NuNER Combined with Global Boundary Matrix for Few-shot Named Entity Recognition

The few-shot named entity recognition task aims to recognize entities using limited labeled data. Traditional methods typically optimize the model in a tagged manner, which may be adversely affected by non-entities and hence the accuracy of the model. To address this issue, we propose a two-stage approach. In the span extraction stage, NuNER is used as an encoder to convert the continuous labeling into a global boundary matrix. In this way, the model can utilize the semantic knowledge and entity information provided by NuNER across multiple domains, as well as the boundary information in the global boundary matrix, thus improving the accuracy of span extraction. In the entity classification phase, a distance-based loss function is introduced to optimize the prototype training. In addition, a marginal loss is employed to ensure that potential false positive spans are kept at a distance from the prototype region, thus improving the accuracy of entity classification. Comparison experiments with multiple baselines demonstrate the superior performance of the proposed model. In addition, ablation studies emphasize the necessity of each component and validate the effectiveness of the model.