DepGLM: Depression degree recognition on social media based on large language models.

Background: Depression is recognized as one of the most prevalent mental health disorders globally. Early detection and intervention can significantly reduce its negative impact on patients. Data from social media platforms provide new opportunities for early warning, intervention, and support for individuals with depression. However, traditional discriminative methods are limited in both the accuracy and interpretability of depression identification. Most studies focus on binary classification tasks, lacking more fine-grained identification. Methods: We model depression degree recognition as a multi-level text classification task and a multi-level text generation task, exploring more fine-grained identification in the Chinese context. We construct the depression degree ((Depression Weibo) dataset and the first multi-level interpretable depression degree instruction dataset. Based on these datasets and ChatGLM3, a Chinese open-source large language model, we train DepGLM. We also compare how fine-tuned large language models perform against multiple baseline models across different tasks. Results: The results show that the text classification model, DepGLM-tc, demonstrates superior accuracy compared to current state-of-the-art (SOTA) discriminative approaches, achieving a weighted F1 score of 84%. Meanwhile, the text generation model, DepGLM-tg, approaches the accuracy of SOTA discriminative methods and is capable of generating high-quality explanations comparable to ChatGPT. Conclusion: This study demonstrates that large language models fine-tuned for specific domains more accurately identify depression degree. Moreover, our model demonstrates enhanced performance in classification tasks.