CoMAE: A Multi-factor Hierarchical Framework for Empathetic Response Generation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 813–824 August 1–6, 2021. ©2021 Association for Computational Linguistics 813 CoMAE: A Multi-factor Hierarchical Framework for Empathetic Response Generation Chujie Zheng†, Yong Liu‡, Wei Chen‡, Yongcai Leng‡, Minlie Huang†∗ †The CoAI group, DCST, Institute for Artificial Intelligence, †State Key Lab of Intelligent Technology and Systems, †Beijing National Research Center for Information Science and Technology, †Tsinghua University, Beijing 100084, China ‡Sogou Inc., Beijing, China chujiezhengchn@gmail.com, aihuang@tsinghua.edu.cn Abstract The capacity of empathy is crucial to the suc- cess of open-domain dialog systems. Due to its nature of multi-dimensionality, there are various factors that relate to empathy expres- sion, such as communication mechanism, dia- log act and emotion. However, existing meth- ods for empathetic response generation usu- ally either consider only one empathy fac- tor or ignore the hierarchical relationships be- tween different factors, leading to a weak abil- ity of empathy modeling. In this paper, we propose a multi-factor hierarchical framework, CoMAE, for empathetic response generation, which models the above three key factors of empathy expression in a hierarchical way. We show experimentally that our CoMAE- based model can generate more empathetic re- sponses than previous methods. We also high- light the importance of hierarchical modeling of different factors through both the empirical analysis on a real-life corpus and the exten- sive experiments. Our codes and used data are available at https://github.com/ chujiezheng/CoMAE. 1 Introduction Empathy, which refers to the capacity to under- stand or feel what another person is experiencing (Rothschild, 2006; Read, 2019), is a critical capa- bility to open-domain dialog systems (Zhou et al., 2018b). As shown in previous research, empathetic conversational models can improve user satisfac- tion and receive more positive feedback in numer- ous domains (Klein, 1998; Liu and Picard, 2005; Brave et al., 2005; Fitzpatrick et al., 2017; Liu et al., 2021). Recently, there have also been numer- ous works devoted to improving the dialog models’ ability to understand the feelings of interlocutors (Rashkin et al., 2019; Lin et al., 2019; Majumder ∗Corresponding author. context response DA EM CM DA EM Figure 1: Our proposed hierarchical framework: Co- MAE (right). The directed arrows denote dependencies. We also present the framework (left) of EmpTransfo (Zandie and Mahoor, 2020) for comparison. et al., 2020), which makes the dialog models more empathetic to a certain extent. However, empathy is a multi-dimensional con- struct (Davis et al., 1980) rather than merely recog- nizing the interlocutor’s emotion (Lin et al., 2019) or emotional responding (Zhou et al., 2018a). It consists of two broad aspects related to cognition and affection (Omdahl, 2014; Paiva et al., 2017). The cognitive aspect requires understanding and interpreting the situation of the interlocutor (El- liott et al., 2018), which is reflected in the dialog act taken in the conversation (De Vignemont and Singer, 2006), such as questioning (e.g., What’s wrong with it?), consoling (e.g., You’ll get through this), etc. The affective aspect relates to properly expressing emotion in reaction to the experiences and feelings shared by the …

Content truncated to 500 words