Generating Topical and Emotional Responses Using Topic Attention

Zhanzhao Zhou, Maofu Liu, Zhenlian Zhang, Yang Fu, Junyi Xiang

Abstract: As an indispensable influencing factor of human-computer interaction experience, emotional cognitive behaviors in dialogues have aroused spread concern of researchers. However, existing emotional dialogue generation models tend to generate generic and universal responses. To address this problem, this paper proposes a topical and emotional chatting machine (TECM) that generates not only high-quality but also emotional responses. TECM utilizes the information obtained by the topic model as a prior knowledge to guide the generation of the responses, and the topic information is used as the input of the topic attention mechanism to improve the quality of responses. TECM also adopts a method of emotion category embedding to generate emotional responses. The empirical study on automatic evaluation metrics shows that TECM can generate diverse, informative and emotional responses.