Speaker: Giorgio Magri (CNRS and University of Paris 8)
Title: What is the proper model of probabilistic phonology?
Time: Friday, November 8th, 3:30pm – 5pm
Location: 32-155
Abstract: Phonology has traditionally focused on categorical alternations. More recently, phonology has extended its empirical coverage to quantitative data such as gradient judgements and phonologically conditioned variation. This empirical extension requires a corresponding theoretical extension from categorical to probabilistic models of phonology. Perhaps the main open question in phonological theory in the next decade is how to properly characterize the probabilistic model underlying natural language phonology. This talk makes three contributions towards addressing this question. Part I provides a derivation of MaxEnt phonology from first principles. Part II (based on joint work with Arto Anttila) argues that MaxEnt nonetheless makes no sense as a model of probabilistic phonology because it severely over-generates. If even the best is not good enough, we better try something different. Part III starts to explore the alternative strategy of defining probabilistic phonological grammars by sampling from an underlying typology of categorical grammars. This class of models (which includes stochastic HG and OT) is shown to provide tight linguistic predictions and to have learnability properties comparable to those of MaxEnt. The talk is a commercial for a forthcoming textbook on categorical and probabilistic constraint-based phonology.