Reproduce “STAMPsy: Towards SpatioTemporal-Aware Mixed-Type Dialogues for Psychological Counseling” on OpenResearch

About this paper

Jieyi Wang, Yue Huang, Zeming Liu, Dexuan Xu, Chuan Wang, Xiaoming Shi, Ruiyuan Guan, Hongxing Wang and 2 more

Researchers from Peking University and collaborators introduced STAMPsy, the first Chinese mixed-type psychological counseling dialogue dataset, which integrates spatiotemporal states and goal sequences. They developed Self-STAMPsy, an iterative self-feedback framework that leverages helping skills, spatiotemporal awareness, and adaptive RAG to generate more effective and empathetic responses, achieving superior performance on automatic and manual evaluation metrics compared to baseline LLMs.

Key insights

• Explicit guidance through counselor helping skills and client goal clarification is crucial for enabling AI systems to conduct more structured and therapeutically relevant conversations, as shown by significant performance drops when this module was ablated.
• Incorporating spatiotemporal awareness allows for the generation of significantly more empathetic and personalized responses, demonstrating its direct impact on perceived helpfulness and emotional attunement.
• An iterative self-feedback loop enhances dialogue quality and consistency, suggesting that self-reflection mechanisms can improve the overall coherence and effectiveness of generative dialogue models in sensitive domains.

Summary by alphaXiv.