Memory-Augmented Dialogue State Tracker in Task-Oriented Dialogue System

Abstract

Dialogue state tracker is a core component of task-oriented dialogue system, which tracks users' goals during interaction between users and systems. Though many models have been applied to task-oriented dialogue systems and made some progress, these models still have poor performance in multi-domain conversation. To improve memory ability of dialogue state tracker, we propose Mem-DST (Memory-augmented dialogue state tracker), which is based on memory networks. Experiments on Multiwoz show that our model perform well in multi-domain dialogue, as well as in single domain dialogue. Besides, it is an important progress that our model gets considerable joint accuracy in both scenes.