Interpretation as Abduction

1
Interpretation as Abduction

• Maurizio Atzori
• atzori_at_di.unipi.it

Interpretation as Abduction (1993) Jerry R.
Hobbs, Mark Stickel, Douglas Appelt, and Paul
Martin
2
Summary

• Abduction in NLP
• The TACITUS Project
• The Abductive Commonsense Inference Text
  Understanding System
• Weighted Abduction
• Some Local Pragmatics

3
What is abduction?

• Deduction A, A ? B
• B
• Induction A(a1), A(a2),..., B(a1), B(a2),
  B(a3),...
• " x . A(x) Þ B(x)
• Abduction is inference to the best explanation

B, A ? B A
4
Logic as the Language of Thought

• The six keys of Cognitive processes
• Conjunction of concepts (P ? Q)
• Modus Ponens
• Recognition of Obvious Contradictions
• Predicate-Argument Relations
• We can relate different concept together
• Universal Instantiation
• In other words First-order logic!
• With no double negations or contrapositives

5
Nonmonotonic Logic as the Reasoning of Thought

• Monotonic logic KB A Þ KBÈX A
• Nonmonotonic KB A Þ KBÈX A
• E.g. negation as failure
• KB
• bird(x) Ù Øabnormal_bird(x)Þ fly(x)
• pinguin(x) Þ abnormal_bird(x)
• bird(a)
• fly(a) ? true
• KB KB È pinguin(a)
• fly(a) ? false

6
Discourse Understanding

• People understand discourse because they know so
  much
• How is knowledge used in the interpretation of
  discourse?
• We need to build a KB of commonsense and domain
  knowledge
• Local pragmatics
• Reference resolution
• Interpretation of compound nominals
• Syntactic/lexical ambiguity
• Metonymy resolution

7
Sentence Interpretation

• Prove the logical form of the sentence
• Together with the constraints that predicates
  impose on their arguments
• Allowing for coercions
• Merging redundancies where possible
• Making assumptions where necessary

8
Concrete Example

• A cargo train running from Lima to Lorohia was
  derailed before dawn today after hitting a
  dynamite charge.
• Inspector Eulogio Flores died in the explosion.
• The police reported that the incident took place
  past midnight in the Carahuaichi-Jaurin area.
• Incident Location Peru Carahuaichi-Jaurin
  (area)
• Incident Type Bombing
• Physical Target Description cargo train
• Physical Target Effect Some damage cargo
  train
• Human Target Name Eulogio Flores

9
Concrete Example Inferences

• Hitting a dynamite charge booming
• The target train that hit the charge
• The human target in the explosion
• Incident hitting of the dynamite charge
• In order to get the location

10
TACITUS

• Syntactic analysis / Semantic translation
  component (DIALOGIC)
• Obtained mergin a large grammar of English with a
  semantic translator for all the rules (DIAGRAM
  Project, Linguistic String Project)
• Produce a logical form of the sentence (no KB)
• Pragmatic component
• Produces an elaborated logical form inferences,
  assumptions, coreferences are explicited (KB)
• Task component
• Outputs the desired answer (e.g. diagnosis or
  database entries)

11
Most- or least-specific abduction?

• In many AI application, most-specific abduction
  is used
• E.g.
• In NLP application
• Sometimes least-specific abduction is better
• E.g. fluid we dont want to abduce lube oil
• Sometimes most-specific is better
• E.g. alarm sounded. Flow obstructed and the
  alarm is for the lube oil pressure we want to
  abduce that the flow is of lube oil

12
Weighted Abduction desiderata

• A new abduction scheme (3 features)
• Goals should be assumable
• Assumption at various levels of specificity
• Redundacy of text should be taken into account
  (yielding more economic proofs)

13
Weighted Abduction solution

• Every conjunct in the logical form of the
  sentence is given an assumability cost
• If cost(Q)c then cost(P1) is w1c
• If (...,x,y,...) ...,q(x)20,q(y)10,...
• Then (...,x,...) ...,q(x)10,...
• leading to minimality through redundancies
• Eg.

14
Weighted Abduction examples

• How much does it cost to prove Q?
• C, or 0.6 if we already know P1 or P2
• Q1? Least-specific 10
• Q1 Ù Q2? Most-specific! 18 instead of 20!

