Whamit!

The Weekly Newsletter of MIT Linguistics

Colloquium 9/11 - Maria Gouskova (NYU)

Speaker: Maria Gouskova (NYU)
Title: A Computational Learner for Complex Segment Representations
Time: Friday, September 11th, 3:30pm – 5pm

Zoom Link: (Please email ling-coll-org@mit.edu for more information)

Abstract: Phonological analysis often entails decisions about sequences such as [ts]: is it two consonants or a complex segment? Arguments for complex segments range from phonotactics to inventory structure, typology, and etymology. But while typology and etymology are accessible to linguists, they are not accessible to language learners. Phonotactics and inventories also do not always offer clear guidance. How, then, do learners discover complex segments? I describe a learning model based on lexical statistics. The model starts with a lexicon and simplex segment representations only. For any CC sequence, the model calculates inseparability: the likelihood of occurring together vs. separately. High inseparability is a property of complex segments in a range of languages. After showing a few cases, I consider alternatives: learning from natural classes, phonotactics, and phonetics. I also discuss evidence from several languages that the right distributions are in morphemes, not in phonological words or in connected speech, which has implications for the acquisition timeline.