Comparison of Bayesian Discriminative and Generative Models for Dialogue State Tracking

摘要

In this paper, we describe two dialogue state tracking models competing in the 2012 Dialogue State Tracking Challenge (DSTC). First, we detail a novel discriminative dialogue state tracker which directly estimates slot-level beliefs using deterministic state transition probability distribution. Second, we present a generative model employing a simple dependency structure to achieve fast inference. The models are evaluated on the DSTC data, and both significantly outperform the baseline DSTC tracker.