Evolutionary Self-Supervised Contradiction Detection for Biomedical NLI

TMLR Paper9224 Authors

Abstract: Identifying conflicting claims in biomedical literature is critical for advancing scientific understanding, yet the scarcity of high-quality training data remains a significant challenge. We introduce EvoNLI, an evolutionary algorithm that learns how to transform entailing sentence pairs into challenging contradictions by mutating words until a frozen teacher model confidently flips its prediction, while preserving topical coherence. EvoNLI, applied to PubMed randomized controlled trials (RCTs), generates SciCon, a dataset of premise–hypothesis pairs whose labels achieve 94.4\% agreement across expert judgments in an audit by five domain experts. Fine-tuning large language models on SciCon improves contradiction ROC-AUC consistently across eight biomedical NLI benchmarks. EvoNLI and SciCon are publicly available to support evidence synthesis and robust biomedical natural language inference, and to advance robust domain-specific contradiction detection.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Prayag_Tiwari1

Submission Number: 9224