Barcelona, 21 January 2005

1
Towards Wide-Coverage Semantic Interpretation of
TextsJohan BosSchool of InformaticsUniversity
of EdinburghScotland,UK
2
Semantic Interpretation

• Syntactic analysis
• Semantic analysis
• Inference

S
NP
VP
DET
VP
N
MOD
AUX
VP
PN
POS
Ralph s car did not start
3
Semantic Interpretation

• Syntactic analysis
• Semantic analysis
• Inference

4
Semantic Interpretation

• Syntactic analysis
• Semantic analysis
• Inference

Ralph s car didnt start ------------------------
-- Ralph has a car
5
We can do this for small domains

• Applications
• Machine Translation
• Spoken Dialogue Systems

6
But can we do this on a larger scale?

• Achieving wide coverage/robustness
• Semantic interpretation of
• newspaper texts
• web pages
• Five years ago impossible

7
Wide-Coverage Semantic Interpretation

• Talk outline
• 1. Semantic Formalism
• 2. Combinatorial Categorial Grammar
• 3. Wide-Coverage Parsing
• 4. Inference
• 5. Conclusion

8
Computational Semantics

• SemanticsThe branch of logic and linguistics
  concerned with meaning
• Computational SemanticsUsing a computer to build
  meaning representations, and reason with the
  result (inference)

9
Discourse Representation Theory

• DRT well understood formalism
• Scope, anaphora, presupposition, tense, etc.
• Kamp 81, Kamp Reyle 93, Van der Sandt 92
• Representations (DRSs) can be build using
  traditional tools
• Lambda calculus
• Underspecification
• Model-theoretic interpretation
• Inference possible
• Translation to first-order logic

10
DRT example

• Discourse Representation Structures (DRS)
• Box structures as a partial model
• Structure plays role in pronoun resolution
• Phenomena not covered
• Plurals
• Tense Aspect
• Ellipsis

A spokesman lied.
11
Talk Outline

• 1. Semantic Formalism
• 2. Combinatorial Categorial Grammar
• 3. Wide-Coverage Parsing
• 4. Inference
• 5. Conclusion

12
Combinatorial Categorial Grammar

• Semantic interpretation requires syntactic
  analysis
• We use CCG (Steedman 2001) because
• Deals with complex cases of coordination and
  long-distance dependencies
• Lexicalised, therefore easy to implement
• Fast wide-coverage parser available

13
CCG type-driven lexicalised grammar
14
CCG small set of combinatorial rules

• Forward Application (FA)
• Backward Application (BA)
• Generalised Forward Composition (FC)
• Backward Crossed Composition (BC)
• Type Raising (TR)
• Coordination

15
CCG derivation

• NP/Na Nspokesman S\NPlied

16
CCG derivation

• NP/Na Nspokesman S\NPlied

17
CCG derivation

• NP/Na Nspokesman S\NPlied
• ------------------------------- (FA)
• NP a spokesman

18
CCG derivation

• NP/Na Nspokesman S\NPlied
• ------------------------------- (FA)
• NP a spokesman
• ----------------------------------------
  (BA)
• S a spokesman lied

19
Coordination in CCG

• s/(s\np)Artie (s\np)/nplikes (x\x)/xand
  s/(s\np)Tony (s\np)/nphates npbeans
• ------------------------------------ (FC)
  --------------------------------------- (FC)
  
• s/np Artie likes
  s/npTony hates
• 
  --------------------------------------------------
  ----- (FA)
• 
  (s/np)\(s/np)and Tony hates
• ----------------------------------
  -----------------------------------------------
  (BA)
• 
  s/np Artie likes and Tony hates
• 
  ----------------------------------------
  -------------- (FA)
• 
  s Artie likes and Tony
  hates beans

20
CCG lexical semantics
21
CCG derivation

• NP/Na Nspokesman
  S\NPlied
• ?p. ?q. p(x)q(x) ?x.
  ?y.
• --------------------------------------------------
  ------ (FA)
• NP a spokesman
• ?q. q(x)
• ---------------------------------------
  ----------------------------------------- (BA)
• 
  S a spokesman lied

22
Talk Outline

• 1. Semantic Formalism
• 2. Combinatorial Categorial Grammar
• 3. Wide-Coverage Parsing
• 4. Inference
• 5. Conclusion

