Progressive Dialogue State Tracking For Multi-Domain Dialogue Systems

There are two critical observations in multi-domain dialogue state tracking (DST) that have been ignored in most existing work. First, the number of triples (domain-slot-value) in dialogue states generally increases with the growth of dialogue turns. Second, even though dialogue states are accumulating, the difference between two adjacent turns is steadily minor. To model the two observations, we propose to divide the task into two successive procedures: progressive domain-slot tracking and shrunk value prediction. Specifically, domain-slot pairs are first modeled in a multi-level structure that can be predicted progressively based on previous turns. Then, we employ a generative approach to producing dialogue values for the predicted, rather than for all possible, domain-slot pairs. This divide-and-conquer approach not only enables parallelization for predicting domain-slot pairs, but also reduces the number of domain-slot candidates significantly for value prediction. Experimental results on the MultiWOZ datasets confirm that our methodology achieves very favourable improvement over baseline models.