Stochastic Systems Group  

Discriminative linear models for natural language processing
Michael Collins
CSAIL, MIT
Recent work in machine learning approaches to natural language problems has considered discriminative methods such as loglinear (or maximumentropy) models, the perceptron algorithm, and algorithms based on support vector machines.
In this talk I will describe some recent results in this area. In particular, I'll describe a method that generalizes support vector machine methods to supervised training of Markov random fields, hidden markov models, probabilistic contextfree grammars, and other structured models. I will describe a new algorithm for solving the "largemargin" optimization problem defined in (Taskar, Guestrin and Koller 2003). The optimization method makes use of algorithms such as the forwardbackward or insideoutside algorithm, and relies on the application of exponentiated gradient updates (Kivinen and Warmuth, 1997) to quadratic programs.
Problems with this site should be emailed to jonesb@mit.edu