Abstract:In the field of Internet-based medical treatment, AI-based triage represents a key link, which allocates patients to departments according to conditions, disease attributes, medications, etc. We can adopt the BERT with a deep bi-directional Transformer structure for language model pre-training to enhance the word semantics; however, the text description of patients’ conditions offers sparse information, which is not conducive to the full learning of characteristics by BERT. This paper presents DNNBERT, a joint training model integrating knowledge. Combining the advantages of Deep Neural Network (DNN) and the Transformer model, DNNBERT can learn more semantics from text. The experimental results prove that the computing time of DNNBERT is 1.7 times shorter than that of BERT-large; the accuracy rate of DNNBERT is 0.12 higher than the F1 value of ALBERT and 0.17 higher than that of TextCNN. This paper will provide a new idea for sparse feature learning and the applications of deep Transformer-based models to production.