Chinese Medical Named Entity Recognition Method Based on Word-Word Relationship

Abstract

To solve the problem that the current deep learning method is difficult to deal with the recognition of nested entities in Chinese medical text, a deep learning model based on word-word relationship is introduced, and the relationship between words is built by multi-granularity 2D graphs to improve the recognition of nested entities. First, we use BERT (Bidirectional Encoder Representation from Transformers) for pre-training, then we use BiLSTM (directional Long Short-Term Memory) to extract the context information. Then, we merge the token representation information, the word distance information and the word regional information, through use a multi-granularity hole convolution to obtain the role information of different words. Finally, we use decoding layer to predict entity relationships and decode the result. This model is tested on the CMeEE Chinese medical dataset. Compared with the popularity models such as BiLSTM-CRF (Conditional Random Field) and BERT-BiLSTM-CRF, the F1 value is improved by 2.52%. Experimental results show that for Chinese medical named entity recognition with nested entities, this model can better recognize the medical entities in Chinese medical text.