NORTHWESTERN UNIVERSITY SIMILARITY AND FREQUENCY IN PHONOLOGY A DISSERTATION SUBMITTED TO THE GRADUATE SCHOOL IN PARTIAL FULFILLMENT OF THE REQUIREMENTS for the degree DOCTOR OF PHILOSOPHY Field of Linguistics By

Similarity and Frequency in Phonology Stefan Frisch This thesis focuses upon parallels between phonology and phonological processing. I study phonological speech errors and a phonotactic dissimilarity constraint, demonstrating they have analogous similarity and frequency effects. In addition, I show that abstract phonological constraints are influenced by the phonological encoding of lexical items. The results of this thesis are based on a metric of similarity computed using the representations of STRUCTURED SPECIFICATION (Broe 1993). This metric is quantitatively superior to traditional metrics of similarity which are based on feature counting. I also employ a probabilistic model of a gradient linguistic constraint which is based on categorical perception. In this model, the acceptability of a form is gradient, and acceptability is correlated with frequency. The most acceptable forms in a language are the most frequent ones. This constraint model provides a better fit to gradient phonotactic data than traditional categorical linguistic constraints. Together, the similarity metric and gradient constraint model demonstrate that statistical patterns in language can be relevant, principled, and formally modeled in linguistic theory. Using the gradient constraint model, I show that similarity effects in phonotactics are stronger word initially than later in the word. A parallel pattern is experimentally demonstrated for speech errors. I claim that the effect for speech errors follows from the fact that production of segmental material in a lexical item is inherently temporal. I argue that segmental information in lexical representations is sequentially accessed even for abstract phonological purposes, like phonotactics. The effects of word position on similarity in both speech production and phonotactics are accounted for in a connectionist model of lexical access, which does not differentiate the storage of a representation from its use. Structured specification is incompatible with UNDERSPECIFICATION (Kiparsky 1982, Archangeli 1984). In underspecification, features are left blank in a linguistic representation to capture redundancy relationships and phonological markedness. I demonstrate that models of similarity in phonotactics and speech errors which use underspecification do not model the data as well as the similarity metric based on structured specification. iii ACKNOWLEDGMENTS