Persona Expansion with Commonsense Knowledge for Diverse and Consistent Response Generation

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1139–1149 May 2-6, 2023 ©2023 Association for Computational Linguistics Persona Expansion with Commonsense Knowledge for Diverse and Consistent Response Generation Donghyun Kim1 Youbin Ahn1 Wongyu Kim1 Chanhee Lee1 Kyungchan Lee1 Kyong-Ho Lee1 Jeonguk Kim2 Donghoon Shin2 Yeonsoo Lee2 1Department of Computer Science, Yonsei University 2NCSOFT {dhkim92,ybahn,rladnjsrb9999,heederer,steven800,khlee89}@yonsei.ac.kr {jukim,dhshin,yeonsoo}@ncsoft.com Abstract Generating diverse and consistent responses is the ultimate goal of a persona-based dialogue. Although many studies have been conducted, the generated responses tend to be generic and bland due to the personas’ limited descriptive- ness. Therefore, it is necessary to expand the given personas for more attractive responses. However, indiscriminate expansion of personas threaten the consistency of responses and there- fore reduce the interlocutor’s interest in con- versation. To alleviate this issue, we propose a consistent persona expansion framework that improves not only the diversity but also the con- sistency of persona-based responses. To do so, we define consistency criteria to avoid possi- ble contradictions among personas as follows: 1) Intra-Consistency and 2) Inter-Consistency. Then, we construct a silver profile dataset to deliver the ability to conform with the consis- tency criteria to the expansion model. Finally, we propose a persona expansion model with an encoder-decoder structure, which consid- ers the relatedness and consistency among per- sonas. Our experiments on the Persona-Chat dataset demonstrate the superiority of the pro- posed framework. 1 Introduction In the field of open-domain dialogues, maintain- ing long-term interaction with users by generating human-like responses has been a persistent goal. As efforts along this line of research, persona-based dialogue models aim at generating more engag- ing and consistent responses based on personal traits (Li et al., 2016b; Qian et al., 2018). As the most widely used benchmark dataset, Persona-Chat (Zhang et al., 2018) consists of personal profiles, which contain 3-5 persona sentences (hereinafter referred to as "personas"), and dialogues condi- tioned on personas. With the Persona-Chat, many researches have attempted to improve the quality of responses from various perspectives (Song et al., ��������������� ������������������� � ����������������� ����������� ���������������� ������������������ ������� �������� Figure 1: An example of persona expansion with com- monsense. The center circle is the predefined profile, and the dashed circle is the valid expandable range. 2020; Liu et al., 2020; Song et al., 2021; Cao et al., 2022). However, since the descriptiveness of per- sonas is limited, they have difficulties in converting what personas imply into attractive responses. Unlike dialogue models, humans converse be- yond the explicit persona itself through rational rea- soning with commonsense knowledge. The com- monsense reasoning ability is essential for making an interlocutor engage in conversations with a dia- logue agent. Figure 1 illustrates an example of per- sona expansion through commonsense reasoning. Given the persona ‘I am an athlete’, it is possible to infer the characteristic ‘sporty and active’ with high probability. The implications of ‘want to be the best athlete’ and ‘need to train hard’ are also predictable. In order to mimic human reasoning …

Content truncated to 500 words