Constructing and applying a multi-turn psychological support dialogue corpus based on the Helping Skills Chain-of-Thought

With the increasing prominence of mental health issues, automated psychological support dialogue systems have gradually gained attention. However, existing Chinese corpora mostly remain at the level of single-turn Q&A or lack psychological counseling theoretical grounding, making it difficult to cover the progressive interactions common in psychological counseling. Meanwhile, collecting and releasing large-scale real multi-turn dialogues faces challenges related to privacy protection and high costs. To address this, this paper proposes the Helping Skills Chain-of-Thought (HCoT) method, which integrates Helping Skills Theory with Chain-of-Thought prompting. We utilized GPT-4o to rewrite CD-CN single-turn data into a Chinese multi-turn psychological support corpus, HCoT-Corpus. This corpus contains 22,341 dialogues and 211,473 strategy annotations, achieving a systematic expansion in scale, structural depth, and theoretical grounding. Analysis results indicate that HCoT-Corpus demonstrates high structural coherence and multi-strategy collaborative characteristics under the “Exploration-Comfort-Action” three-stage framework. Experimental evaluations show that, compared to baselines like SMILE, the HCoT method achieves the most balanced performance in emotional resonance, strategy application, and structural integrity. Furthermore, HCoT-Chat, fine-tuned on Qwen2.5-7B-Instruct, achieved significant advantages in both automatic metrics and cross-model evaluations. This study demonstrates the HCoT method as a promising path for constructing large-scale, theoretically grounded psychological support dialogue datasets.