In order to retain more characteristic information in the training process, this study uses pre-training word vector and fine-tuning word vector to extend Bi-directional Long Short-Term Memory network (Bi-LSTM), and combines the co-training semi-supervision method to deal with the feature of sparse annotated text in the medical field. An improved model of Co-Training Double word embedding conditioned Bi-LSTM (CTD-BLSTM) is further proposed for Chinese named entity recognition. Experiments show that compared with the original BLSTM and BLSTM-CRF, the CTD-BLSTM model has higher accuracy and recall rate in the absence of corpora, the proposed method can better support the construction of medical knowledge graph and the development of knowledge answering system.