Abstract: In view of the existing multi-party human-computer dialogue system, it is easy to ignore the speakers who may be forgotten in the dialogue process, resulting in weak user interaction initiative and poor user dialogue experience. This paper proposes a multi-party human-computer active dialogue strategy based on knowledge enhancement. By using the knowledge graph as external knowledge, the strategy learns the preferences of specific individuals who may be forgotten in the dialogue, and designs a personalized response strategy combining with the current group dialogue needs to actively promote the participation of individuals and groups in the dialogue. Firstly, the graph attention mechanism is used to learn the representation of specific individuals’ interest entities, and the deep preference representation of individuals is obtained by introducing time weight aggregation. Then, multiple neighbor sets are triggered along the path of the current dialogue topic knowledge subgraph to actively capture the high-level personalized interest topics of individuals. Finally, by analyzing the current context and needs of group dialogue at the semantic level, the group's satisfaction with the candidate topics is assessed, and the optimal dialogue topics that can meet individual preferences and take care of group feelings are obtained. The experimental results show that the multi-party human-computer dialogue system that integrates external knowledge and focuses on specific speakers can effectively improve the content richness of response and participant interaction satisfaction, and promote the continuous multi-party dialogue.

Key words: Human-computer interaction, Multi-party interaction, Dialogue strategy, Knowledge graph