Case-based Problem Solving with a Large Knowledge Base of Learned Cases

Recent experiments indicate that a case-based approach to the problem of word pronunciation is effective as the basis for a system that learns to pronounce English words. More generally, the approach taken here illustrates how a case-based reasoner can access a large knowledge base containing hundreds of potentially relevant cases and consolidate these multiple knowledge sources using numerical relaxation over a structured network. In response to a test item, a search space is first generated and structured as a lateral inhibition network. Then a spreading activation algorithm is applied to this search space using activation levels derived from the case base. In this paper we describe the general design of our model and report preliminary test results based on a training vocabulary of 750 words. Our approach combines traditional heuristic methods for memory organization with connectionist-inspired techniques for network manipulation in an effort to exploit the best of both information-processing methodologies.