Chinese Electronic Medical Record Named Entity Recognition Based on PreAtt-BiLSTM-CRF
Author:Ming Gao, Shaochun Wu Author Unit: Department of Intelligent Information Processing, Shanghai University, Shanghai 200044,China
Abstract:The main task of Chinese electronic medical record named entity recognition (NER) is to identify medical
named entities such as diseases, symptoms, examinations, and treatments in electronic medical records. The
state-of-the-art method is the BiLSTM-CRF method based on deep learning, which uses the BiLSTM model
to extract medical record text features and uses the CRF model to obtain the optimal tag sequence. However,
the model tends to ignore the difference in attention between the “other” category and the medical record
entity category during feature extraction, causing the model to try to mark all words as other classes. This
paper proposes a new calculation method based on the attention mechanism to solve this problem. First, labels
are divided into a category of concern and a category of no concern. On the basis of this, the BiLSTM-CRF
model is assigned different attention and a larger weight is assigned to the entity of interest. The experimental
results show that while the overall F1 value is increased by 1.03%, the model can reduce the recognition
ability of “other” classes, and optimize the accuracy, recall and F1 values of “non-other” categories.
Keywords:Electronic medical record; Named entity recognition; BiLSTM-CRF
Download information [SIZE:1.40 MB Amount of downloads: second]