Jesse Harris defends dissertation

Jesse Harris will defend his dissertation, "Processing Commitments" at 1PM in Dickinson 212 on Monday, November 28. All are welcome.

Congratulations Jesse!

Michael Clauss talks at the LARC/Acquisition Lab meeting

Michael Clauss will give "Modal morphology in Child Tamil," at the LARC/Acquisition Lab meeting on Monday, November 28, at 5:15 in the Partee Room (South College 301).

All are welcome!

Discussion of Wilson and Obdeyn (2009)

Joe Pater writes:

Brian Dillon and I have organized a group discussion of Wilson and Obdeyn's (2009) "Simplifying subsidiary theory: statistical evidence from Arabic, Muna, Shona, and Wargamay." 

The meeting will take place in the Partee room (South College 301) at 2:30 pm November 30th. I've put a copy of the difficult-to-find paper here:

http://people.umass.edu/hgr/WilsonObdeynSeptember2009.pdf

All are welcome!

Colin Wilson gives department colloquium

Colin Wilson, of Johns Hopkins, will give the following talk at the department colloquium on Friday, December 2 at 3:30 in Machmer E-37.

Bayesian inference for constraint-based phonology

Bayesian mathematics and associated algorithms provide a general solution to the problem of inferring structure from incomplete, ambiguous, and noisy data. In this talk, I apply these methods to the specific problem of learning constraint-based grammars of phonology, focusing in particular on the relative roles of the likelihood -- which depends on the language-specific sound pattern -- and the prior -- which embodies assumptions made by the learner independently of the data. Previous research has proposed a rich set of prior assumptions (e.g., that inputs are identical to outputs in early phonological learning, and that certain classes of constraints are biased to be higher-ranked), which are (approximately) enforced by an increasingly complex battery of learning mechanisms. I argue that the prior can be greatly simplified, perhaps even made completely unbiased, by embracing the learner's uncertainty about the inputs and weightings/rankings that underly the observable data. Formal analysis of an unbiased learner, together with simulations from an implementation that uses Gibbs sampling, suggest that a rich prior is not needed to ensure 'restrictiveness' or other empirically-motivated properties of phonological learning.