Structured Featural Representations in a Generative Model of

We present a generative model of phono-tactics, the set of well-formed phoneme sequences in a language. In order to test what formal mechanisms are useful for representing phonotactics in this set-ting, we adopt the “inductive baseline” approach of Gildea and Jurafsky (1996) and Hayes and Wilson (2008), comparing models that do and do not include partic-ular linguistic representations. We formal-ize and test three ideas from the phonol-ogy literature: (i) phonological features can be organized in a feature dependency graph; (ii) phonological restrictions make reference to subsegmental bundles of fea-tures that behave as single units, and (iii) restrictions tend to hold between corre-sponding bundles of features (tier-based interactions). We evaluate our models’ ability to capture phonotactic distributions in the lexicons of 14 languages drawn from the WOLEX corpus (Graff, 2012). Each of the three representational devices consistently improves the predictive prob-abilities assigned to held-out forms, for all languages. We also present novel information-theoretic analyses designed to probe these behaviors in more detail.