Overview
Detect negation cues and their scope in biomedical text with models that can handle imbalanced labels and structured output behavior.
System design
- Fine-tuned BERT, BioBERT, and ClinicalBERT with BIO-style token labeling
- Added CRF and hierarchical classification variants to model label dependencies
- Trained PPO-based sequence-labeling policies using rewards derived from human-annotated spans
- Evaluated cue and scope quality with token-level and span-level consistency checks
Results
- Established strong supervised baselines for negation cue and scope detection
- Improved minority-label behavior with reinforcement-style training over the supervised baseline
- Created a reusable experimental setup for later biomedical QA and evidence-quality work