Towards Persona-Based Empathetic Conversational Models

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6556–6566, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6556 Towards Persona-Based Empathetic Conversational Models Peixiang Zhong1,2,3, Chen Zhang4, Hao Wang4, Yong Liu1,3, Chunyan Miao1,2,3∗ 1Alibaba-NTU Singapore Joint Research Institute 2School of Computer Science and Engineering 3Joint NTU-UBC Research Centre of Excellence in Active Living for the Elderly Nanyang Technological University, Singapore 4Alibaba Group, China peixiang001@e.ntu.edu.sg, zhangchen010295@163.com, cashenry@126.com, stephenliu@ntu.edu.sg, ascymiao@ntu.edu.sg Abstract Empathetic conversational models have been shown to improve user satisfaction and task outcomes in numerous domains. In Psychol- ogy, persona has been shown to be highly cor- related to personality, which in turn influences empathy. In addition, our empirical analysis also suggests that persona plays an important role in empathetic conversations. To this end, we propose a new task towards persona-based empathetic conversations and present the first empirical study on the impact of persona on empathetic responding. Specifically, we first present a novel large-scale multi-domain dataset for persona-based empathetic conversa- tions. We then propose CoBERT, an efficient BERT-based response selection model that ob- tains the state-of-the-art performance on our dataset. Finally, we conduct extensive exper- iments to investigate the impact of persona on empathetic responding. Notably, our re- sults show that persona improves empathetic responding more when CoBERT is trained on empathetic conversations than non-empathetic ones, establishing an empirical link between persona and empathy in human conversations. 1 Introduction Empathy, specifically affective empathy, refers to the capacity to respond with an appropriate emo- tion to another’s mental states (Rogers et al., 2007). In NLP, empathetic conversational models have been shown to improve user satisfaction and task outcomes in numerous domains (Klein, 1998; Liu and Picard, 2005; Wright and McCarthy, 2008; Fitzpatrick et al., 2017; Zhou et al., 2018a). For example, empathetic agents received more posi- tive user ratings, including greater likeability and trustworthiness than controls (Brave et al., 2005). In recent years, neural network based conversa- tional models (Vinyals and Le, 2015; Lowe et al., ∗Corresponding author caring content jealous impressed excited furious prepared annoyed devastated terrified angry confident disgusted anxious grateful faithful ashamed proud afraid disappointed hopeful apprehensive sad lonely joyful embarrassed guilty anticipating surprised trusting nostalgic sentimental 0.0 0.1 0.2 0.3 0.4 0.5 0.6 TF-IDF Similarity between Responses from Two Different Persons TF-IDF Similarity between Two Random Disjoint Sets of Responses Figure 1: TF-IDF similarity between two sets of empa- thetic responses (Rashkin et al., 2019) for each emotion (best viewed in color). For most emotions (28 out of 32), the similarity between responses from two differ- ent speakers (blue) is substantially smaller than the sim- ilarity between two random disjoint sets of responses (orange, averaged over five runs). 2015) are becoming dominant. Zhou et al. (2018a) designed XiaoIce, a popular AI companion with an emotional connection to satisfy the human need for communication, affection, and social belonging. Recently, Rashkin et al. (2019) presented a new dataset and benchmark towards empathetic con- versations and found that both Transformer-based generative models (Vaswani et al., 2017) …

Content truncated to 500 words