Improved non-autoregressive dialog state tracking model

Abstract

Dialogue systems, a powerful tool of human-machine interaction, are widely applied in e-commerce, online education, and cellphone assistant, etc. Dialogue state tracking (DST), updating the state of user goals during dialogue, is a core part of task-oriented dialogue systems. Recent research has made progress in low-latency and good-performance DST neural network models, i.e., non-autoregressive dialogue state tracking model (NADST). However, there are still some rooms for improvement in dialogue state tracking. In this paper, we propose following ways to improve the efficiency of NADST: (1) adding shrinkage residual network into fertility prediction; (2) constructing residual connection between different hierarchical attentions; (3) inserting a relative position encoding into state decoder for improving the performance of state prediction. The results of analysis and experiments indicate that the proposed model is the SOTA non-autoregressive method of dialog state tracking.