Interpolation as Abduction

Hobbs, Stickel, Martin, Edwards 1988

This article describes TACITUS, a text understanding system that is based on abductive reasoning. TACITUS integrates syntax, semantics, and pragmatics into a single system, and they demonstrate how it can be used to handle coreference, propositional attatchment ambiguity, metonymy, and compositional noun interpretation.

TACITUS maintains a first-order predicate calculus representation of the world; and when it is presented with a new sentence, it attempts to prove the it. Usually, the proof will require new axioms; but TACITUS tries to use the minimum number of new axioms necessary. In particular, axioms are assigned "costs," and TACITUS picks the proof for the new sentence with a minimal total axiom cost.

The cost of assuming an axiom without any justification is 1. But if we can use abductive reasoning to produce the axiom, then its cost may be less (abductive reasoning is inferring the antecedant from the consequent). Intuitively, this makes some sense -- if we know that X occured, and Y is a possible cause for X, then we're more likely to think that Y occured.

In order to prevent over-generation, a type hierarchy is used to impose selectional restrictions on predicate expressions.

This approach requires the user to specify a large set of general axioms, encoding common-sense and domain-specific knowledge. As the domain gets larger, the interactions between these axioms gets quite complex, and tweaking the weights gets very difficult. Using a more modern approach, we might cast the system as a bayesian net, where axioms are pieces of the next that can be used for forward/backward probablistic reasoning. But clearly the bayesian net would be quite large, and estimating its parameters would require enormous amounts of data.

@InProceedings{hobbs1988,
  author =       {Jerry R. Hobbs, Mark Stickel, Paul Martin, and Douglas Edwards},
  title =        {Interpretation as Abduction},
  booktitle =    {Proceedings of ACL 1988},
  year =         1988
  url =          {http://acl.ldc.upenn.edu/P/P88/P88-1012.pdf}
}