Abstract
The task of few-shot named entity recognition (NER) is to identify named entities by using limited annotated samples. Meta-learning, as a specific paradigm in the field of machine learning, has shown good results in acquiring the ability to "learn how to learn" and in quickly learning new tasks. However, some methods in the field of meta-learning identify named entities by calculating the word-level similarity between the query set and support set, without fully considering the label semantic information. To address this issue, we propose a method called UnionPromptNER for few-shot named entity recognition in the bridging domain. This method utilizes a joint prompt strategy to acquire label semantics, and then introduces a framework for computing the semantic representation of joint prompts. Through experiments on three different types of datasets, our proposed method achieved the best results in 19 out of 20 different settings compared with a series of previously optimal methods based on the micro F1 metric.