A Structured Semantic Reinforcement method for Task-Oriented Dialogue

Anonymous

Abstract: Recently, many BERT based approachs have been proposed for task-oriented dialogue (TOD) task. Despite their impressive performance, the insufficient utilization of deep semantic information and long-distance context understanding makes it difficult for these methods to digest complex dialogue scenarios for they cannot obtain sufficient evidence from dialogue data to support dialogue decision-making.In this work, we propose a novel structured semantics reinforcement (SSR) method to handle these issues.SSR reorganized the end-to-end TOD structure, which mainly includes two key components: 1. The dialogue symbolic memory, which cache the objects mentioned in the dialogue and the structure under the semantic relationship. 2. semantic projection module, understanding module, based on the previous structured results, determines the source of the slot extraction required for the current task.And our approach achieves state-of-the-art results on dataset MultiWOZ 2.1, where we acquire a joint goal accuracy beyond 60\% and also gains a significant effect on dataset DSTC8.