A Feature-enhanced Model for Chinese Medical Text Classification Based on Improved BERT and Feature Fusion

With the development of internet hospitals and natural language processing technology, medical text classification based on machine learning gains increasing attentions. In this paper, we propose a feature-enhanced Chinese text classification model for intelligent medical consultation, which combines char-word level tokenization and feature fusion. Specifically, we expand the dictionary of BERT and improve its segmentation mechanism to make full use advantages of different segmentation mechanisms and pre-trained models. In addition, we select 243 common diseases and symptoms from different departments as manual features, and apply the attention mechanism, CNN with multiple kernel sizes and max-pooling operation to capture potential correlations and dependencies between patients' self-statement and manual features. In order to verify the effectiveness of the proposed model, we carry out many simulation experiments. Experimental results show that compared to other typical text classification models, the proposed model achieves better classification performance with accuracy of 96.87% and Fl-score of 96.75%. In addition, the proposed model improves the accuracies of most departments by 1% as compared to the Google BERT.