SkimLit – Biomedical Abstract Sentence Classifier

SkimLit is an advanced NLP-based project designed to classify sentences in biomedical abstracts into rhetorical categories like Objective, Methods, Results, Background, and Conclusion. It helps researchers quickly skim through large volumes of biomedical literature by automatically organizing abstract content.

Key Features:

• 📚 Trained on the PubMed 20k RCT dataset for real-world biomedical sentence classification.
• 🧠 Uses a multi-input neural architecture combining:
  • BERT-based token embeddings
  • Character-level BiLSTM representations
  • Sentence positional features (line number, total lines)
• 🔁 Employs attention and gating mechanisms for refined context-aware feature fusion.
• 📈 Achieved 90.57% accuracy, outperforming several baseline models.
• ⚡ Built for efficiency and runs in low-compute environments (no fine-tuning of BERT required).

Architecture Highlights:

• Semantic understanding via frozen pre-trained BERT embeddings
• Character-level context via Bidirectional LSTM
• Structural awareness via sentence position encodings
• Intelligent attention mechanism to weigh feature importance