1 natural language processing chapter 18. 2 outline reference –kinds of reference phenomena...
TRANSCRIPT
1
Natural Language Processing
Chapter 18
2
Outline
• Reference– Kinds of reference phenomena– Constraints on co-reference– Preferences for co-reference– The Lappin-Leass algorithm for
coreference
• Coherence– Hobbs coherence relations– Rhetorical Structure Theory
3
Part I: Reference Resolution
• John went to Bill’s car dealership to check out an Acura Integra. He looked at it for half an hour
• I’d like to get from Boston to San Francisco, on either December 5th or December 6th. It’s ok if it stops in another city along they way
4
Some terminology
• John went to Bill’s car dealership to check out an Acura Integra. He looked at it for half an hour
• Reference: process by which speakers use words John and he to denote a particular person– Referring expression: John, he– Referent: the actual entity (but as a shorthand we
might call “John” the referent).– John and he “corefer”– Antecedent: John– Anaphor: he
5
Discourse Model
• Model of the entities the discourse is about
• A referent is first evoked into the model. Then later it is accessed from the model
John HeCorefer
Evoke Access
6
Many types of reference
• (after Webber 91)• According to John, Bob bought Sue an
Integra, and Sue bought Fred a Legend– But that turned out to be a lie (a speech act)– But that was false (proposition)– That struck me as a funny way to describe
the situation (manner of description)– That caused Sue to become rather poor
(event)– That caused them both to become rather
poor (combination of several events)
7
Reference Phenomena
• Indefinite noun phrases: generally new– I saw an Acura Integra today– Some Acura Integras were being unloaded…– I am going to the dealership to buy an Acura Integra
today. (specific/non-specific)• I hope they still have it• I hope they have a car I like
• Definite noun phrases: identifiable to hearer because– Mentioned: I saw an Acura Integra today. The
Integra was white– Identifiable from beliefs: The Indianapolis 500– Inherently unique: The fastest car in …
8
Reference Phenomena: Pronouns
• I saw an Acura Integra today. It was white• Compared to definite noun phrases,
pronouns require more referent salience.– John went to Bob’s party, and parked next to a
beautiful Acura Integra– He got out and talked to Bob, the owner, for more
than an hour.– Bob told him that he recently got engaged and
that they are moving into a new home on Main Street.
– ??He also said that he bought it yesterday.– He also said that he bought the Acura yesterday
9
Salience Via Structural Recency
• E: So, you have the engine assembly finished. Now attach the rope. By the way, did you buy the gas can today?
• A: Yes• E: Did it cost much?• A: No• E: Good. Ok, have you got it attached
yet?
10
More on Pronouns
• Cataphora: pronoun appears before referent:– Before he bought it, John checked over
the Integra very carefully.
11
Inferrables
• I almost bought an Acura Integra today, but the engine seemed noisy.
• Mix the flour, butter, and water.– Kneed the dough until smooth and shiny– Spread the paste over the blueberries– Stir the batter until all lumps are gone.
12
Discontinuous sets
• John has an Acura and Mary has a Suburu. They drive them all the time.
13
Generics
• I saw no less than 6 Acura Integras today. They are the coolest cars.
14
Pronominal Reference Resolution
• Given a pronoun, find the referent (either in text or as a entity in the world)
• We will approach this today in 3 steps– Hard constraints on reference– Soft constraints on reference– An algorithm that uses these constraints
15
Why people care
• Classic: "text understanding"• Information extraction, information
retrieval, summarization…
16
What influences pronoun resolution?
• Syntax• Semantics/world knowledge
17
Why syntax matters
• John kicked Bill. Mary told him to go home.
• Bill was kicked by John. Mary told him to go home.
• John kicked Bill. Mary punched him.
18
Why syntax matters
• John kicked Bill. Mary told him to go home.
• Bill was kicked by John. Mary told him to go home.
• John kicked Bill. Mary punched him.
John
19
Why syntax matters
• John kicked Bill. Mary told him to go home.
• Bill was kicked by John. Mary told him to go home.
• John kicked Bill. Mary punched him.Bill
20
Why syntax matters
• John kicked Bill. Mary told him to go home.
• Bill was kicked by John. Mary told him to go home.
• John kicked Bill. Mary punched him.
Bill
21
Why syntax matters
• John kicked Bill. Mary told him to go home.
• Bill was kicked by John. Mary told him to go home.
• John kicked Bill. Mary punched him.
Grammatical role hierarchy
22
Why syntax matters
• John kicked Bill. Mary told him to go home.
• Bill was kicked by John. Mary told him to go home.
• John kicked Bill. Mary punched him.
Grammatical role
parallelism
23
Why semantics matters
The city council denied the demonstrators a permit because they {feared|advocated} violence.
24
Why semantics matters
The city council denied the demonstrators a permit because they {feared|advocated} violence.
25
Why semantics matters
The city council denied the demonstrators a permit because they {feared|advocated} violence.
26
Why knowledge matters
• John hit Bill. He was severely injured.
