Response Generation with Personal Attributes and Act Information
Haitao Gui, Zhongqing Wang · 2024
Despite the remarkable success of personalized response generation models, there is a lack of effective research on how to make dialogue systems more personalized by integrating dialogue act data and personal attributes. In this study, we propose a dialogue acts and personal attributes-based self-supervised model for response generation. To carry out the task, we first add personal attributes and dialogue acts to the sitcom dataset. Then, we employ a self-supervised training technique to give a pre-trained language generation model with knowledge of the unique structural features of dialogue texts, with several mask strategies to merge historical dialogues, personal attributes, and dialogue acts. After that, we construct a personalized template representing personal attributes and dialogue acts. Finally, we take historical dialogues and filled templates as inputs to generate personalized responses with the self-supervised generation model. The experimental results indicate that our response generation model outperforms other strong baselines on several metrics, demonstrating the effectiveness of personal attributes, acts, and our personalized response generation model.