Chinese clinical entity recognition based on pre-trained models

摘要

In the Chinese clinical domain, there are problems such as blurred word boundaries, multiple meanings of one word, insufficient access to character information and a large number of unregistered words, making it more difficult to recognize named entities for this domain. In view of this, this paper proposes a glyph-based enhanced information model, using convolutional neural network and ALBERT to pre-train the language model to obtain the enhanced character information vector and also introduces an attention mechanism on the BiLSTM-CNN-CRF structure to solve the problem that the traditional model extracts features ignoring the importance and location information of words. Finally, experiments are conducted on the CCKS2018 dataset, and the results show that the model outperforms other commonly used models in Chinese clinical recognition.