Formulaic Expression Desensitization Boosts Scientific Sentence Classification on Small Datasets

Formulaic expression desensitization augments small scientific paper datasets by abstracting domain-specific expressions to reduce model overfitting to surface forms, improving generalization for problem and method sentence extraction. The approach combines synthetic data generation with context enrichment to compensate for limited labeled data.