EmpHi: Generating Empathetic Responses with Human-like Intents

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1063 - 1074 July 10-15, 2022 ©2022 Association for Computational Linguistics EmpHi: Generating Empathetic Responses with Human-like Intents Mao Yan Chen∗and Siheng Li∗and YujiuYang† Shenzhen International Graduate School, Tsinghua University {chenmaoy19, lisiheng21}@mails.tsinghua.edu.cn yang.yujiu@sz.tsinghua.edu.cn Abstract In empathetic conversations, humans express their empathy to others with empathetic intents. However, most existing empathetic conversa- tional methods suffer from a lack of empathetic intents, which leads to monotonous empathy. To address the bias of the empathetic intents distribution between empathetic dialogue mod- els and humans, we propose a novel model to generate empathetic responses with human- consistent empathetic intents, EmpHi for short. Precisely, EmpHi learns the distribution of po- tential empathetic intents with a discrete la- tent variable, then combines both implicit and explicit intent representation to generate re- sponses with various empathetic intents. Exper- iments show that EmpHi outperforms state-of- the-art models in terms of empathy, relevance, and diversity on both automatic and human evaluation. Moreover, the case studies demon- strate the high interpretability and outstanding performance of our model. Our code are avali- able at https://github.com/mattc95/EmpHi. 1 Introduction Empathy is a basic yet essential human ability in our daily life. It is a capacity to show one’s car- ing and understanding to others. Many types of research have been conducted on empathetic ex- pression to enhance the empathy ability of Arti- ficial Intelligence, e.g., computational approach for empathy measurement (Sharma et al., 2020), empathetic expression understanding in newswire (Buechel et al., 2018), online mental health support (Sharma et al., 2021), etc. In this work, we fo- cus on the task of generating empathetic responses (Rashkin et al., 2019; Lin et al., 2019; Majumder et al., 2020) in open-domain conversation. Existing empathetic dialogue models pay more attention to the emotion-dependent response gen- eration (Lin et al., 2019; Majumder et al., 2020). ∗ Equal contribution. † Corresponding author(yang.yujiu@sz.tsinghua.edu.cn). I just started college again, and while I'm doing great in school, it has lead me to feel very lonely with a lack of social life. I am sorry to hear that! I know I broke up with my ex, but I can't help but feel irritated when he talks about going on dates. A while back my cat knocked over and broke my mother’s urn. Oh my goodness, I have a cat, I know how you feel. I am sorry to hear that . I hope you find someone to help you. That is so annoying. I would be upset too. Acknowledging Sympathizing Agreeing EmpHi Output Figure 1: EmpHi generates empathetic responses with human-like empathetic intents (text in blue box), while existing empathetic dialogue models generate responses with contextually irrelevant and monotonous empathy (text in orange box). However, using emotion alone to generate re- sponses is coarse-grained, and the model needs to incorporate empathetic intent information. On the one hand, the speaker often talks with a particu- lar emotion while the listener shows their empathy …

Content truncated to 500 words