Computing Linguistically-based Computing Linguistically-based Textual InferencesTextual Inferences

Martin Forst

Palo Alto Research CenterJoint work with D. Bobrow, C. Condoravdi, L. Karttunen,

T. H. King, V. de Paiva, A. Zaenen

LORIA, Nancy

March 20, 2008

OverviewOverview

IntroductionMotivationLocal Textual Inference

PARC’s XLE systemProcess pipelineAbstract Knowledge Representation (AKR)

Conceptual and temporal structureContextual structure and instantiability

Semantic relationsEntailments and presuppositionsRelative polarity

Entailment and Contradiction (ECD)

Demo!

MotivationMotivation

A measure of understanding a text is the ability to make inferences based on the information conveyed by it. We can test understanding by asking questions about the text.

A long-standing goal of computational linguistics is to build a system for answering natural language questions.If the question is Did Shackleton reach the South Pole?, the sentence

Shackleton failed to get to the South Pole. contains the answer.

A successful QA system has to recognize semantic relations between sentences.

None of the current search engines (Google, Yahoo!) is capable of delivering a simple NO answer in such cases.

The system I describe in this talk makes the correct inference.

Local Textual Inference Local Textual Inference

PASCAL RTE Challenge (Ido Dagan, Oren Glickman) 2005, 2006

PREMISE

CONCLUSIONTRUE/FALSE

Rome is in Lazio province and Naples is in Campania.

Rome is located in Lazio province.

TRUE ( = entailed by the premise)

Romano Prodi will meet the US President George Bush in his capacity as the president of the European commission.

George Bush is the president of the European commission.

FALSE (= not entailed by the premise)

PARC ECD (Entailment and PARC ECD (Entailment and Contradiction Detection)Contradiction Detection)

Text: Kim hopped.

Hypothesis: Someone moved.

Answer: YES.

Text: Sandy touched Kim.

Hypothesis: Sandy kissed Kim.

Answer: UNKNOWN.

Text: Sandy kissed Kim.

Hypothesis: No one touched Kim.

Answer: NO.

World Knowledge World Knowledge

Romano Prodi will meet the US President George Bush in his capacity as the president of the European commission.

George Bush is the president of the European commission.

FALSE (= not entailed by the premise on the correct anaphoric resolution)

G. Karas will meet F. Rakas in his capacity as the president of the European commission.

F. Rakas is the president of the European commission.

TRUE (= entailed by the premise on one anaphoric resolution)

OverviewOverview

IntroductionMotivationLocal Textual Inference

PARC’s XLE systemProcess pipelineAbstract Knowledge Representation (AKR)

Conceptual and temporal structureContextual structure and instantiability

Semantic relationsEntailments and presuppositionsRelative polarity

Entailment and Contradiction (ECD)

Demo!

XLE System ArchitectureXLE System ArchitectureText Text (A)KR (A)KR

1. Parse text to LFG c-/f-structure pairsc-structures are context-free trees; f-structures are AVMs

Represent syntactic/semantic features (e.g. tense, number)

Localize arguments (e.g. long-distance dependencies, control)

2. Rewrite f-structures to AKR clauses

Collapse syntactic alternations (e.g. active-passive)

Flatten embedded linguistic structure to clausal form

Map to concepts and roles in some ontology

Represent intensionality, scope, temporal relations

Capture commitments of existence/occurrence

3. Rewrite AKR to target KR

XLE PipelineXLE Pipeline

Process OutputText-Breaking Delimited Sentences

NE recognition Type-marked Entities (names, dates, etc.)

