Kishore Papineni
IBM Research
We are interested in classification problems with a finite number of classes. A classifier is any probability distribution on the classes conditioned on the observations. We want our classifier to discriminate between the true class and the imposters. We define a measure of discrimination of an arbitrary conditional probability model on a set of labeled training data. We then consider maximizing discrimination on a parametric family of exponential models that arises naturally in the maximum entropy framework. We show that this objective function is globally convex in Rn and is piece-wise linear. We propose a solution that involves solving a series of linear programming problems. We will compare this framework to that of maximum entropy.