Medical named entity recognition based on dilated convolutional neural network

Abstract

Named entity recognition (NER) is a fundamental and important task in natural language processing. Existing methods attempt to utilize convolutional neural network (CNN) to solve NER task. However, a disadvantage of CNN is that it fails to obtain the global information of texts, leading to an unsatisfied performance on medical NER task. In view of the disadvantages of CNN in medical NER task, this paper proposes to utilize the dilated convolutional neural network (DCNN) and bidirectional long short-term memory (BiLSTM) for hierarchical encoding, and make use of the advantages of DCNN to capture global information with fast computing speed. At the same time, multiple feature words are inserted into the medical text datasets for improving the performance of medical NER. Extensive experiments are done on three real-world datasets, which demonstrate that our method is superior to the compared models.