Gaja Jarosz, from Yale University, will give the department colloquium talk this Friday, April 27, at 3:30PM in Machmer E-37. You can learn more about Professor Jarosz at http://pantheon.yale.edu/~gjs42/.
The title and abstract for her talk are:
Learning Hidden Structure in OT and HG
Computational models of learning with violable constraints have led to significant progress in understanding how learners acquire the complex system of knowledge that is phonology; however, a number of significant challenges remain. This talk addresses one of the major outstanding problems, learning in the face of hidden structure, and examines the challenges this problem poses for successful and efficient learning. The particular kind of hidden structure I focus on in this talk is structural ambiguity, which includes any kind of latent structure assigned when parsing a phonological string, such as metrical feet, (sub)syllabic constituents, and autosegmental representations. I examine Robust Interpretive Parsing (Tesar and Smolensky 1998), a well-known approach to structural ambiguity in OT, and show that its extension to probabilistic constraint-based grammars as first described by Boersma (2003) is problematic. The problem occurs because the proposed formulation fails to take advantage of the rich information contained in the learner's stochastic grammar. I propose two modifications to the learning algorithm, one trivial and another more involved, and show that both lead to improvements in the success rates of Stochastic OT and noisy HG learners. I then examine another important aspect of these algorithms' performance: their efficiency. The results of this second evaluation are complex but indicate an efficiency advantage for the proposed modifications as well. I discuss the implications for the relative merits of OT and HG as well as for the evaluation of computational models of phonological learning more generally.
