A DEFAULT HIERARCHY FOR PRONOUNCING ENGLISH

In this correspondence, we study the principles governing the power and efficiency of the default hierarchy, a system of knowledge acquisition and representation. The default hierarchy consists of clear and accessible rules, like an expert-made system, but trains automatically when exposed to data, like a neural network. The hierarchy has both general (default) and specific rules. In training, specific rules are learned only if they are exceptions to general rules; in using the hierarchy, default rules are used when no relevant specific rules are found. Our application of the default hierarchy to the task of pronouncing written English reveals interesting properties of the default hierarchy architecture. We find that the hierarchy performs best when there is free access to general rules, and that it is less than 1/4 the size of a comparable nonhierarchial rule set while no less accurate in pronunciation. Evaluating the hierarchy as a pronouncer of English, we find that its rules capture several key features of English spelling. Moreover, the default hierarchy pronounces English better than the neural network NETtalk, and almost as well as expert-devised systems (DECtalk and the Naval Research Laboratory's pronunciation system).