With the advent of the information age, electronic medical record resources are growing exponentially and have diverse data forms. In this paper, we propose SoftLexicon-RoBERTa-BiLSTM-CRFS (SLRBC) to implement electronic medical records named entity recognition. It can embed known lexical information into a character-level-based model to obtain lexical information, and yet avoid the impact of superposition errors caused by word-level-based models due to word separation. The pre-trained model RoBERTa enables the model to obtain a better embedding representation, which can effectively improve the accuracy of the model. We conduct experiments on the public datasets CCKS2018 and CCKS2019, and achieve better results to verify the effectiveness of the model.