Concept-based Persona Expansion for Improving Diversity of Persona-Grounded Dialogue

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3471–3481 May 2-6, 2023 ©2023 Association for Computational Linguistics Concept-based Persona Expansion for Improving Diversity of Persona-Grounded Dialogue Donghyun Kim1 Youbin Ahn1 Chanhee Lee1 Wongyu Kim1 Kyong-Ho Lee1 Donghoon Shin2 Yeonsoo Lee2 1Department of Computer Science, Yonsei University 2NCSOFT {dhkim92,ybahn,heederer,rladnjsrb9999,khlee89}@yonsei.ac.kr {dhshin,yeonsoo}@ncsoft.com Abstract A persona-grounded dialogue model aims to improve the quality of responses to promote user engagement. However, because the given personas are mostly short and limited to only a few informative words, it is challenging to utilize them to generate diverse responses. To tackle this problem, we propose a novel persona expansion framework, Concept-based Persona eXpansion (CPX). CPX takes the original per- sona as input and generates expanded personas that contain conceptually rich content. We con- stitute CPX with two task modules: 1) Concept Extractor and 2) Sentence Generator. To train these modules, we exploit the duality of two tasks with a commonsense dataset consisting of a concept set and the corresponding sentences which contain the given concepts. Extensive ex- periments on persona expansion and response generation show that our work sufficiently con- tributes to improving the quality of responses in diversity and richness. 1 Introduction A persona-grounded dialogue model aims to gener- ate more human-like and engaging responses based on given traits called persona (Zhang et al., 2018a). As efforts of this research line, many recent works have explored various approaches to improving the quality of persona-based responses (Liu et al., 2020; Song et al., 2020; Kim et al., 2020). In spite of these efforts, there remain some limi- tations in persona-grounded dialogue models. They are in need of generating more diverse responses. Due to the predefined personas being mostly short and limited to only a few informative words, the re- sponses based on these personas tend to be generic and monotonous. To tackle this issue, COMPAC (Majumder et al., 2020) expands the predefined personas with a commonsense knowledge graph about events, ATOMIC (Sap et al., 2019). With the expanded personas, the dialogue agent generates Original Persona Concept-based Expansion Event-based Expansion I like to go hunting. What are you going to do on this vacation? I am going to have fun. I am going to camp outdoor while hunting. I want to have fun. I want to go camping, hunting, and fishing. Figure 1: An example of responses generated by differ- ent persona expansion strategies for the same query. more diverse responses which contain inferential knowledge stemming from the original persona. However, there is room for improvement in the expansion strategy of COMPAC, which takes an event-based approach as means for persona expan- sion, which captures causal inferences from per- sonas. The event-based persona expansion is done with the utilization of ATOMIC as a training dataset for the COMET commonsense transformer (Bosse- lut et al., 2019). As an example, in Figure 1, the expanded persona of the agent (i.e., I want to have fun) can be obtained from the given original per- …

Content truncated to 500 words