Enhancing Emotional Support Conversation via Strategy Planner and Chain-of-Thought

The goal of Emotional Support Conversations (ESC) is to provide comfort and advice to people experiencing emotional distress through daily conversations, and it has gained significant attention in recent years. With the advancement of Large Language Models (LLMs), utilizing LLMs for emotional support conversations has become common. However, despite their excellent performance in language generation, recent research indicates that LLMs may have inherent biases towards certain emotional support strategies and tend to offer generic empathy and advice, which undermines the quality of emotional support. To address this issue, we propose a Strategy-planning Chain-of-Thought method (SplanCoT), using Pre-trained Language Models (PLMs) to plan support strategies for LLMs. Additionally, we construct Chain-of-Thought (CoT) according to the general process of ESC, breaking down the complex ESC task into multiple sub-steps and introduce strategy information. Experimental results on the ESConv dataset show that SplanCoT significantly outperforms previous state-of-the-art methods.