Using Dempster-Shafer Evidence Theory for Dialog State Tracking

In a task oriented spoken dialogue system, the automatic speech recognition (ASR) and spoken language understanding (SLU) modules usually provide multiple uncertain results, which may be related or divergent. Previously, researchers used classical probability theory based approach to solve this problem, but it is difficult to handle results combination and conflict. In this paper, we describe a novel target-based state tracking algorithm based on Dempster-Shafer(D-S) theory to deal with ASR and SLU uncertain issues. The n-best recognition results of different dialog turns are combined to update the current dialog state using evidence theory. Our method can be easily integrated into existing dialogue system. The effectiveness of our approach is demonstrated on a song-on-demand task, and performs better than traditional probability based methods.