Effectively Incorporating Knowledge in Open-Domain Dialogue Generation

Wenji Zhou, Tingting He, Miao Zhang, Rui Fan

Abstract: Dialogue generation is one of the most important parts in the dialogue system. Generating useful and informative responses in conversation has become a research hotspot. Previous work has proved that incorporating external knowledge is conducive to generating meaningful responses. But how to make full use of the existing information to select the most appropriate knowledge is a current research difficulty. In this paper, we propose a dialogue generation model with a lightweight knowledge routing module to sample knowledge needed for the conversation. In the knowledge routing module, we not only use the interactive information between the knowledge and the dialogue context utterance, but also the interactive information between each piece of knowledge for knowledge selection. Subsequently, the selected knowledge is incorporated into the dialogue generation model to generate responses. The experimental results show that our model tends to generate more meaningful and informative responses compared with baseline models.