Adversarial training based lattice LSTM for Chinese clinical named entity recognition

Abstract

Clinical named entity recognition (CNER), which intends to automatically detect clinical entities in electronic health record (EHR), is a committed step for further clinical text mining. Recently, more and more deep learning models are used to Chinese CNER. However, these models do not make full use of the information in EHR, for these models are either word-based or character-based. In addition, neural models tend to be locally unstable and even tiny perturbation may mislead them. In this paper, we firstly propose a novel adversarial training based lattice LSTM with a conditional random field layer (AT-lattice LSTM-CRF) for Chinese CNER. Lattice LSTM is used to capture richer information in EHR. As a powerful regularization method, AT can be used to improve the robustness of neural models by adding perturbations to the training data. Then, we conduct experiments on the proposed neural model with dataset of CCKS-2017 Task 2. The results show that the proposed model achieves a highly competitive performance (with an F1 score of 89.64%) compared to other prevalent neural models, which can be a reinforced baseline for further research in this field.

Keywords: Adversarial training; CRF; Clinical named entity recognition; Lattice LSTM.