Chinese Medical Record Entity Recognition Based on Lexicon and Self-attention

Named Entity Recognition (NER) is a pivotal challenge in the field of natural language processing. In the medical field, NER plays a crucial role in identifying medical entities, such as diagnoses and drugs, from electronic medical records. To address the issue of incomplete and inaccurate word recognition in Chinese electronic medical records, we propose a BERT-BiLSTM-CRF model that incorporates both a lexicon and self-attention mechanism.In our approach, we first utilize the BERT model for preprocessing to generate word vectors. These vectors capture important semantic information about the words. Next, we employ a BiLSTM network to further capture contextual information from the input sequence. This helps to enhance the understanding of the text's meaning. To reinforce the significance of the sequence information, we introduce a self-attention layer that learns to prioritize relevant contextual details. Subsequently,the extracted features are fed into a CRF model for final training. The CRF layer takes into account the dependencies among the labels and improves the overall sequence labeling accuracy. Given the specialized nature of the medical domain, we integrate a medical lexicon, incorporating external knowledge features into the model. This inclusion enables better recognition of medical terms and enhances performance on medical-related entities.Through extensive experiments on the CCKS2019 dataset, our proposed model demonstrates favorable performance. It get relatively excellent results in identifying medical entities accurately and effectively from Chinese electronic medical records.