DeBERTa-BiLSTM: a multi-label classification model for depression emotions.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Mental Health Informatics

Background:

• Multi-label depressive emotion classification is complex due to overlapping emotions and subtle linguistic cues.
• Existing methods struggle with long-range dependencies and differentiating active/passive depressive signals.

Purpose of the Study:

• To develop an effective and computationally efficient framework for multi-label depressive emotion detection.
• To accurately identify explicit active and latent passive depressive emotions from text using advanced NLP models.

Main Methods:

• Evaluated transformer-based and hybrid architectures including BERT, RoBERTa, T5, BART, and DeBERTa with BiLSTM.
• Proposed DeBERTa-BiLSTM model integrating disentangled self-attention and BiLSTM for contextual and sequential learning.
• Utilized the DepressionEmo dataset with eight clinically relevant depression emotions.

Main Results:

• The DeBERTa-BiLSTM model achieved superior performance with F1-Micro of 0.83 and F1-Macro of 0.80.
• Demonstrated robust detection of both frequent and minority emotion labels with micro-precision 0.81 and micro-recall 0.85.
• Showcased improved inference efficiency compared to seq2seq BART, with reduced time per sample.

Conclusions:

• The DeBERTa-BiLSTM model offers a promising approach for accurate and efficient multi-label depressive emotion detection.
• Further validation is needed across broader mental health domains and larger datasets.
• The model's performance highlights the potential of integrating advanced deep learning architectures for mental health text analysis.