Cause-Aware Empathetic Response Generation via Chain-of-Thought Fine-Tuning
Xinhao Chen, Chong Yang, Man Lan, Li Cai, Yang Chen, Tu Hu, Xinlin Zhuang, Aimin Zhou
TL;DR
This work tackles empathetic response generation by addressing the missing link of emotion-cause reasoning and listener-role awareness. It introduces Cause-Aware CoT Fine-Tuning Generation (CFEG), which combines cause-aware Chain-of-Thought prompts, emotion-cause knowledge from COMET directed by ECPE spans, and instruction tuning to stabilize reasoning on a modest LLM (LLaMA-7b) with LoRA. The approach achieves state-of-the-art results on EmpatheticDialogues in both automatic and human evaluations, outperforming strong baselines including ChatGPT variants, while emphasizing the importance of cause-oriented knowledge and listener perspective. The findings highlight the practical potential of structured CoT prompts and ECPE-informed knowledge to improve empathy, coherence, and informativeness in conversational agents, with implications for safer, more explainable empathetic AI systems.
Abstract
Empathetic response generation endows agents with the capability to comprehend dialogue contexts and react to expressed emotions. Previous works predominantly focus on leveraging the speaker's emotional labels, but ignore the importance of emotion cause reasoning in empathetic response generation, which hinders the model's capacity for further affective understanding and cognitive inference. In this paper, we propose a cause-aware empathetic generation approach by integrating emotions and causes through a well-designed Chain-of-Thought (CoT) prompt on Large Language Models (LLMs). Our approach can greatly promote LLMs' performance of empathy by instruction tuning and enhancing the role awareness of an empathetic listener in the prompt. Additionally, we propose to incorporate cause-oriented external knowledge from COMET into the prompt, which improves the diversity of generation and alleviates conflicts between internal and external knowledge at the same time. Experimental results on the benchmark dataset demonstrate that our approach on LLaMA-7b achieves state-of-the-art performance in both automatic and human evaluations.