In this paper, we propose a Chinese Named Entity Recognition (NER) model based on Multihead AttentionMechanism (MAM), RoBERTa-BiLSTM-MHA-CRF, which combines the deep semantic representationcapability of RoBERTa, the global context modelling capability of MAM, and the advantage of BiLSTM-CRF incapturing sequence dependencies, to provide a novel solution for the Chinese NER task. , and the advantageof BiLSTM-CRF in capturing sequence dependencies, providing a novel solution for the Chinese NER task. Theexperimental results on the MSRA dataset show that the model outperforms the mainstream NER model inthe key metric of F1 value, with an F1 value of 95.43%. The ablation experiments further validate that thesemantic understanding capability of RoBERTa, the long-distance dependency modelling capability of themulti-attention mechanism, and the role of CRF for global label optimisation all have important contributionsin performance improvement. Compared with traditional methods, the model not only significantly enhancesthe ability to recognise complex entity boundaries, but also improves the model's ability to comprehensivelyunderstand contextual information. The results show that the RoBERTa-BiLSTM-MHA-CRF model is able toeffectively solve the semantic ambiguity and long-distance dependency problems in the Chinese NER task,and has high academic research value and application potential. Future work will focus on exploring themodel's adaptability in specific domains and its performance in low-resource scenarios to extend itscapabilities in practical applications.