Reproduce “MentalGLM Series: Explainable Large Language Models for Mental Health Analysis on Chinese Social Media” on OpenResearch

About this paper

Wei Zhai, Nan Bai, Qing Zhao, Jianqiang Li, Fan Wang, Hongzhi Qi, Meng Jiang, Xiaoqin Wang and 2 more

This research develops the MentalGLM series, the first open-source, explainable large language models tailored for multi-task mental health analysis on Chinese social media. The models, fine-tuned using the newly constructed C-IMHI dataset, achieve high accuracy, outperforming existing deep learning and general LLMs on tasks like suicide risk detection (85.12% F1) and cognitive distortion detection, while providing credible decision-making explanations and demonstrating generalizability to clinical data.

Key insights

• A two-stage fine-tuning process, combining general mental health domain adaptation with specific Chinese social media data, is crucial for achieving high performance and explainability in domain-specific LLMs.
• Large Language Models can achieve competitive or superior predictive performance compared to specialized deep learning models in mental health tasks while inherently offering the ability to generate human-interpretable explanations.
• High-quality instruction datasets, particularly those enriched with expert-guided explanations, are essential for effectively adapting LLMs to perform complex, explainable tasks in sensitive domains like mental health.

Summary by alphaXiv.