23
The Clark Curran Parser

• Use standard statistical techniques
• Robust wide-coverage parser
• State-of-the-art performance
• Clark Curran (ACL 2004)
• Grammar derived from CCGbank
• 409 different categories
• Hockenmaier Steedman (ACL 2002)

24
Combining CCG with DRT

• Supply semantic representations for CCG
  categories
• most frequent categories only (ca. 300)
• using lambda calculus
• Interpret combinatorial rules in terms of
  functional application
• Input CCG derivation
• Output Lambda-expression comprising a DRS
• Results 96 coverage WSJ

25
System Overview
Raw Text
Tokeniser
Tokenised Text
CCG-parser
CCG Derivation
DRS-builder
DRSs
26
Talk Outline

• 1. Semantic Formalism
• 2. Combinatorial Categorial Grammar
• 3. Wide-Coverage Parsing
• 4. Inference
• 5. Conclusion

27
Inference

• Computational semantics without inferenceis not
  computational semantics!
• Inference tasks
• Consistency checking
• Informativeness checking
• Use first-order logic
• Theorem proving
• Model building
• Background knowledge required

28
Why First-Order Logic?

• Why not use higher-order logic?
• Better match with formal semantics
• But Undecidable/no fast provers available
• Why not use weaker logics?
• Modal/description logics (decidable fragments)
• But Cant cope with all of natural language
• Why use first-order logic?
• Undecidable, but good inference tools available
• DRS translation to first-order logic

29
(No Transcript)
30
Yin and Yang of Inference

• Theorem Proving and Model Building function as
  opposite forces
• Assume ?, a logical formula, representing a
  certain discourse ?
• If a theorem prover succeeds in finding a proof
  for ??, then ? is inconsistent
• If a model builder succeeds to construct a model
  for ?, then ? is consistent

31
Inference Example

• DRS
• FOL ?x?y(spokesman(x) lie(y) agent(y,x))
• Model D d1,d2,d3
  F(spokesman)d1,d2 F(lie)d3
  F(agent)(d3,d2)

32
Background Knowledge

• Background Knowledge
• Use hypernym relations from WordNet to build an
  ontology
• Create MiniWordNets for small texts
• Convert these into first-order axioms
• MiniWordNet for example text
• There is no asbestos in our products now. Neither
  Lorillard nor the researchers who studied the
  workers were aware of any research on smokers of
  the Kent cigarettes.

33
MiniWordNet
34
MiniWordNet
?x(asbestos(x)?physical_thing(x)) ?x(asbestos(x)??
cigarette(x))
35
System Overview
Raw Text
Tokeniser
Tokenised Text
CCG-parser
CCG Derivation
DRS-builder
DRSs
36
System Overview
Raw Text
Tokeniser
Theorem Prover (Vampire7)
proofs
Tokenised Text
Model Builder (Paradox)
models
CCG-parser
CCG Derivation
WordNet
DRS-builder
Inference
DRSs
37
Inference Results

• Consistency checking on WSJ
• 95 of DRSs consistent
• Inconsistent cases due to
• Inconsistencies in WordNet
• Errors in CCG analysis
• Inadequate semantic analysis (N coordination)
• WSD
• Use consistency check as filter for ambiguity
  resolution

38
Example inconsistent DRS

• Germans vote for a party , rather than a person ,
  and
• __________________________________
• x12 x16 x15 x10 x6 x13 x4
• ----------------------------------
• germans(x16)
• nn(x16,x12)
• vote(x12)
• party(x15)
• for(x12,x15)
• ______________
• x14
• __ --------------
• person(x14)
• for(x12,x14)
• ______________

39
Talk Outline

• 1. Semantic Formalism
• 2. Combinatorial Categorial Grammar
• 3. Wide-Coverage Parsing
• 4. Inference
• 5. Conclusion

40
Conclusion

• Results
• Converting raw text into semanticrepresentations
  (coverage 96)
• Implementing inference using tools from automated
  reasoning (FOL)
• Automatic background knowledge generation
  (MiniWordNets)
• System useful for NLP applications
• Question answering, semantic web, summarisation,
  machine translation

41
Future Work

• Computing background knowledge
• Use resources other than WordNet(FrameNet,
  VerbNet, PrepNet, XWN)
• Measuring semantic accuracy
• Build gold standard corpus
• Measuring semantic adequacy
• FRACAS natural lang. inference test suite
• Entailment task (PASCAL challenge)