关键词: 关系抽取, 小样本学习, 预训练模型, 对比学习, 多头注意力

Abstract: Few-shot relation extraction aims to identify the semantic relationship between entities in text when a limited number of labelled examples were available. To overcome the feature-alignment deficiency and task-adaptation weakness caused by single-view contrastive learning and static graph structures, a few-shot relation extraction model fusing multi-view contrastive learning and dynamic graph generation was proposed. During the pre-training stage, sentence-anchored and label-anchored strategies were introduced through multi-view contrastive learning to enhance the alignment between instance features and relation-label features. In task the-generation stage, a graph-generation module was used to construct task-specific graphs, and a multi-head attention guidance layer was applied to dynamically reallocate feature importance thereby, improving the model's adaptability to few-shot and cross-domain tasks. Experiments on the FewRel 1.0, FewRel 2.0, and NYT-25 datasets were carried out. The results showed that the proposed method achieved high accuracy in multiple N-way K-shot settings, demonstrated strong general andization task adaptability, and confirmed its effectiveness in few-shot and cross-domain scenarios.

Key words: relation extraction, few-shot learning, pre-trained model, contrastive learning, multi-head attention