NER has become an important fundamental link in numerous NLP tasks such as information extraction, machine translation, question-answering systems, semantic understanding, and knowledge graph construction. Based on the application of pre-trained language models in the NER field, for low-resource domains and minority languages, due to the lack of sufficient labeled data, the training and performance of the model are greatly limited. This study proposes an improved NER framework that integrates dynamic mask enhancement, multi-granularity attention mechanism and adversarial domain adaptation. On the BERT-BiLSTM-CRF infrastructure, the recognition accuracy of entity boundaries is improved through the dynamic mask strategy, the cross-level attention fusion mechanism is designed to enhance the ability of long-distance dependency modeling, and domain adversarial training is introduced to solve the problem of domain transfer. Experiments on standard datasets such as CoNLL2003, MSRA-NER and CCKS2019 show that the improved model has advantages.