Analogical Modeling of Pronunciation: Vowels in English Monosyllables

English spelling is related to its pronunciation by complicated and numerous rule-like patterns. Previously, most work on this relationship has focused on eager learning, the a priori extraction of rules that are later applied to predict pronunciation. Psychological research, meanwhile, has tended away from a rules-based approach to reading, favoring analogy-based models. Readers seem to incorporate knowledge of similar words (whether regularly or irregularly pronounced) when reading a word with which they are unfamiliar.This thesis takes an exemplar-based or lazy-learning approach to the task of predicting vowel pronunciations in English monosyllables, using Analogical Modeling of Language to discover regularities within a large mass of data and select a pronunciation by a formalized measure of similarity. It uses data from a large natural-language database, basing its predictions on the spelling of the vowel nucleus itself and on neighboring consonants. Related systems, using related approaches, have reached accuracies from 9o% to 95%; the grounds for expecting similar results are discussed here. The thesis also proposes problems to be addressed by future work.

Thesis Author: Rasmussen, Nathan E.

Year Completed: 2005

Committee Members: Alan Melby, Royal Skousen

Thesis Chair: Deryle Lonsdale