For robust spoken conversational interaction, many dialog state tracking algorithms have been developed. Few studies, however, have reported the strengths and weaknesses of each method. The Dialog State Tracking Challenge (DSTC) is designed to address this issue by comparing various methods on the same domain. In this paper, we present a set of techniques that build a robust dialog state tracker with high performance: wide-coverage and well-calibrated data selection, feature-rich discriminative model design, generalization improvement techniques and unsupervised prior adaptation. The DSTC results show that the proposed method is superior to other systems on average on both the development and test datasets.
展开▼
