Empathetic Dialog Generation with Fine-Grained Intents

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 133–147 November 10–11, 2021. ©2021 Association for Computational Linguistics 133 Empathetic Dialog Generation with Fine-Grained Intents Yubo Xie and Pearl Pu School of Computer and Communication Sciences École Polytechnique Fédérale de Lausanne, Switzerland {yubo.xie, pearl.pu}@epfl.ch Abstract Empathetic dialog generation aims at generat- ing coherent responses following previous di- alog turns and, more importantly, showing a sense of caring and a desire to help. Existing models either rely on pre-defined emotion la- bels to guide the response generation, or use deterministic rules to decide the emotion of the response. With the advent of advanced language models, it is possible to learn sub- tle interactions directly from the dataset, pro- viding that the emotion categories offer suf- ficient nuances and other non-emotional but emotional regulating intents are included. In this paper, we describe how to incorporate a taxonomy of 32 emotion categories and 8 ad- ditional emotion regulating intents to succeed the task of empathetic response generation. To facilitate the training, we also curated a large- scale emotional dialog dataset from movie sub- titles. Through a carefully designed crowd- sourcing experiment, we evaluated and demon- strated how our model produces more empa- thetic dialogs compared with its baselines. 1 Introduction Empathy is considered to be an innate ability of hu- man beings (Roth-Hanania et al., 2011) and plays an important role in people’s social communica- tion (Valente, 2016). It has been shown that in- tegrating empathy into dialog systems could im- prove user experience for human-computer inter- action (Liu and Picard, 2005). One of the empa- thetic components is the capacity to respond with an appropriate emotion to another person’s mental states (Shamay-Tsoory et al., 2009). In this regard, many existing neural dialog systems (Zhou et al., 2018; Huang et al., 2018; Zhou and Wang, 2018; Colombo et al., 2019; Song et al., 2019; Shen and Feng, 2020) generate emotional responses condi- tioned on a pre-specified emotion label. However, this might be impractical when deploying the chat- bots in reality, since an extra label is required as Someone cut me off in traffic! 😡 (angry) You should fight that guy! 😡 (angry) I’m so mad at him! 😡 (angry) Haha! That’s so funny! 😁 (joyful) What’s wrong with you? 😡 (angry) Did you report to the police? 🤔 (questioning) Well, probably it’s not so big a deal… 😒 (neutral) (a) (b) (c) ✗ ✗ ✓ Figure 1: Three ways of responding to a speaker’s utter- ance. Note that simply following the speaker’s emotion state (a) or reversing it (b) still leaves the speaker in angry state (or even escalates the situation). Respond- ing with questioning (c) successfully calms down the speaker and drives the conversation to a more manage- able direction. input. Other neural dialog systems (Asghar et al., 2018; Li and Sun, 2018; Zhong et al., 2019; Lin et al., 2019; Li et al., 2020) adopt manually de- fined rules, either explicitly or implicitly, to de- cide the emotion …

Content truncated to 500 words