RSP-DST: Revisable State Prediction for Dialogue State Tracking

Task-oriented dialogue systems depend on state tracking to keep track of the intentions users in course conversations. Although recent models exhibit good performance, errors predicting value each slot at current turn these are easily carried over next turn, and unlikely be revised resulting error propagation. In this paper, we propose a revisable prediction for tracking, which constructs two-stage process composed an original revising prediction. The jointly previous context predict turn. Then, order avoid existing continuing utilizes revise errors, alleviating Experiments conducted MultiWOZ 2.0, 2.1, 2.4 results indicate that our model outperforms state-of-the-art works, achieving new performances with 56.35, 58.09, 75.65% joint goal accuracy, respectively, has significant improvement (2.15, 1.73, 2.03%) best results.