An Automated Learner for Phonology and Morphology

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

This basis of this project is a machine-implemented algorithm that learns the phonological and morphological rules governing inflectional paradigms. The algorithm, following exposure to a set of learning data, can generate appropriately inflected forms for novel stems. The system has the capacity to generate multiple guesses for a given stem, and offers gradient "well-formedness judgments" for each, which correlate well with the well-formedness judgments of human speakers. The project will involve two activities. First, the algorithm will be used in exploring areas of phonology and morphology that have hitherto been underinvestigated because they require the kind of detailed rule systems that can best be discovered by machine. Pilot evidence, primarily from Italian, indicates that people learn the phonological generalizations of paradigms in surprising detail, in a way that is approximated by the machine algorithm. The project will explore such detailed systems further though large-scale "Wug"-testing experiments, in which subjects provide inflected forms for nonce stems. Tests will be carried out in four areas. *English past tenses* are important because they relate to a large body of work in computational modeling and psycholinguistics. *Spanish diphthongization* exemplifies a process that is subject to irregularity, but is phonological rather than morphological. *Korean stem-final consonant alternations* are involved in a major diachronic restructuring, which the algorithm will be used to model. *V ~ O alternations in Polish* bring evidence from the learning algorithm (and parallel judgments and productions of native speakers) to bear on theoretical issues of abstractness and opacity. The learning algorithm will be developed further, incorporating resources from computational linguistics: richer phonological representations, more sophisticated methods of finding phonological rules, and a morphological parser to serve as a front end. The revised algorithm will also incorporate elements of phonological theory: a projection-based theory of long-distance rules (permitting the algorithm to discover vowel-to-vowel processes), and a conception of derived-environment processes that proceeds naturally from the algorithm. The proposed project will (a) increase our knowledge of how phonology and morphology work at the level of detailed generalizations; (b) provide a test for phonological theories; (c) contribute to the education of linguistics students in developing lexical databases, programming, and carrying out experiments; (d) provide databases usable for testing learners developed by other researchers.