Comparison of Bayesian Discriminative and Generative Models for Dialogue State Tracking