Morphological Analysis Word stems + features

LFG parsing Functional Representation

Semantic Processing Scope, Predicate-argument structure

AKR Rules Abstract Knowledge Representation

Alignment Aligned T-H Concepts and Contexts

Entailment and Contradiction Detection

YES / NO / UNKNOWN

XLE PipelineXLE Pipeline

• Mostly symbolic system• Ambiguity-enabled through packed representation of analyses• Filtering of dispreferred/improbable analyses is possible

• OT marks• mostly on c-/f-structure pairs, but also on c-structures• on semantic representations for selectional preferences

• Statistical models• PCFG-based pruning of the chart of possible c-structures• Log-linear model that selects n-best c-/f-structure pairs

morphological analyses

c-structures

c-/f-structure pairs

CSTRUCTURE OT marks PCFG-based chart pruning

“general” OT markslog-linear model

F-structures F-structures vs.vs. AKR AKR

Nested structure of f-structures vs. flat AKRF-structures make syntactically, rather than conceptually, motivated

distinctions Syntactic distinctions canonicalized away in AKR

Verbal predications and the corresponding nominalizations or deverbal adjectives with no essential meaning differences

Arguments and adjuncts map to roles

Distinctions of semantic importance are not encoded in f-structures Word sensesSentential modifiers can be scope taking (negation, modals,

allegedly, predictably)Tense vs. temporal reference

Nonfinite clauses have no tense but they do have temporal reference

Tense in embedded clauses can be past but temporal reference is to the future

F-Structure to AKR MappingF-Structure to AKR Mapping

Input: F-structures

Output: clausal, abstract KR

Mechanism: packed term rewritingRewriting system controls lookup of external ontologies via Unified Lexicon compositionally-driven transformation to AKR

Transformations:Map words to Wordnet synsetsCanonicalize semantically equivalent but formally distinct

representationsMake conceptual & intensional structure explicitRepresent semantic contribution of particular constructions

Basic structure of AKRBasic structure of AKR

Conceptual StructurePredicate-argument structures

Sense disambiguation

Associating roles to arguments and modifiers

Contextual StructureClausal complements

Negation

Sentential modifiers

Temporal StructureRepresentation of temporal expressions

Tense, aspect, temporal modifiers

Conceptual StructureConceptual Structure

Captures basic predicate-argument structures

Maps words to WordNet synsets

Assigns VerbNet roles

subconcept(talk:4,[talk-1,talk-2,speak-3,spill-5,spill_the_beans-1,lecture-1])role(Actor,talk:4,Ed:1)subconcept(Ed:1,[male-2])alias(Ed:1,[Ed])role(cardinality_restriction,Ed:1,sg)

Shared by “Ed talked”, “Ed did not talk” and “Bill will say that Ed talked.”

Canonicalization in conceptual structureCanonicalization in conceptual structure

subconcept(tour:13,[tour-1])role(Theme,tour:13,John:1) role(Location,tour:13,Europe:21) subconcept(Europe:21,[location-1])alias(Europe:21,[Europe]) role(cardinality_restriction,Europe:21,sg)subconcept(John:1,[male-2])alias(John:1,[John])role(cardinality_restriction,John:1,sg)

subconcept(travel:6,[travel-1,travel-2,travel-3,travel-4,travel-5,travel-6])

role(Theme,travel:6,John:1)

role(Location,travel:6,Europe:22)

subconcept(Europe:22,[location-1])

alias(Europe:22,[Europe])

role(cardinality_restriction,Europe:22,sg)

subconcept(John:1,[male-2])

alias(John:1,[John])

role(cardinality_restriction,John:1,sg)

“John took a tour of Europe.”“John traveled around Europe.”

Contextual StructureContextual Structure

context(t)

context(ctx(talk:29))

context(ctx(want:19))

top_context(t)

context_relation(t,ctx(want:19),crel(Topic,say:6))

context_relation(ctx(want:19),ctx(talk:29),crel(Theme,want:19))

Bill said that Ed wanted to talk.

Use of contexts enables flat representations

Contexts as arguments of embedding predicates Contexts as scope markers

Concepts and ContextsConcepts and Contexts

Concepts live outside of contexts.

Still we want to tie the information about concepts to the contexts they relate to.

Existential commitmentsDid something happen?

e.g. Did Ed talk? Did Ed talk according to Bill?

Does something exist?e.g. There is a cat in the yard. There is no cat in the yard.

InstantiabilityInstantiability

An instantiability assertion of a concept-denoting term in a context implies the existence of an instance of that concept in that context.

An uninstantiability assertion of a concept-denoting term in a context implies there is no instance of that concept in that context.

If the denoted concept is of type event, then existence/nonexistence corresponds to truth or falsity.

