A General Approach to Specificity in Default Reasoning

Abstract

We present an approach addressing the notion of specificity, or of preferring a more specific default sentence over a less specific one, in commonsense reasoning. Historically, approaches have either been too weak to provide a full account of defeasible reasoning while accounting for specificity, or else have been too strong and fail to enforce specificity. Our approach is to use the techniques of a weak system, as exemplified by System Z, to isolate minimal sets of conflicting defaults. From the specificity information intrinsic in these sets, a default theory in a target language is specified. In this paper we primarily deal with theories expressed (ultimately) in Default Logic. However other approaches would do just as well, as we illustrate by also considering Autoepistemic Logic and variants of Default Logic. In our approach, the problems of weak systems, such as lack of adequate property inheritance and (occasional) unwanted specificity relations, are avoided. Also, difficulties inherent with stronger systems, in particular, lack of specificity are addressed. This work differs from previous work in specifying priorities in Default Logic, in that we obtain a theory expressed in Default Logic, rather than ordered sets of rules requiring a modification to Default Logic.