A Linguistically Based Semantic Bias for Theory Revision
Clifford A. Brunk · 1998
We present a new approach to theory revision that uses a linguistically based semantics to help detect and correct errors in classification rules. The idea is that preferring linguistically cohesive revisions will enhance the comprehensibility and ultimately the accuracy of rules. We explain how to associate terms in the rules with elements in a lexical class hierarchy and use distance within the hierarchy to estimate linguistic cohesiveness. We evaluate the utility of this approach empirically using two relational domains. 1 INTRODUCTION Theory revision is the task of making a theory, which may contain errors, consistent with a set of examples. It is one of the tasks performed by a knowledge engineer during the creation of a rule-based expert system. Within the machine learning community there has been a focus on developing approaches to automate the task of using a set of examples to identify and repair errors in classification rules (e.g., SEEK2 (Ginsberg, Weiss & Politakis 1985), ...