Few-Shot Joint Extraction of Entity-Relation Triples in the Domain of Ancient Chinese Medical Texts

Entity–relation extraction aims to identify structured triples from unstructured text. However, most existing models are evaluated on large, densely annotated corpora; when applied to vertical domains with scarce labeled data, fully su-per-vised models tend to generalize poorly. Existing few-shot learning ap-proaches often simplify the triple extraction task into relation classification with known entities, or overlook the common real-world scenario where a sen-tence con-tains multiple triples. To address these challenges, this paper focuses on ancient Chinese medicine texts and proposes a multi-granularity fused pro-totype net-work. The model fully integrates semantic information between rela-tions and entities during prototype construction to enhance the discriminability among different prototypes. A pointer network is further employed to enable the ex-traction of multiple triples from a single text. Experiments on a self-constructed dataset of ancient Chinese medicine texts demonstrate the ef-fectiveness of the proposed approach.