Enhancing Emotional Support Conversation with Cognitive Chain-of-Thought Reasoning

Yaru Cao, Zhuang Chen, Guanqun Bi, Yulin Feng, Min Chen, Fucheng Wan, Minlie Huang, Hongzhi Yu

Abstract: Emotional support conversation can reduce mental stress and provide social benefits. However, the help seeker’s mental state often lies beneath the surface utterance in conversation, making it difficult to understand the seeker’s deep status and provide effective emotional support through mere dialogue modeling. To address this challenge, we propose CogChain, a cognitive chain-of-thought (CoT) reasoning framework that mimics a human supporter’s cognitive process for emotional support conversation. Specifically, CogChain employs a chained structure to analyze the seeker’s issues, infer internal thoughts, determine behavioral intentions, and select appropriate strategies to achieve support goals. We further design an in-context learning pipeline using large language models to efficiently generate CogChain for any given dialogue content. To validate the effectiveness of CogChain, we incorporate it in the frequently used ESConv dataset and accordingly train enhanced machine supporters with CoT reasoning ability. Extensive automatic and human evaluations show that, CogChain not only improves the machine supporter’s performance for in-domain seen scenarios but also enhances its generalizability to out-of-domain unseen scenarios, demonstrating the importance of cognitive reasoning for emotional support conversation (Our resources are available at https://github.com/YaruCao-AI/CogChain).