Efficient Dialogue State Tracking based on Dialogue-level State Prediction and Historical Dialogue State Deletion

The dialogue state tracking module is a crucial component of task-oriented dialogue systems. Recently, some Dialogue State Tracking (DST) methods have used the previous dialogue state as auxiliary input, resulting in errors that propagate and subsequently affect predictions. This paper proposes utilizing dialogue-level state as the prediction target and randomly removing historical dialogue state during training. The experiments demonstrate that this approach can effectively enhance the performance of the DST algorithm, alleviate error propagation, and achieve competitive results on both noisy (MultiWOZ 2.1) and clean (MultiWOZ 2.4) datasets.