Obtaining a minimal set of rewrite rules

Abstract

In this paper the authors describe a new approach to rewrite rule extraction and analysis, using Minimal Representation Graphs. This approach provides a mechanism for obtaining the smallest possible rule set – within a context-dependent rewrite rule formalism – that describes a set of discrete training data completely, as an indirect approach to obtaining optimal accuracy on an unseen test set. They demonstrate the application of this technique for a pronunciation prediction task.