Knowledge Graph Driven Dialogue Management for Task-oriented Dialogue

Task-oriented dialogue systems have received much attention nowadays. As a core part of the dialogue system, dialogue management often uses a set of predefined slots to represent the dialogue state and decides what to reply to users. However, due to the flat and independent structure between different slots, it is difficult for the system to handle complex dialogue tasks. Besides, the slots are usually coded by human expert for specific task domain, which is hard for domain migration. This paper implements a knowledge graph-driven dialog management method, which introduces the knowledge graph to represent the dialogue state instead of slots. In order to solve the problem of excessive dialogue state space, we investigate a hierarchical entangled state tracking method, which can handle complex conversation and is easy for domain migration. Finally, the effectiveness of the proposed method is verified on two different dialogue tasks of electricity account application and electricity purchasing.