Emp-RFT: Empathetic Response Generation via Recognizing Feature Transitions between Utterances

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4118 - 4128 July 10-15, 2022 ©2022 Association for Computational Linguistics Emp-RFT: Empathetic Response Generation via Recognizing Feature Transitions between Utterances Wongyu Kim1, Youbin Ahn2, Donghyun Kim2, and Kyong-Ho Lee2 1Department of Artificial Intelligence, 2Department of Computer Science Yonsei University, Seoul, Republic of Korea {rladnjsrb9999, ybahn, dhkim92, khlee98}@yonsei.ac.kr Abstract Each utterance in multi-turn empathetic dia- logues has features such as emotion, keywords, and utterance-level meaning. Feature transi- tions between utterances occur naturally. How- ever, existing approaches fail to perceive the transitions because they extract features for the context at the coarse-grained level. To solve the above issue, we propose a novel approach of recognizing feature transitions between ut- terances, which helps understand the dialogue flow and better grasp the features of utterance that needs attention. Also, we introduce a re- sponse generation strategy to help focus on emotion and keywords related to appropriate features when generating responses. Experi- mental results show that our approach outper- forms baselines and especially, achieves signif- icant improvements on multi-turn dialogues. 1 Introduction Humans have empathy which is the ability to under- stand situations others have experienced and emo- tions they have felt from the situations (Eisenberg and Strayer, 1987). That ability also enables to in- terest and console others while sharing a conversa- tion. Thus, empathetic response generation task has been considered noteworthy. Figure 1 shows an ex- ample of a multi-turn empathetic dialogue dataset, EmpatheticDialogues (Rashkin et al., 2019) con- structed to solve the task. A speaker talks about one of 32 emotion labels and a situation related to the emotion label, and a listener empathizes, responding to the speaker. Existing approaches (Rashkin et al., 2019; Lin et al., 2019; Majumder et al., 2020; Li et al., 2020; Kim et al., 2021) for the task achieve promising results but show limitations when dialogues become long because they extract features from the concatenation of all tokens in the context at the coarse-grained level. However, at the fine-grained level, each utter- ance in multi-turn empathetic dialogues has fea- tures such as emotion, keywords that each denote My sister came to visit from out of state. It was first time we have seen her in nearly in decade! That must have been exciting to see her after so long. Dose she live very far away? Yes, she lives on the other side of the country. We had a nice dinner, but my boyfriend is a loud chewer and a sloppy eater. It was quite embarrassing! That is good, I am glad you got her back. That is a shame. Does your sister have a bad impression of your boyfriend having dinner with you? Speaker Listener Speaker Response A (Gold) Response B Figure 1: An example of EmpatheticDialogues with response A and B. Response B is from one of state-of- the-art models. Highlighted words are keywords. what an interlocutor feels and primarily says, and utterance-level meaning that …

Content truncated to 500 words