Cost(Q1)10 Cost(Q2)10
15
Weighted Abduction et cetera

• (" x) lube-oil(x) Þ fluid(x)
• It is abductively unuseful
• Flow obstructed. Metal particles in lube oil
  filter
• ( x) lube-oil(x) but we cannot infer fluid(x) ?
• (" x) fluid(x) Þ lube-oil(x)
• It works but we havent such an axiom
• It is false!
• (" x) fluid(x) Ù etc1(x) Þ lube-oil(x)
• etc(x) is something like abnormal (special)
  fluid
• It can only be assumed, never proved

16
Local Pragmatics Phenomena

• Definite Reference
• I bought a new car last week. The car is already
  giving me trouble.
• I bought a new car last week. The vehicle is
  already giving me trouble.
• I bought a new car last week. The engine is
  already giving me trouble.
• The engine of my new car is already giving me
  trouble.
• KB
• (" x) car(x) Þ vehicle(x)
• (" x) car(x) Þ ( x) vehicle(x)

17
Lexical Ambiguity

• John wanted a loan. He went to the bank.
• KB
• bank1(x) Þ bank(x) banca
• bank2(x) Þ bank(x) riva
• loan(y) Þ financial-institution(x) Ù issue(x,y)
• financial-institution(x) Ù etc1(x) Þ bank1(x)
• river(z) Þ bank2(x) Ù borders(x,z)

18
Lexical Ambiguity Abduction

• ... Ù bank(x) Ù ...
• bank1(x) Þ bank(x)
• financial-institution(x) Ù etc1(x) Þ bank1(x)
• loan(y) Þ financial-institution(x) Ù issue(x,y)
• loan(L)

19
Compound Nominals

• Turpentine jar.
• ( x, y) turpentine(y) Ù jar(x) Ù nn(y, x)
• KB
• (" y) liquid(y) Ù etc1(y) Þ turpentine(y)
• (" e1, x, y) function(e1, x) Ù contain(e1, x, y)
  Ù liquid(y) Ù etc2 (e1, x, y) Þ jar(x)
• If the function of something is to contain
  liquid, then it may be a jar
• (" e1, x, y) contain(e1, x, y) Þ nn(y, x)

20
Compound Nominals Abduction

• turpentine(y) Ù nn(y, x) Ù jar(x)
• liquid(y) Ù etc1(y) Þ turpentine(y)
• contain(e1, x, y) Þ nn(y, x)
• liquid(y) Ù function(e1, x) Ù contain(e1, x, y)
  Ù etc2 (e1, x, y) Þ jar(x)

21
Other Local Pragmatics

• Exploiting Redundancy
• Coreference Problems
• Distinguishing the Given and the New

22
Integration with other approaches

• Interpretation as abduction
• Parsing as deduction
• It becomes possible to integrate syntax,
  semantics and pragmatics in a very thorough and
  elegant way.

23
Applications

• Text understanding
• TACITUS Project at SRI
• Equipment failure reports
• Naval operations reports
• Terrorist reports
• Question Answering!
• FALCONs postprocessor makes use of this
  abductive framework
• Select the right answer among some candidate
  documents

24
References (1/3)

• Hobbs, Jerry R., 2001. Abduction in Natural
  Language Understanding, to appear in L. Horn and
  G. Ward (eds.), Handbook of Pragmatics, Blackwell
• Thomason, Richmond H., and Jerry R. Hobbs, 1997.
  Interrelating Interpretation and Generation in an
  Abductive Framework, Proceedings, AAAI Fall
  Symposium Workshop on Communicative Action in
  Humans and Machines, Cambridge, Massachusetts,
  November 1997, pp. 97-105
• Hobbs, Jerry R., 1992. Metaphor and Abduction, in
  A. Ortony, J. Slack, and O. Stock, eds.,
  Communication from an Artificial Intelligence
  Perspective Theoretical and Applied Issues,
  Springer-Verlag, Berlin, pp. 35-58. Also
  published as SRI Technical Note 508, SRI
  International, Menlo Park, California. August 1991

25
References (2/3)

• Hobbs, Jerry R., Douglas E. Appelt, John Bear,
  Mabry Tyson, and David Magerman, 1991. The
  TACITUS System The MUC-3 Experience, SRI
  Technical Note 511, SRI International, Menlo
  Park, California. November 1991
• Stickel, M.E., 1991. A Prolog-like inference
  system for computing minimum-cost abductive
  explanations in natural-language interpretation.
  Annals of Mathematics and Artificial Intelligence
  4 (1991), 89-106
• Hobbs, Jerry R., and Megumi Kameyama, 1990.
  Translation by Abduction, in H. Karlgren, ed.,
  Proceedings, Thirteenth International Conference
  on Computational Linguistics, Helsinki, Finland,
  Vol. 3, pp. 155-161, August, 1990

26
References (3/3)

• Tyson, Mabry, and Jerry R. Hobbs, 1990.
  Domain-Independent Task Specification in the
  TACITUS Natural Language System, Technical Note
  488, Artificial Intelligence Center, SRI
  International, May 1990
• Hobbs, Jerry R., 1990. An Integrated Abductive
  Framework for Discourse Interpretation,
  Proceedings, AAAI Spring Symposium on Abduction,
  Stanford, California, March 1990
• Hobbs, Jerry R., 1989. The Use of Abduction in
  Natural Language Processing, Proceedings, Nagoya
  International Symposium on Knowledge Information
  and Intelligent Communication, Nagoya, Japan,
  November 1989
• Hobbs, Jerry R., and Paul Martin 1987. Local
  Pragmatics. Proceedings, International Joint
  Conference on Artificial Intelligence, pp.
  520-523. Milano, Italy, August 1987.