Computing Textual Inferences

1. Computing Textual Inferences Cleo Condoravdi Palo Alto Research Center Georgetown University Halloween, 2008

2. Overview • Motivation • Local Textual Inference • Textual Inference Initiatives and refinements • PARC’s BRIDGE system • XLE • Abstract Knowledge Representation (AKR) • Conceptual and temporal structure • Contextual structure and instantiability • Semantic relations • Entailments and presuppositions • Relative polarity • Entailment and Contradiction (ECD) • (Interaction with external temporal reasoner)

3. Access to content: existential claimsWhat happened? Who did what to whom? • Microsoft managed to buy Powerset. •  Microsoft acquired Powerset. • Shackleton failed to get to the South Pole. • Shackleton did not reach the South Pole. • The destruction of the file was not illegal. • The file was destroyed. • The destruction of the file was averted. • The file was not destroyed.

4. Access to content: monotonicityWhat happened? Who did what to whom? • Every boy managed to buy a small toy. •  Every small boy acquired a toy. • Every explorer failed to get to the South Pole. • No experienced explorer reached the South Pole. • No file was destroyed. • No sensitive file was destroyed. • The destruction of a sensitive file was averted. • A file was not destroyed.

5. Access to content: temporal domainWhat happened when? • Ed visited us every day last week. •  Ed visited us on Monday last week. • Ed has been living in Athens for 3 years. • Mary visited Athens in the last 2 years. • Mary visited Athens while Ed lived in Athens. • The deal lasted through August, until just before the government took over Freddie. (NYT, Oct. 5, 2008) •  The government took over Freddie after August.

6. Grammatical analysis for access to content • Identify “Microsoft” as the buyer argument of the verb “buy” • Identify “Shackleton” as the traveler argument of the verb “get to” • Identify lexical relation between “destroy” and “destruction” • Identify syntactic relation between verbal predication “destroy the file” and nominal predication “destruction of the file” • Identify the infinitival clause as an argument of “manage” and “fail” • Identify the noun phrases “every”, “a”, “no” combine with • Identify the phrases “every day”, “on Monday”, “last week” as modifiers of “visit” • Identify “has been living” as a present progressive

7. Knowledge about words for access to content • The verb “acquire” is a hypernym of the verb “buy” • The verbs “get to” and “reach” are synonyms • Inferential properties of “manage”, “fail”, “avert”, “not” • Monotonicity properties of “every”, “a”, “no”, “not” • Restrictive behavior of adjectival modifiers “small”, “experienced”, “sensitive” • The type of temporal modifiers associated with prepositional phrases headed by “in”, “for”, “on”, or even nothing (e.g. “last week”, “every day”)

8. Toward NL Understanding • Local Textual Inference • 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. • Veridicality reasoning • Did an event mentioned in the text actually occur? • Temporal reasoning • When did an event happen? How are events ordered in time? • Spatial reasoning • Where are entities located and along which paths do they move? • Causality reasoning • Enablement, causation, prevention relations between events

9. Local Textual Inference • PASCAL RTE Challenge (Ido Dagan, Oren Glickman) 2005, 2006 • PREMISE • CONCLUSION • TRUE/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 hiscapacity as the president of the European commission. • George Bush is the president of the European commission. • FALSE (= not entailed by the premise)

10. PARC Entailment and Contradiction Detection (ECD) • Text: Kim hopped. • Hypothesis: Someone moved. • Answer: TRUE • Text: Sandy touched Kim. • Hypothesis: Sandy kissed Kim. • Answer: UNKNOWN • Text: Sandy kissed Kim. • Hypothesis: No one touched Kim. • Answer: NO • Text: Sandy didn’t wait to kiss Kim. • Hypothesis: Sandy kissed Kim. • Answer: AMBIGUOUS

11. Linguistic meaning vs. speaker meaning • Not a pre-theoretic but rather a theory-dependent distinction • Multiple readings • ambiguity of meaning? • single meaning plus pragmatic factors? • The diplomat talked to most victims • The diplomat did not talk to all victims • UNKNOWN / YES I don’t know which. • You can have the cake or the fruit. • You can have the fruit YES UNKNOWN

12. 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)

