View Seminar Video
Abstract
The first part of this talk explains why non-local context-sensitive interactions cause standard techniques for learning probabilistic grammars to be inconsistent, and motivates the more general log-linear (aka MaxEnt) models used here. The second part of the talk introduces pseudo-likelihood, and describes experiments learning stochastic unification-based grammars from corpora. The third part of the talk discusses implications of this work and the relationship between these models and optimality theory.