Exploring Synthetic Data Generation on Extremely Low-Resource Clinical Temporal Relation Extraction

Resumen

Temporal relation extraction (TRE) in clinical texts is essential but challenging due to limited annotated data and the domain-specific nature of medical language. This paper explores diversity/fidelity trade-off in data augmentation using large language models to generate synthetic training data under minimal supervision. We compare two approaches, Triplet-based Generation (TBG) and Paraphrase-based Generation (PBG), and evaluate their effectiveness across discriminative models. Our findings show that minimal manual annotation, even as little as one document, can significantly improve domain relevance and model performance. These results highlight the importance of supervision and data augmentation strategy selection when applying LLMs to domain-specific tasks like clinical TRE.