Gabe Doyle (UCSD) will be presenting at the P-Interest meeting today at noon in the Greenberg Room (460-126). There will be lunch (out) after the talk for those who are interested.
Computational models of phonological constraint acquisition
Abstract: Phonology, whether approached from a rule-based or Optimality Theory viewpoint, relies on a set of rules or constraints that shape the sound patterns of a language. But where does this set come from? The most common, sometimes unstated, solution is to treat the set as innate and language-universal. This universality has some explanatory benefits, but it is a strong assumption, and one influenced in part by the lack of viable methods for learning constraints. In this talk, I propose two computational models for markedness constraint acquisition in an Optimality Theory framework. The first uses minimal phonological structure to learn a set of constraint violations that can be used to identify probable constraints. The second uses a similar learning structure but includes a basic grammar for constraints (similar to Smith 2004’s schemas) to jointly learn violations and constraint identities. These methods, tested on Wolof vowel harmony and English plurals, learn systems of constraints that explain observed data equally well as the constraints in a standard phonological analysis, with a violation structure that largely corresponds with the standard constraints. This suggests that constraint learning may be possible as a part of acquisition, and the innateness and universality assumption may be able to be weakened. This is joint work with Klinton Bicknell (Northwestern) and Roger Levy (UCSD).