Enhanced Prototypical Network for Few-Shot Named Entity Recognition

Few-shot Named Entity Recognition (NER) aims to perform NER tasks with limited data. Prototypical networks have already demonstrated excellent performance in addressing issues in few-shot settings. Traditional prototypical networks face issues such as inaccurate representation of O-class entities and poor distribution of entity prototypes. This paper presents an Enhanced Prototypical Network (EPN) for few-shot Named Entity Recognition, mainly including O-class subclass clustering module and feature consistency evaluation module. EPN divides O-class entities into multiple subclasses, each with its own prototype, to represent O-class entities more accurately. It also optimizes the distribution of entities in the feature space for better entity classification. Experimental results show that the model proposed in this paper improves performance on mainstream datasets.