Automatic knowledge base refinement for classification systems

摘要

An automated approach to knowledge base refinement, an important aspect of knowledge acquisition is described. Using empirical performance analysis, SEEK2 extends the capabilities of its predecessor rule refinement system, SEEK [17]. In this paper, the progress made since the original SEEK program is described: (a) SEEK2 works with a more general class of knowledge bases than SEEK, (b) SEEK2 has an automatic refinement capability, it can perform many of the basic tasks involved in knowledge base refinement without human interaction, (c) a metalanguage for knowledge base refinement has been specified which describes knowledge about the refinement process. Methods for estimating the expected gain in performance for a refined knowledge base and prospective test cases are described and some results are reported. An approach to justifying refinement heuristics is discussed.