Aiming at the problems of low recognition accuracy and poor model recall of unstructured data in named entity recognition tasks, a named entity recognition method based on BERT-MD-RNN-CRF is proposed: firstly, the BERT pre-trained language model is used to extract character vectors containing context information in the text, and the vectors are used as model input vectors. Then, the multidimensional recurrent neural network and attention network are used to operate on the span information matrix to enhance the semantic connection between the spans. Conditional random field CRF is used to learn the internal relationships between labels to ensure their sequence; Finally, the augmented span information matrix is classified by span to identify named entities. Experiments show that compared with the traditional span method, this method can effectively enhance the semantic dependency feature between spans, thereby improving the recall rate of named entity recognition. This method improves the recall rate by 0.31% compared with the traditional method on the ACE2005 English dataset, and achieves a higher Fi value.