Scholarlink

<p>The increasing number of mental disorders is a serious problem in the modern world and can even lead to suicide if left untreated. In the age of digitalization, we move part of our lives to social networks, where we share the good and bad moments. This allows for the detection of early mental disorders (such as depression, excessive stress, or social phobia) that the user may not even be aware of. We address the problem of effectively using large language models (LLMs) to detect mental disorders. In particular, we propose modifying LLMs by adding appropriate layers to improve the classification parameters and build classifiers for discovering mental disorders from texts. We have performed experiments with PHI-2, PHI-3, Mistral, Flan-T5, and LLaMA 2/3/3.1, showing that such an approach gives a better prediction performance than zero-shot/few-shot for LLMs and classification by BERT-based models. The proposed architecture makes it possible to return a label thus giving a specific classification of the case, rather than an ambiguous answer text.</p>

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

LLM-based classifiers for discovering mental disorders

  • 期刊论文

  • 发表日期：

    2025年4月4日

查看全文
  • Arkadiusz Nowacki,
  • Wojciech Sitek,
  • Henryk Rybiński

摘要

The increasing number of mental disorders is a serious problem in the modern world and can even lead to suicide if left untreated. In the age of digitalization, we move part of our lives to social networks, where we share the good and bad moments. This allows for the detection of early mental disorders (such as depression, excessive stress, or social phobia) that the user may not even be aware of. We address the problem of effectively using large language models (LLMs) to detect mental disorders. In particular, we propose modifying LLMs by adding appropriate layers to improve the classification parameters and build classifiers for discovering mental disorders from texts. We have performed experiments with PHI-2, PHI-3, Mistral, Flan-T5, and LLaMA 2/3/3.1, showing that such an approach gives a better prediction performance than zero-shot/few-shot for LLMs and classification by BERT-based models. The proposed architecture makes it possible to return a label thus giving a specific classification of the case, rather than an ambiguous answer text.