综合信息系统研究中心知识库

Dialogue state tracking (DST), as an essential component of task-oriented dialogue systems, refers to keeping track of the user's intentions as a conversation progresses. Typical methods formulate it as a classification task with fixed pre-defined slot-value pairs, or generate slot-value candidates given the dialogue history. Most of them have limitations on considering interactions of slots with utterance sentences and other slots progressively. To tackle this problem, we propose a Dialogue State Tracker with Hierarchical Temporal Slot Interactions (DST-HTSI) to capture slot-related semantic information from utterance sentences and slots. It firstly captures interactive information among slots within a turn and across turns by applying hierarchical slot interactions. Then a temporal slot interaction module is employed to establish slot dependencies along the time. Finally, a GRU is applied as the decoder to generate values for each slot correspondingly. Furthermore, we also leverage pre-trained language models as the backbone of our model. Experiments show that DST-HTSI outperforms previous state-of-the-art on MultiWOZ 2.2 and WOZ 2.0, and achieves competitive results on MultiWOZ 2.1.