13. Recognizing textual entailments • Monotonicity Calculus, Polarity, Semantic Relations • Much of language-oriented reasoning is tied to specific words, word classes and grammatical features • Class 1: “fail”, “refuse”, “not”, … • Class 2: “manage”, “succeed”, … • Tenses, progressive –ing form, … • Representation and inferential properties of modifiers of different kinds • throughout July vs. in July • for the last three years vs. in the last three years • sleep three hours -- duration • sleep three times -- cardinality

14. XLE 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 CSTRUCTURE OT marks PCFG-based chart pruning c-structures “general” OT marks log-linear model c-/f-structure pairs

15. Ambiguity is rampant in language • Alternatives multiply within and across layers… Abstract KR C-structure F-structure Linguistic Sem KRR

16. Oops: Strong constraints may reject the so-far-best (= only) option Statistics X What not to do • Use heuristics to prune as soon as possible X X X C-structure F-structure Linguistic sem Abstract KR KRR X Fast computation, wrong result

17. Options multiplied out The sheep-sg liked the fish-sg. The sheep-pl liked the fish-sg. The sheep-sg liked the fish-pl. The sheep-pl liked the fish-pl. Options packed sgpl sgpl The sheep liked the fish Manage ambiguity instead How many sheep? How many fish? The sheep liked the fish. Packed representation: • Encodes all dependencies without loss of information • Common items represented, computed once • Key to practical efficiency with broad-coverage grammars

18. System Overview LFG Parser syntactic F-structure string “A girl hopped.” rewrite rules AKR (Abstract Knowledge Representation)‏

19. XLE Pipeline

20. XLE System ArchitectureText  AKR Parse text to f-structures • Constituent structure • Represent syntactic/semantic features (e.g. tense, number) • Localize arguments (e.g. long-distance dependencies, control) • 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

21. AKR representation concept term WordNet synsets thematic role instantiability facts event time A collection of statements

22. F-structures vs. AKR • Nested structure of f-structures vs. flat AKR • F-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 senses • Sentential 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

23. F-Structure to AKR Mapping • Input: F-structures • Output: clausal, abstract KR • Mechanism: packed term rewriting • Rewriting system controls • lookup of external ontologies via Unified Lexicon • compositionally-driven transformation to AKR • Transformations: • Map words to Wordnet synsets • Canonicalize semantically equivalent but formally distinct representations • Make conceptual & intensional structure explicit • Represent semantic contribution ofparticular constructions

24. F-Structure to AKR Mapping • Input: F-structures • Output: clausal, abstract KR • Mechanism: packed term rewriting • Rewriting system controls • lookup of external ontologies via Unified Lexicon • compositionally-driven transformation to AKR • Transformations: • Map words to Wordnet synsets • Canonicalize semantically equivalent but formally distinct representations • Make conceptual & intensional structure explicit • Represent semantic contribution ofparticular constructions

25. Basic structure of AKR • Conceptual Structure • Predicate-argument structures • Sense disambiguation • Associating roles to arguments and modifiers • Contextual Structure • Clausal complements • Negation • Sentential modifiers • Temporal Structure • Representation of temporal expressions • Tense, aspect, temporal modifiers

26. Ambiguity management with choice spaces girl with a telescope seeing with a telescope

27. Conceptual 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.”

28. temporalRel(startsAfterEndingOf,say:6,Now) temporalRel(startsAfterEndingOf,say:6,talk:21) “Bill will say that Ed talked.” Temporal Structure • Matrix vs. embedded tense • temporalRel(startsAfterEndingOf,Now,talk:6) Shared by “Ed talked.” and “Ed did not talk.”

29. 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) Canonicalization in conceptual structure “John took a tour of Europe.” “John traveled around Europe.”

30. Contextual Structure • Use of contexts enables flat representations • Contexts as arguments of embedding predicates • Contexts as scope markers • 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.

31. Concepts and Contexts • Concepts live outside of contexts. • Still we want to tie the information about concepts to the contexts they relate to. • Existential commitments • Did 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.

