Partner Personas Generation for Dialogue Response Generation

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5200 - 5212 July 10-15, 2022 ©2022 Association for Computational Linguistics Partner Personas Generation for Dialogue Response Generation Hongyuan Lu, Wai Lam, Hong Cheng, Helen M. Meng The Chinese University of Hong Kong {hylu,wlam,hcheng,hmmeng}@se.cuhk.edu.hk Abstract Incorporating personas information allows di- verse and engaging responses in dialogue re- sponse generation. Unfortunately, prior works have primarily focused on self personas and have overlooked the value of partner personas. Moreover, in practical applications, the avail- ability of the gold partner personas is often not the case. This paper attempts to tackle these issues by offering a novel framework that lever- ages automatic partner personas generation to enhance the succeeding dialogue response gen- eration. Our framework employs reinforcement learning with a dedicatedly designed critic net- work for reward judgement. Experimental re- sults from automatic and human evaluations indicate that our framework is capable of gen- erating relevant, interesting, coherent and infor- mative partner personas, even compared to the ground truth partner personas. This enhances the succeeding dialogue response generation, which surpasses our competitive baselines that condition on the ground truth partner personas. 1 Introduction Building informative and engaging dialogue agents (Zhang et al., 2020; Roller et al., 2021) is a pop- ular research direction within the area of natural language processing. For the sake of engagement, diverse and consistent responses (Song et al., 2020, 2021) are important factors, and personas informa- tion (Zhang et al., 2018) gives rise to both. There are two types of personas, namely self persona and partner persona. The former refers to a self profile consisting of several sentences represent- ing the dialogue agents. Such a persona allows producing consistent responses rather than solely relying on the personas that are randomly learned and embedded in the model parameters (Kim et al., 2020). The latter refers to a profile that represents the users. Leveraging such partner personas has been empirically shown to be helpful for dialogue response selection (Gu et al., 2021). Unfortunately, the existence of partner personas suffers from the cold start (Schein et al., 2002; Zhang et al., 2014; Li et al., 2021) at the begin- ning of the conversation. Most of the works, if not all, (Li et al., 2016b; Mazaré et al., 2018; Song et al., 2019; Gu et al., 2019; Zhao et al., 2019; Madotto et al., 2019; Liu et al., 2020; Majumder et al., 2020; Wu et al., 2020a; Song et al., 2020) have been either overlooking partner personas or simply focusing on the impractical situation where partner personas guarantee to exist. In contrast, our work does not suffer from the practical issue when partner personas are missing during inference, and our proposed framework surpasses the baseline that conditions on the ground truth partner personas. To our knowledge, this is the first attempt to for- mulate partner personas generation for improved performance on the downstream dialogue response generation. Our work is motivated by the under- …

Content truncated to 500 words