Scholarlink

<p>Detecting depression in low-resource and code-mixed social media text remains highly challenging, due to scarcity of annotated datasets and complexity of mixed language pattern. The study presents Multi-MentalBERT, a hybrid deep learning framework to perform dual-level multi-class classification across code-mixed and monolingual datasets. It first classifies the text as depressive, non-depressive or neutral, and then predicts the specific Major Depressive Disorder (MDD) category for each depressive instance. The framework integrates multilingual contextual features from MuRIL with mental health specific embeddings of MentalBERT through Cross-Attention fusion, supported by Hierarchical attention and Deep Joint Autoencoder to create a unified representation learning. The framework achieved a macro F1-score of 91.03% for three-class depression detection and 87.75% for MDD classification on Kannada-English dataset, and obtained 95.87% macro F1-score for MDD classification on English Reddit depressive dataset. The result outperforms the baseline models and demonstrates the effectiveness of the approach in automated mental health analysis across linguistic boundaries.</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

A hybrid deep learning framework for multilingual depression detection and symptom classification from social media text

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • C. H. Shwetha,
  • K. Pushpalatha

摘要

Detecting depression in low-resource and code-mixed social media text remains highly challenging, due to scarcity of annotated datasets and complexity of mixed language pattern. The study presents Multi-MentalBERT, a hybrid deep learning framework to perform dual-level multi-class classification across code-mixed and monolingual datasets. It first classifies the text as depressive, non-depressive or neutral, and then predicts the specific Major Depressive Disorder (MDD) category for each depressive instance. The framework integrates multilingual contextual features from MuRIL with mental health specific embeddings of MentalBERT through Cross-Attention fusion, supported by Hierarchical attention and Deep Joint Autoencoder to create a unified representation learning. The framework achieved a macro F1-score of 91.03% for three-class depression detection and 87.75% for MDD classification on Kannada-English dataset, and obtained 95.87% macro F1-score for MDD classification on English Reddit depressive dataset. The result outperforms the baseline models and demonstrates the effectiveness of the approach in automated mental health analysis across linguistic boundaries.