Colin Wilson, Linguistics Colloquium: Learning phonological classification

Colloquium | November 27 | 3:10-5 p.m. | 370 Dwinelle Hall

 Colin Wilson, Johns Hopkins University

 Department of Linguistics

Many idiosyncratic properties of a lexical item — such as its inflection/declension class or grammatical gender, the allomorphs that it selects, and the non-automatic alternations that it undergoes — can be partly predicted from its phonological form. Patterns of this kind can be analyzed with probabilistic models in which weighted phonological properties induce distributions over lexical classifications. Previous work has demonstrated that native speakers have some knowledge of phonologically-based classification; however, the range of predictive properties that are present in natural lexicons, and those that can be internalized by speakers, are topics of ongoing research. This talk introduces a model that learns phonological predictors from positive examples and that serves as a common inductive framework within which theories that vary in restrictiveness can be formalized and evaluated. Across a range of cases drawn from the literature, the model induces compact and interpretable grammars that rival hand-crafted alternatives in accounting for native speaker judgments.

 terry.regier@berkeley.edu, 510-642-2757