Joint contrastive learning with semantic enhanced label referents for few-shot NER

Few-shot named entity recognition (NER) faces a challenge of obtaining more generalized class representations with limited labeled instances to recognize named entities. However, few support instances are insufficient to encompass the intra-class diverse entity semantics, and learning the similarity between query instances and support instances in isolation predisposes dispersed class distributions and fuzzy boundaries of certain classes, leading to misclassification. Meanwhile, there are differences in the meanings of the same class labels under different target domains, which leads to low accuracy of label referents generated based on label words, affecting the similarity metric between query instances and label referents. To overcome the problem, we propose a new method, Joint Contrastive learning with semantic Enhanced Label Referents for few-shot NER (JCELRNER). Specifically, we propose a strategy for representing the class using multiple referents, including support instance tokens and label prototypes. A joint contrastive loss objective that simultaneously is proposed to optimize the distributional distance between query instances and all class referents in the same semantic space by jointly considering token-to-token and token-to-label correlations, promoting the intra-class compactness. Meanwhile, JCELRNER proposes a dual semantic enhancement module to generate reliable label referents that adapt to different target domains by fusing entity and context semantic information. Extensive experiments show that by promoting the intra-class relevance of referents of the same class and enhancing the label referents with dual semantic information, the model achieves a comprehensive and accurate understanding of the entire class, thereby enhancing the generalization ability to novel classes, outperforming the other competitive approach.