ECU-BRE: NLI-Based Biomedical Relation Extraction with EC Supervision and Uncertainty-Aware Inference

Biomedical relation extraction (RE), which identifies semantic relationships among biomedical entities, plays a vital role in knowledge discovery. However, the task remains challenging due to ambiguous entity mentions, imbalanced relation distributions, and limited annotated data in the biomedical domain. A promising direction reformulates RE as a natural language inference (NLI) task, enabling models to better exploit semantic cues and reduce dependence on manual annotations. Although this reformulation enhances generalization, existing NLI-based approaches often neglect contradiction signals and tend to make overconfident, unreliable predictions–particularly in low-resource biomedical settings. To address these issues, we propose ECU-BRE, a framework that integrates entailment–contradiction supervision with uncertainty-aware inference. By introducing contrastive signals between entailment and contradiction during training, the model learns to better distinguish entity pairs expressing valid relations from those that do not. At inference time, Monte-Carlo dropout is used to estimate uncertainty and improve robustness–crucial for high-stakes biomedical applications. Experiments on the ChemProt and DDI datasets show that ECU-BRE delivers strong performance in diverse low-resource scenarios, outperforms existing baselines in most settings, and remains competitive under full supervision.