MoEL: Mixture of Empathetic Listeners

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 121–132, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 121 MoEL: Mixture of Empathetic Listeners Zhaojiang Lin, Andrea Madotto, Jamin Shin, Peng Xu, Pascale Fung Center for Artificial Intelligence Research (CAiRE) Department of Electronic and Computer Engineering The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong {zlinao,amadotto,jmshinaa,pxuab}@connect.ust.hk, pascale@ece.ust.hk Abstract Previous research on empathetic dialogue sys- tems has mostly focused on generating re- sponses given certain emotions. However, being empathetic not only requires the abil- ity of generating emotional responses, but more importantly, requires the understanding of user emotions and replying appropriately. In this paper, we propose a novel end-to- end approach for modeling empathy in dia- logue systems: Mixture of Empathetic Lis- teners (MoEL). Our model first captures the user emotions and outputs an emotion distri- bution. Based on this, MoEL will softly com- bine the output states of the appropriate Lis- tener(s), which are each optimized to react to certain emotions, and generate an empathetic response. Human evaluations on empathetic- dialogues (Rashkin et al., 2018) dataset con- firm that MoEL outperforms multitask training baseline in terms of empathy, relevance, and fluency. Furthermore, the case study on gen- erated responses of different Listeners shows high interpretability of our model. 1 Introduction Neural network approaches for conversation mod- els have shown to be successful in scalable training and generating fluent and relevant re- sponses (Vinyals and Le, 2015). However, it has been pointed out by Li et al. (2016a,b,c); Wu et al. (2018b) that only using Maximum Likelihood Es- timation as the objective function tends to lead to generic and repetitive responses like “I am sorry”. Furthermore, many others have shown that the in- corporation of additional inductive bias leads to a more engaging chatbot, such as understanding commonsense (Dinan et al., 2018), or modeling consistent persona (Li et al., 2016b; Zhang et al., 2018a; Mazare et al., 2018a). Meanwhile, another important aspect of an en- gaging human conversation that received rela- Emotion: Angry Situation I was furious when I got in my first car wreck. Speaker I was driving on the interstate and another car ran into the back of me. Listener Wow. Did you get hurt? Sounds scary. Speaker No just the airbags went off and I hit my head and got a few bruises. Listener I am always scared about those airbags! I am so glad you are ok! Table 1: One conversation from empathetic dialogue, a speaker tells the situation he(she) is facing, and a lis- tener try to understand speaker’s feeling and respond accordingly tively less focus is emotional understanding and empathy (Rashkin et al., 2018; Dinan et al., 2019; Wolf et al., 2019). Intuitively, ordinary social con- versations between two humans are often about their daily lives that revolve around happy or sad experiences. In such scenarios, people generally tend to respond in a way …

Content truncated to 500 words