NER in the CSIRO Adverse Drug Event Corpus

Overview

This project focuses on Named Entity Recognition (NER) for biomedical texts, specifically on detecting Adverse Drug Reactions (ADR) in the CADEC dataset. The dataset consists of user-generated medical forum posts labeled with different medical entities.

We employ BioBERT, a transformer-based biomedical model, and enhance it with Focal Loss to address class imbalance, significantly improving the detection of rare medical entities.

---

Features

• Transformer-based NER model for biomedical text classification.
• Fine-tuned BioBERT model to extract medical entities from patient-reported drug reviews.
• Focal Loss implementation to improve recognition of underrepresented entities.
• Performance evaluation based on standard metrics like Accuracy, Precision, Recall, and F1-score.

---

Results Summary

Metric	BioBERT + Cross-Entropy	BioBERT + Focal Loss
Accuracy	88.92%	89.15% (+0.23%)
F1-score	88.04%	88.20% (+0.16%)
Rare Entity (B-Finding)	0%	21.51%
Rare Entity (I-Disease)	1.68%	17.33%

---

The notebook includes:

• Data preprocessing & tokenization
• Model training (BioBERT + CrossEntropy / Focal Loss)
• Performance evaluation
• Visualization of results

---

Model Performance

Standard BioBERT Model (Cross-Entropy Loss)

• High overall accuracy but struggles with rare entities.

BioBERT + Focal Loss

• Improves recognition of rare entities (e.g., diseases, symptoms, ADRs).
• Slight increase in accuracy and F1-score with better class balance.