27
Margaret Thatcher admires Hillary Clinton, and George
W. Bush absolutely worships her.
Why Knowledge Matters
28
Hard constraints on coreference
• Number agreement– John has an Acura. It is red.
• Person and case agreement– *John and Mary have Acuras. We love them
(where We=John and Mary)
• Gender agreement– John has an Acura. He/it/she is attractive.
• Syntactic constraints– John bought himself a new Acura (himself=John)– John bought him a new Acura (him = not John)
29
Pronoun Interpretation Preferences
• Selectional Restrictions– John parked his Acura in the garage. He
had driven it around for hours.
• Recency– John has an Integra. Bill has a Legend.
Mary likes to drive it.
30
Pronoun Interpretation Preferences
• Grammatical Role: Subject preference– John went to the Acura dealership with
Bill. He bought an Integra.– Bill went to the Acura dealership with
John. He bought an Integra– (?) John and Bill went to the Acura
dealership. He bought an Integra
31
Repeated Mention preference
• John needed a car to get to his new job. He decided that he wanted something sporty. Bill went to the Acura dealership with him. He bought an Integra.
32
Parallelism Preference
• Mary went with Sue to the Acura dealership. Sally went with her to the Mazda dealership.
• Mary went with Sue to the Acura dealership. Sally told her not to buy anything.
33
Verb Semantics Preferences
• John telephoned Bill. He lost the pamphlet on Acuras.
• John criticized Bill. He lost the pamphlet on Acuras.
• Implicit causality– Implicit cause of criticizing is object.– Implicit cause of telephoning is subject.
34
Verb Preferences
• John seized the Acura pamphlet from Bill. He loves reading about cars.
• John passed the Acura pamphlet to Bill. He loves reading about cars.
35
Pronoun Resolution Algorithm
• Lappin and Leass (1994): Given he/she/it, assign antecedent.
• Implements only recency and syntactic preferences
• Two steps– Discourse model update
• When a new noun phrase is encountered, add a representation to discourse model with a salience value
• Modify saliences.
– Pronoun resolution• Choose the most salient antecedent
36
Salience Factors and Weights
• From Lappin and Leass
Sentence recency 100
Subject emphasis 80
Existential emphasis 70
Accusative (direct object) emphasis
50
Ind. Obj and oblique emphasis 40
Non-adverbial emphasis 50
Head noun emphasis 80
37
Recency
• Weights are cut in half after each sentence is processed
• This, and a sentence recency weight (100 for new sentences, cut in half each time), captures the recency preferences
38
Lappin and Leass (cont)
• Grammatical role preference– Subject > existential predicate nominal > object >
indirect object > demarcated adverbial PP
• Examples– An Acura Integra is parked in the lot (subject)– There is an Acura Integra parked in the lot (ex. pred
nominal)– John parked an Acura Integra in the lot (object)– John gave his Acura Integra a bath (indirect obj)– In his Acura Integra, John showed Susan his new CD
player (demarcated adverbial PP)
• Head noun emphasis factor gives above 80 points, but followed embedded NP nothing:– The owner’s manual for an Acura Integra is on John’s
desk
39
Lappin and Leass Algorithm
• Collect the potential referents (up to 4 sentences back)
• Remove potential referents that do not agree in number or gender with the pronoun
• Remove potential references that do not pass syntactic coreference constraints
• Compute total salience value of referent from all factors, including, if applicable, role parallelism (+35) or cataphora (-175).
• Select referent with highest salience value. In case of tie, select closest.
40
Example
• John saw a beautiful Acura Integra at the dealership. He showed it to Bob. He bought it.
rec Subj Exist
Obj Ind-obj
Non-adv
Head N
Total
John 100 80 50 80 310
Integra 100 50 50 80 280
dealership
100 50 80 230
Sentence 1:
41
After sentence 1
• Cut all values in half
Referent Phrases Value
John {John} 155
Integra {a beautiful Acura Integra}
140
dealership
{the dealership 115
42
He showed it to Bob
• He specifies male gender• So Step 2 reduces set of referents to only
John.• Now update discourse model:• He in current sentence (recency=100),
subject position (=80), not adverbial (=50) not embedded (=80), so add 310:
Referent Phrases Value
John {John, he1} 155+310
Integra {a beautiful Acura Integra}
140
dealership
{the dealership 115
43
He showed it to Bob
• Can be Integra or dealership.• Need to add weights:
– Parallelism: it + Integra are objects (dealership is not), so +35 for integra
– Integra 175 to dealership 115, so pick Integra
• Update discourse model: it is nonembedded object, gets 100+50+50+80=280:
44
He showed it to Bob
Referent Phrases Value
John {John, he1} 465
Integra {a beautiful Acura Integra, it1}
420
dealership
{the dealership} 115
45
He showed it to Bob
• Bob is new referent, is oblique argument, weight is 100+40+50+80=270
Referent Phrases Value
John {John, he1} 465
Integra {a beautiful Acura Integra, it1}
420
Bob {Bob} 270
dealership
{the dealership} 115
46
He bought it
• Drop weights in half:
Referent Phrases Value
John {John, he1} 232.5
Integra {a beautiful Acura Integra, it1}
210
Bob {Bob} 135
dealership
{the dealership} 57.5He2 will be resolved to John, and it2 to Integra
47
A search-based solution
• Hobbs 1978: Resolving pronoun references
• We are skipping this topic Fall 2007
48
Hobbs 1978
• Assessment of difficulty of problem• Incidence of the phenomenon• A simple algorithm that has become a
baseline• See handout
• Skipping Fall 2007
49
A parse tree
•Skipping Fall 2007
50
Hobbs’s point
…the naïve approach is quite good. Computationally speaking, it will be a long time before a semantically based algorithm is sophisticated enough to perform as well, and these results set a very high standard for any other approach to aim for.
