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

Area of Science:

• Computational linguistics
• Mental health informatics
• Natural Language Processing

Background:

• Depression is a global mental health concern requiring early detection and intervention.
• Social media data offers potential for early depression warning but traditional methods lack accuracy and interpretability.
• Existing studies often use binary classification, missing nuanced depression severity identification.

Purpose of the Study:

• To develop a fine-grained depression degree recognition model using multi-level text classification and generation.
• To explore enhanced identification in the Chinese social media context.
• To create novel datasets for depression degree classification and interpretable instruction-following.

Main Methods:

• Modeled depression degree recognition as multi-level text classification and generation tasks.
• Constructed the Depression Weibo dataset and a multi-level interpretable instruction dataset.
• Fine-tuned the ChatGLM3 large language model to create DepGLM (DepGLM-tc and DepGLM-tg).

Main Results:

• DepGLM-tc achieved a weighted F1 score of 84%, surpassing current state-of-the-art discriminative models.
• DepGLM-tg demonstrated comparable accuracy to SOTA discriminative methods.
• The text generation model produced high-quality explanations, similar to ChatGPT.

Conclusions:

• Fine-tuned large language models offer superior accuracy for identifying depression severity.
• The developed DepGLM model shows enhanced performance in depression classification tasks.
• This approach advances nuanced depression detection using social media data.