PLEM: Prototype Learning With Evidence Match for Improving Few-Shot Document-Level Relation Extraction

Few-shot document-level relation extraction (FSDLRE) aims to develop a model with the ability to generalize to new categories in the context of document-level relation extraction, using a small number of support samples. Among others, metric based meta-learning methods are widely used in FSDLRE, which involve constructing class prototypes using the contextual representation of the entire document and the representation of entity pairs for relation classification. However, in relation classification, only a subset of sentences in a document, known as evidence, is required to determine the relationship category of entity pairs. In this paper, we propose a prototype learning method with evidence match (PLEM). By introducing an evidence matching auxiliary task in the process of relation prototype construction, the model is guided to focus more on the semantics of evidence sentences when building prototypes, thereby enhancing the relation prototypes. We further design task-specific evidence prototypes, enabling the model to adapt to the evidence semantic space of different relation categories. Extensive experimental results demonstrate that PLEM outperforms the state-of-the-art methods, achieving an average improvement of 1.23% in Macro F1 across various settings of two FSDLRE benchmarks.