DeBERTa-Enhanced Extreme Multi-Label Classification for Biomedical Articles

This study introduces a novel approach to the extreme multi-label classification of biomedical articles, leveraging the advanced capabilities of the DeBERTa (Decoding-enhanced BERT with Disentangled Attention) model. Aimed at addressing the challenges of classifying research articles into Medical Subject Headings (MeSH) labels, our methodology encompasses a comprehensive preprocessing strategy, model adaptation for multi-label classification, and a rigorous evaluation framework. Our DeBERTa-enhanced architecture demonstrates notable success in achieving a Test F1 Accuracy of 85.58%, underscoring its effectiveness in balancing precision and recall across a diverse array of MeSH categories. Through a detailed analysis of model performance across various labels, we identify strengths and highlight areas for future improvement. This study not only contributes to the field of natural language processing (NLP) applications in biomedical literature but also sets the stage for further advancements in multi-label classification techniques.