
155
115
After adding “he”
• He showed it to Bob.
Referent
Value
John
Referring
expressions
{John, he}
Acura
{Acura}
140
dealership {dealership}
465
115
Handling “it”
• He showed it to Bob.
• The salience value of “it” is 280.
• Two new factors:
– Role parallelism:
– Cataphora (??):
35
-175
Referent
John
Expressions
{John, he}
Value
465
Acura
{Acura}
140
dealership {dealership}
115
After adding “it”
• He showed it to Bob.
• The salience value of “it” is 280.
• Two new factors:
– Role parallelism:
– Cataphora (??):
35
-175
Referent
John
Expressions
{John, he}
Value
465
Acura
{Acura, it}
140+280
+35=455
115
dealership {dealership}
Handling “Bob”
• He showed it to Bob.
• The salience value of “Bob” is 270.
Referent
John
Expressions
{John, he}
Value
465
Acura
{Acura, it}
455
dealership {dealership}
115
After adding “Bob”
• He showed it to Bob.
• The salience value of “Bob” is 270.
Referent
Expressions
value
John
{John, he}
465
Acura
{Acura, it}
455
Bob
{Bob}
270
dealership
{dealership}
115
Moving on to the 3rd sentence
• He bought it.
Referent
John
Expressions
{John, he}
value
232.5
Acura
{Acura, it}
227.5
Bob
{Bob}
135
dealership
{dealership}
57.5
 He (John) bought it (Acura).
Core of the algorithm
• For each referring expression
– Calculate the saliency value, x.
– Collect all the referents that comply with
agreement and syntactic constraints.
– If the set is not empty, choose the one with
the highest salience value, and increase the
reference value by x.
– If the set is empty, add a new referent to the
discourse model, and set its value to x.
Algorithms for reference resolution
• Heuristics approaches:
– Lappin & Leass (1994)
– Hobbs (1978)
– Centering Theory (Grosz, Joshi, Weinstein
1995, and various)
• Machine learning approaches
Summary of reference resolution
•
•
•
•
•
Some terms: referents, referring expression
Discourse model
Types of referring expression
Types of referents
Constraints and preference for reference
resolution
• Some algorithms for reference resolution