32. Instantiability • 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.

33. Negation “Ed did not talk” • Contextualstructure • context(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

34. Relations between contexts • Generalized entailment: veridical • If c2 is veridical with respect to c1, the information in c2 is part of the information in c1 • Lifting rule: instantiable(Sk, c2) => instantiable(Sk, c1) • Inconsistency: antiveridical • If c2 is antiveridical with respect to c1, the information in c2 is incompatible with the info in c1 • Lifting rule: instantiable(Sk, c2) => uninstantiable(Sk, c1) • Consistency: averidical • If c2 is averidical with respect to c1, the info in c2 is compatible with the information in c1 • No lifting rule between contexts

35. Determinants of context relations • Relation depends on complex interaction of • Concepts • Lexical entailment class • Syntactic environment • Example Hedidn’t remembertoclose the window. He doesn’t remember thatheclosed the window. He doesn’t remember whether heclosed the window. • Heclosed the window. • Contradicted by 1 • Implied by 2 • Consistent with 3

36. Embedded 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.

37. Embedded examples in real text • From Google: • Song, Seoul's point man, did not forget to persuade the North Koreans to make a “strategic choice” of returning to the bargaining table... Song persuaded the North Koreans… The North Koreans made a “strategic choice”…

38. Semantic relations • Presupposition (Factive verbs, Implicative verbs) • It is surprising that there are no WMDs in Iraq. • It is not surprising that there are no WMDs in Iraq. • Is it surprising that there are no WMDs in Iraq? • If it is surprising that there are no WMDs in Iraq, it is because we had good reasons to think otherwise. • Entailment (Implicative verbs) • It has been shown that there are no WMDs in Iraq. • It has not been shown that there are no WMDs in Iraq. • Has it been shown that there are no WMDs in Iraq? • If it has been shown that there are no WMDs in Iraq, the war has turned out to be a mistake.

39. Factives Class Inference Pattern Positive Negative

40. Implicatives Class Inference Pattern Two-way implicatives One-way implicatives

41. Implicatives under Factives • It is surprising that Bush dared to lie. Bush lied. It is not surprising that Bush dared to lie.

42. Ability Noun (ability/means) = --Implicative Have + Chance Noun = --Implicative (chance/opportunity) = ++/--Implicative Character Noun (courage/nerve) Chance Noun = ++/--Implicative (chance/opportunity) Asset Noun (money) = ++/--Implicative Take + Effort Noun (trouble/initiative) = ++/--Implicative (chance/opportunity) Chance Noun = ++/--Implicative Use + = ++/--Implicative Asset Noun (money) Chance Noun (chance/opportunity) = +-/-+Implicative Waste + (money) Asset Noun = ++/--Implicative Miss + (chance/opportunity) Chance Noun = +-/-+Implicative + Seize Chance Noun (chance/opportunity) = ++/--Implicative Phrasal Implicatives

43. Conditional verb classes • Joe had the chutzpah to steal the money.⇝Joe stole the money. Two-way implicative with “character nouns” “character noun” (gall, gumption, audacity…)

44. Relative 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

45. Example: polarity propagation • “Ed did notforget to force Dave to leave.” • “Dave left.”

46. leave subj Dave + not comp - forget + subj comp Ed force + subj obj comp Ed Dave leave subj Dave

47. Summary of basic structure of AKR Conceptual Structure Terms representing types of individuals and events, linked to WordNet synonym sets by subconcept declarations. Concepts typically have roles associated with them. Ambiguity is encoded in a space of alternative choices. Contextual Structure t is the top-level context, some contexts are headed by some event term Clausal complements, negation and sentential modifiers also introduce contexts. Contexts can be related in various ways such as veridicality. Instantiability declarations link concepts to contexts. Temporal Structure Locating events in time. Temporal relations between events.

48. ECD • ECD operates on the AKRs of the passage and of the hypothesis • ECD operates on packed AKRs, hence no disambiguation is required for entailment and contradiction detection • If one analysis of the passage entails one analysis of the hypothesis and another analysis of the passage contradicts some other analysis of the hypothesis, the answer returned is AMBIGUOUS • Else: If one analysis of the passage entails one analysis of the hypothesis, the answer returned is YES • If one analysis of the passage contradicts one analysis of the hypothesis, the answer returned is NO • Else: The answer returned is UNKNOWN

49. AKR (Abstract Knowledge Representation)

50. More specific entails less specific