As one of the core tasks of natural language processing, NER plays a key role in many fields such as information extraction, knowledge graph construction, machine translation and Q&A system. BiLSTM-CRF combines the advantages of BiLSTM and CRF, and achieves excellent performance in NER task. Through forward and reverse LSTM network, it can make full use of the information before and after the text, and make a more comprehensive representation of the word in each position. In this study, based on BiLSTM-CRF, on the basis of giving full play to its advantages, DQN is used to correct labels with large uncertainties, so as to make up for the shortcomings of BiLSTM-CRF model. The advantages of the algorithm are verified by comparison experiment and migration experiment, and it is proved that the algorithm has good generalization ability, improves training efficiency and model performance, can play a role in NER tasks in different fields, and adapt to changing business requirements and data environment.