Scholarlink

This paper presents an investigation into the fine-tuning of large language models (LLMs) on the D4 dataset, a Chinese dialogue dataset specifically curated for depression-diagnosis-oriented conversations via mixture of specialized experts. Recognizing the absence of large language model fine-tuning on this dataset, we aim to leverage the power of LLMs to develop a dialogue system capable of diagnosing depression through multi-turn conversations. Our approach involves fine-tuning several foundation LLMs, including Qwen-MoE, Deepseek-MoE, and LLaMA-MoE, on the D4 dataset to perform four critical tasks: response generation, topic prediction, dialogue summarization, and classification of depression severity and suicide risk. The objective is to construct a simulated clinical consultation dialogue system that can accurately diagnose depression, offering a realistic and empathetic user experience. Through extensive experiments, we demonstrate the effectiveness and superiority of our fine-tuned models in achieving high performance on the D4 dataset, showcasing the potential of LLMs in advancing mental health diagnostics.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DepLLM: Fine-Tuning Large Language Models with a Chinese Dialogue Dataset for Depression Diagnosis via Mixture of Specialized Experts

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Chaolin Xiong,
  • Hualiang Li,
  • Dawei Peng,
  • Ziyue Lin,
  • Wenhan Yang,
  • Yalong Wang

摘要

This paper presents an investigation into the fine-tuning of large language models (LLMs) on the D4 dataset, a Chinese dialogue dataset specifically curated for depression-diagnosis-oriented conversations via mixture of specialized experts. Recognizing the absence of large language model fine-tuning on this dataset, we aim to leverage the power of LLMs to develop a dialogue system capable of diagnosing depression through multi-turn conversations. Our approach involves fine-tuning several foundation LLMs, including Qwen-MoE, Deepseek-MoE, and LLaMA-MoE, on the D4 dataset to perform four critical tasks: response generation, topic prediction, dialogue summarization, and classification of depression severity and suicide risk. The objective is to construct a simulated clinical consultation dialogue system that can accurately diagnose depression, offering a realistic and empathetic user experience. Through extensive experiments, we demonstrate the effectiveness and superiority of our fine-tuned models in achieving high performance on the D4 dataset, showcasing the potential of LLMs in advancing mental health diagnostics.