51
Hobbs’s point
Yet there is every reason to pursue a semantically based approach. The naïve algorithm does not work. Any one can think of examples where it fails. In these cases it not only fails; it gives no indication that it has failed and offers no help in finding the real antecedent.
(p. 345)
52
Reference Resolution: Summary
• Lots of other algorithms and other constraints– Centering theory: constraints which focus on
discourse state, and focus. (read on your own)
– Hobbs: ref. resolution as by-product of general reasoning (later in these notes)
– Mitkov et al. (e.g.) Machine learning
53
Part II: Text Coherence
54
What Makes a Discourse Coherent?
The reason is that these utterances, when juxtaposed, will not exhibit coherence. Almost certainly not. Do you have a discourse? Assume that you have collected an arbitrary set of well-formed and independently interpretable utterances, for instance, by randomly selecting one sentence from each of the previous chapters of this book.
55
Better?
Assume that you have collected an arbitrary set of well-formed and independently interpretable utterances, for instance, by randomly selecting one sentence from each of the previous chapters of this book. Do you have a discourse? Almost certainly not. The reason is that these utterances, when juxtaposed, will not exhibit coherence.
56
Coherence
• John hid Bill’s car keys. He was drunk• ??John hid Bill’s car keys. He likes
spinach
57
What makes a text coherent?
• Appropriate use of coherence relations between subparts of the discourse -- rhetorical structure
• Appropriate sequencing of subparts of the discourse -- discourse/topic structure
• Appropriate use of referring expressions
58
Hobbs 1979 Coherence Relations
• Result • Infer that the state or event asserted
by S0 causes or could cause the state or event asserted by S1.– John bought an Acura. His father went
ballistic.
59
Hobbs: “Explanation”
• Infer that the state or event asserted by S1 causes or could cause the state or event asserted by S0
• John hid Bill’s car keys. He was drunk
60
Hobbs: “Parallel”
• Infer p(a1,a2...) from the assertion of S0 and p(b1,b2…) from the assertion of S1, where ai and bi are similar, for all I.
• John bought an Acura. Bill leased a BMW.
61
Hobbs “Elaboration”
• Infer the same proposition P from the assertions of S0 and S1:
• John bought an Acura this weekend. He purchased a beautiful new Integra for 20 thousand dollars at Bill’s dealership on Saturday afternoon.
62
An Inference-Based Algorithm
• Abduction A B; B; infer A (unsound)• All Jaguars are fast. John’s car is fast.
Abductively infer: John’s car is a Jaguar.• Defeasible: John’s car is a Porsche,
though.• When we use abduction to recognize
discourse coherence, we want the best explanation.
• Probabilities, heuristics, or both (Hobbs)
63
Example
• See lecture
64
Rhetorical Structure Theory
• One theory of discourse structure, based on identifying relations between segments of the text– Nucleus/satellite notion encodes
asymmetry– Some rhetorical relations:
• Elaboration (set/member, class/instance, whole/part…)
• Contrast: multinuclear• Condition: Sat presents precondition for N• Purpose: Sat presents goal of the activity in N
65
Relations
• A sample definition– Relation: evidence– Constraints on N: H might not believe N as much
as S think s/he should– Constraints on Sat: H already believes or will
believe Sat
• An example:The governor supports big business.He is sure to veto House Bill 1711.
66
Automatic Rhetorical Structure Labeling
• Supervised machine learning– Get a group of annotators to assign a set
of RST relations to a text– Extract a set of surface features from the
text that might signal the presence of the rhetorical relations in that text
– Train a supervised ML system based on the training set
67
Features
• Explicit markers: because, however, therefore, then, etc.
• Tendency of certain syntactic structures to signal certain relations: Infinitives are often used to signal purpose relations: Use rm to delete files.
• Ordering• Tense/aspect• Intonation
68
Some Problems with RST
• How many Rhetorical Relations are there?
• How can we use RST in dialogue as well as monologue?
• RST forces an artificial tree structure on discourses
• Difficult to get annotators to agree on labeling the same texts
69
Summary
• Reference– Kinds of reference phenomena– Constraints on co-reference– Preferences for co-reference– The Lappin-Leass algorithm for
coreference
• Coherence– Hobbs coherence relations– Rhetorical Structure Theory