Few-Shot Named Entity Recognition Based on the Collaborative Graph Attention Network

Few-shot Named Entity Recognition (NER) aims to extract entity information from limited annotated samples, addressing the scarcity of data in specialized domains. However, existing few-shot NER methods relying on data augmentation struggle to adequately augment semantic features, limiting their learning and representation capabilities. To overcome this, we introduce a novel few-shot NER encoder based on a Collaborative Graph Attention network (ColGAT). This encoder utilizes a collaborative graph-based data augmentation mechanism to thoroughly extract latent semantic features of entities within sentences, enabling precise entity recognition. Furthermore, to facilitate information interaction between support and query sets, we develop an entity classifier with Match Processing (MP), where adaptive weights allow the support set to flexibly adapt to different query instances, enhancing entity classification performance. Our model achieved an average F1 result of 65.87% across six datasets, surpassing the second-ranked model by 2.19% and achieving state-of-the-art performance, demonstrating significant improvements over previous methods.