Label Prompt Guiding for Two-Stage Few-Shot Named Entity Recognition

Although the current two-stage prototypical network approach based on pretrained language models has achieved success in few-shot named entity recognition (NER) tasks, the span boundary at the entity span detection stage is still difficult to locate, leading to the extraction of redundant or even incorrect entity spans, which greatly affects the subsequent entity type classification stage; moreover, how to improve the prototypical networks in the entity type classification stage so that it can better express semantic information remains a challenging issue. In this paper, we propose Label Prompt Guiding framework for two-stage few-shot NER (LPGNER), dividing the few-shot NER task into two stages: entity span detection and entity type classification. During training phase, the model is trained separately for the span detection and entity type classification stages by constructing a concise but effective prompt label template that provides additional labeling information for given samples. In the inference process, a span detector is used to extract and predict each input's entity span, then convert the input using the prompt template, and use the well-trained entity type classification model to further predict its type. The experimental results show that LPGNER improves upon baseline methods on the large-scale few-shot NER dataset FEW-NERD; at the same time, its generalizability across different domain benchmarks on cross-domain datasets has been verified.