Speakers: Richard Futrell and Tim O’Donnell
Title: A Tier-Based Probabilistic Phonotactics Model
Date/Time: Monday, May 5, 5:30pm
Location: 32-D831
We present work in progress on a probabilistic generative model of English phonotactics. Augmenting an underlying feature-based N-gram model, we implement a tier-based representation of the kind studied in autosegmental phonology (e.g. Goldsmith, 1976) that allows nonlocal interactions of certain features as in vowel harmony. Local and nonlocal interactions are controlled via a feature geometry embedded in the model. To evaluate our model, we used Mechanical Turk to gather a large dataset of wellformedness ratings for 1000 monosyllabic nonce words. Our generative tier-based model achieves a higher correlation with these human ratings than BLICK. We also test our model on data from Daland et al. (2011), which tests the ability to explain sonority effects, and get performance comparable to the state of the art.