<p>Recently, the widespread application of in-context learning in few-shot named entity recognition tasks has garnered significant attention. Existing research endeavors to induce large language models into self-learning for prediction by introducing external knowledge. However, these methods neglect to seek the evidence of entity classification explicitly within sentences or model the classification evidence implicitly. Thus we model evidence explicitly and propose a novel approach called CoE-NER, which applies chain of evidence to few-shot NER task. First, we employ a strategy of inducing inference from gold labels to identify entity classification evidence at both syntactic and semantic levels. Then, we utilize the support set with the explicit chain of evidence as few-shot demonstrations and stimulate the self-learning of large language models to predict. Our method excavates more comprehensive evidence in sentences to support the recall of positive samples. Additionally, to incorporate semantic information about labels into prompts, we introduce a label conversion module, which removes misleading terms and standardizes labels. Experimental results on two real-world datasets (Few-NERD and CoNLL’03) demonstrate significant improvements achieved by our method.</p>