Do Entailment Models know about Reasoning Temporal Ordering on Clinical Texts?

Resumen

This paper investigates the use of entailment-based methods for Clinical Temporal Relation Extraction (CTRE), addressing challenges such as data scarcity, label imbalance, and domain-specific complexity. By reframing the task as a Natural Language Inference (NLI) problem, the approach reduces annotation requirements and improves generalization across datasets. Experiments with the THYME and E3C corpora show that NLI-based models outperform traditional classifiers in lowresource settings, demonstrating strong transferability and resilience to class imbalance, making them an effective solution for CTRE in clinical narratives.