Discriminative dialog state tracking has become a hot topic in dialog research community recently. Compared to generative approach, it has the advantage of being able to handle arbitrary dependent features, which is very appealing. In this paper, we present our approach to the DSTC2 challenge. We propose to use discriminative Markovian models as a natural enhancement to the stationary discriminative models. The Markovian structure allows the incorporation of 'transitional' features, which can lead to more efficiency and flexibility in tracking user goal changes. Results on the DSTC2 dataset show considerable improvements over the baseline, and the effects of the Markovian dependency is tested empirically.
展开▼