NegationNegation

Contextual structurecontext(t)context(ctx(talk:12)) new context triggered by negationcontext_relation(t, ctx(talk:12), not:8)antiveridical(t,ctx(talk:12)) interpretation of negation

Local and lifted instantiability assertions instantiable(talk:12, ctx(talk:12)) uninstantiable (talk:12, t) entailment of negation

“Ed did not talk”

Relations between contextsRelations between contexts

Generalized entailment: veridicalIf c2 is veridical with respect to c1,

the information in c2 is part of the information in c1Lifting rule: instantiable(Sk, c2) => instantiable(Sk, c1)

Inconsistency: antiveridicalIf c2 is antiveridical with respect to c1,

the information in c2 is incompatible with the info in c1Lifting rule: instantiable(Sk, c2) => uninstantiable(Sk, c1)

Consistency: averidicalIf c2 is averidical with respect to c1,

the info in c2 is compatible with the information in c1No lifting rule between contexts

Determinants of context relationsDeterminants of context relations

Relation depends on complex interaction ofConceptsLexical entailment classSyntactic environment

ExampleHe didn’t remember to close the window. He doesn’t remember that he closed the window. He doesn’t remember whether he closed the window.

He closed the window.Contradicted by 1Implied by 2Consistent with 3

OverviewOverview

IntroductionMotivationLocal Textual Inference

PARC’s XLE systemProcess pipelineAbstract Knowledge Representation (AKR)

Conceptual and temporal structureContextual structure and instantiability

Semantic relationsEntailments and presuppositionsRelative polarity

Entailment and Contradiction (ECD)

Demo!

Embedded clausesEmbedded clauses

The problem is to infer whether an embedded event is instantiable or uninstantiable on the top level.

It is surprising that there are no WMDs in Iraq.

It has been shown that there are no WMDs in Iraq.

==> There are no WMDs in Iraq.

FactivesFactives

Class Inference Pattern

Positive

Negative

++/-+ forget that forget that X ⇝ X, not forget that X ⇝ X

+-/-- pretend that pretend that X not ⇝ X, not pretend that X not ⇝ X

ImplicativesImplicatives

++/-- manage to

+-/-+ fail to

manage to X ⇝ X, not manage to X not ⇝ X

fail to X not ⇝ X, not fail to X ⇝ X

++ force to force X to Y ⇝ Y

+- prevent from prevent X from Ying not Y⇝

-- be able to not be able to X not X⇝

-+ hesitate to not hesitate to X X⇝

Class Inference Pattern

Two-wayimplicatives

One-wayimplicatives

Implicatives under FactivesImplicatives under Factives

It is surprising that Bush dared to lie.

It is not surprising that Bush dared to lie.

Bush lied.

Phrasal ImplicativesPhrasal Implicatives

Have

Take

Ability NounChance Noun

Character Noun

= --Implicative= --Implicative= ++/--Implicative

Miss Chance Noun = +-/-+Implicative

Seize Chance Noun = ++/--Implicative

Chance Noun

Effort NounAsset Noun

= ++/--Implicative= ++/--Implicative

= ++/--Implicative

Use Chance NounAsset Noun

= ++/--Implicative= ++/--Implicative

WasteChance Noun

Asset Noun

= +-/-+Implicative

= ++/--Implicative

+

+

+

+

+

+

(ability/means)

(chance/opportunity)(courage/nerve)

(chance/opportunity)(money)(trouble/initiative)

(chance/opportunity)(money)

(chance/opportunity)(money)

(chance/opportunity)

(chance/opportunity)

Phrasal Implicatives - ExamplePhrasal Implicatives - Example

Joe had the chutzpah to steal the money. ⇝ Joe stole the money.

Two-way implicativewith “character nouns”

“character noun”(gall, gumption, audacity…)

Relative PolarityRelative Polarity

Veridicality relations between contexts determined on the basis of a recursive calculation of the relative polarity of a given “embedded” context

Globality: The polarity of any context depends on the sequence of potential polarity switches stretching back to the top context

Top-down: Each complement-taking verb or other clausal modifier, based on its parent context's polarity, either switches, preserves or simply sets the polarity for its embedded context

Example: polarity propagationExample: polarity propagation

“Ed did not forget to force Dave to leave.”

