Hierarchical Dialog State Tracking with Unknown Slot Values

Abstract

Dialog state tracking (DST) is the key component of goal-driven Spoken Dialog Systems. Almost all existing dialog state trackers are unable to handle unknown slot values. The continuous emergence of unknown slot values in dialogs is inevitable. If unknown slot values cannot be accurately detected and distinguished, the dialog states cannot be correctly updated in real time. This paper proposes a hierarchical dialog state tracking framework to model the dialog state tracking with unknown slot values. Three levels are included in the framework. Unknown slot values are identified at the first-level by a two-layer cascaded neural network as well as known slot values. Distributions for unknown and known slot values are updated separately in the second level and integrated in the third level. Experimental results on DSTC and WOZ2.0 datasets show that the proposed framework achieves good performance. Especially, the detection and distinction of unknown slot values greatly improve the final performance of dialog state tracking, illustrating the effectiveness of our proposed framework for addressing the problem of unknown slot values.