The Weekly Newsletter of MIT Linguistics

Exp/Comp Group (11/4) - Huteng Dai (Rutgers)

Date/Time: Friday (11/4) from 2-3:30pm
Location: 32-D831 and on Zoom
Speaker: Huteng Dai (Rutgers)
Title: A Neo-Trubetzkoyan approach to phonotactic learning in the presence of exceptions
Abstract: Lexicalized exceptions are a major source of noise in phonological acquisition. In a positive-evidence-only setting, it is common to cope with exceptions with indirect negative evidence from distributional information (Clark & Lappin 2010). Most distribution-sensitive models assume a probabilistic grammar that evaluates the grammaticality of words by their predicted likelihood (Hayes & Wilson 2008). However, a probabilistic grammar conflates all words into the same spectrum of probability and grammaticality. As a result, short attested exceptions become more ‘grammatical’ than longer grammatical words with lower probabilities (Daland 2015). This can be problematic because it blurs the boundary between exceptions and grammatical words. In this talk, I spell out a Neo-Trubetzkoyan algorithm that learns a categorical grammar in the presence of exceptions with respect to a restrictive Subregular (Heinz 2010) hypothesis space and iterative Observed/Expected comparison. I argue that this approach is at least as good as, and appears to be superior to the “Probabilistic grammar + Probabilistic inference” approaches in handling exceptions in the case study of Turkish nonlocal vowel phonotactics.