“Dave left.”

Ed

subj

obj

subj comp

comp

comp

subj

not

force

Dave

leave

forget

Ed

+

-

+

+

subj

Dave

leave

OverviewOverview

IntroductionMotivationLocal Textual Inference

PARC’s XLE systemProcess pipelineAbstract Knowledge Representation (AKR)

Conceptual and temporal structureContextual structure and instantiability

Semantic relationsEntailments and presuppositionsRelative polarity

Entailment and Contradiction (ECD)

Demo!

AKR (Abstract Knowledge AKR (Abstract Knowledge Representation)Representation)

More specific entails less specificMore specific entails less specific

How ECD worksHow ECD works

Kim hopped.

Someone moved.

Text:

Hypothesis:Alignment

Specificitycomputation

Elimination ofElimination ofH facts that areH facts that are

entailed by T facts.entailed by T facts.

Kim hopped.

Someone moved.

Text:

Hypothesis:

Kim hopped.Text:

Hypothesis:

t

t

t

t

t

t

Context

Someone moved.

Alignment and specificity computationAlignment and specificity computation

Specificitycomputation

Alignment

Every (↓) (↑) Some (↑) (↑)

Every boy saw a small cat.

Every small boy saw a cat.

Text:

Hypothesis:

Every boy saw a small cat.

Every small boy saw a cat.

Text:

Hypothesis:

Every boy saw a small cat.

Every small boy saw a cat.

Text:

Hypothesis:

t

t

t

t

t

t

Context

Elimination of entailed termsElimination of entailed terms

Every boy saw a small cat.

Every small boy saw a cat.

Text:

Hypothesis:

t

t

Every boy saw a small cat.

Every small boy saw a cat.

Text:

Hypothesis:

t

t

Every boy saw a small cat.

Every small boy saw a cat.

Text:

Hypothesis:

t

t

Context

Contradiction:Contradiction:instantiable --- uninstantiableinstantiable --- uninstantiable

AKR modificationsAKR modifications

AKR0

P-AKR

Q-AKR

simplify

augment

Oswald killed Kennedy. => Kennedy died.

Kim managed to hop. => Kim hopped.

normalize

The situation improved.

The situation became better.

=>

ConclusionConclusion

Local textual inference is a good test bed for computational semantics.It is task-oriented. It abstracts away from particular meaning

representations and inference procedures.

It allows for systems that make purely linguistic inferences, others may bring in world knowledge and statistical reasoning.

This is a good time to be doing computational semantics.Purely statistical approaches have plateaued.

There is computing power for deeper processing.

Success might even pay off in real money.

DemoDemo

CreditsCredits

ASKER teamDaniel BobrowBob CheslowCleo CondoravdiDick Crouch (now at Powerset)Martin ForstRonald Kaplan (now at Powerset)Lauri KarttunenValeria de PaivaAnnie Zaenen

InternsRowan NairnMatt PadenKarl Pichotta

AQUAINT

ReferencesReferences• D. G. Bobrow, B. Cheslow, C. Condoravdi, L. Karttunen, T.H.

King, R. Nairn, V. de Paiva, C. Price, and A. Zaenen. PARC's Bridge and Question Answering System, Proceedings of the Grammar Engineering Across Frameworks (GEAF07) Workshop, pp. 46-66, CSLI Publications.

• Bobrow, D. G., C. Condoravdi, V. de Paiva, L. Karttunen, T. H. King, L. Price, R. Nairn, L.Price, A. Zaenen. Precision-focused Textual Inference, Proceedings of ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pp. 16-21.

• Crouch, Dick and Tracy Holloway King. Semantics via F-Structure Rewriting. Proceedings of LFG06, CSLI On-line Publications, pp. 145-165.

• Rowan Nairn, Cleo Condoravdi and Lauri Karttunen. Computing Relative Polarity for Textual Inference. Proceedings of ICoS-5 (Inference in Computational Semantics). April 20-21, 2006. Buxton, UK.

AQUAINT