Schema-Guided Multi-Domain Dialogue State Tracking with Graph Attention Neural Networks

Dialog state tracking, which is one of the most crucial modules in taskoriented dialog systems, predicts the dialog states based on a dialog history. For multidomain DST, lacking dependencies reasoning across domains and slots is also a major obstacle due to increasingly complex interactive dialogues. Existing methods that extract the values from a pre-defined slot value list or generate values based on the dialog context have difficulty tracking slot values that are not explicitly mentioned in the pre-defined value list and the dialogue history. In this paper, we propose a Multi-step Dialogue State Tracker with Schema-Guided Graph Convolutional Network (MusDST) that incorporates a multi-step prediction procedure for dialogue state tracking. In this procedure, a novel transfer mechanism is proposed to track the slot values that are not pre-defined in the list or explicitly mentioned in the dialogue history. Also, we model slot relations using graph convolutional network to reason slot values over schema graphs. Experiments demonstrate that our proposed MusDST achieves state-of-the-art performance on both MultiWOZ 2.0 and MultiWOZ 2.1 datasets. Also, we show the capability of our model in tracking not explicitly mentioned slot values.