Abstract: To solve the problem that the current dialogue system is mainly passive response and still unable to carry out active dialogue well, a knowledge driven human-machine active dialogue management strategy is proposed, which simulates human communication mode and divedes the dialogue into two sub-tasks: topic switching and topic depth. A personalized dialogue management strategy is designed to realize active guidance and topic transfer in multi-round dialogues. The proposed strategy determines the time of the system's active dialogue based on the emotional state of human-machine interaction, and uses the knowledge graph as the background knowledge information to actively search the multi-hop neighbor set of the dialogue entities that are triggered by the knowledge graph, so as to determine the next interaction content. For topics of users' negative emotions, new topics are actively sought for through outward communication method. For topics with users' positive emotions, the current topic can be deeply respond to through cohesion. The experimental results show that the initiative of model dialogue is improved by the strategy while balancing global dialogue coherence and local topic consistency, which is a new reference for the development of human-machine active dialogue system.

Key words: dialogue system,  knowledge-driven,  active dialogue,  affective state,  dialogue management strategy