Rationale and methods for abductive reasoning in natural-language interpretation

Abstract

By determining those added assumptions sufficient to make the logical form of a natural-language sentence provable, abductive inference can be used in the interpretation of sentences to determine the information to be added to the listener's knowledge, i.e., what the listener should learn from the sentence. Some new forms of abduction are more appropriate to the task of interpreting natural language than those used in the traditional diagnostic and design synthesis applications of abduction. In one new form, least specific abduction, only literals in the logical form of the sentence can be assumed. The assignment of numeric costs to axioms and assumable literals permits specification of preferences on different abductive explanations. Least specific abduction is sometimes too restrictive. Better explanations can sometimes be found if literals obtained by backward chaining can also be assumed. Assumption costs for such literals are determined by the assumption costs of literals in the logical form and functions attached to the antecedents of the implications. There is a new Prolog-like inference system that computes minimum-cost explanations for these abductive reasoning methods.

This research is supported by the Defense Advanced Research Projects Agency, under Contract N00014-85-C-0013 with the Office of Naval Research, and by the National Science Foundation, under Grant CCR-8611116. The views and conclusions contained herein are those of the author and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the Defense Advanced Research Projects Agency, the National Science Foundation, or the United States government. Approved for public release. Distribution unlimited.