Research And Application Of Semantic Matching Technology For Medical Question Answering System

In the era of the epidemic,people are more inclined to consult doctors,inquire about their illnesses and obtain treatment advice from doctors through online consultation.However,there is a lack of face-to-face observation and communication between online consultation doctors and patients,so it is difficult to assess the patient’s condition and provide accurate analysis and diagnostic suggestions.At the same time,online consultation requires patients to describe their symptoms,and different patients may not describe their own symptoms.As a result,the patient information obtained by doctors may be missing or wrong.Nowadays,various online medical platforms have accumulated a large amount of medical consultation data,and patients often ask some common medical questions.The answers to these common medical questions can be obtained from the consultation data of standard medical questions and answers,to provide patients with Provide more comprehensive and accurate medical diagnosis advice.Therefore,it is very necessary and feasible to develop a high-quality medical question answering system.By matching the standard medical questions most like the patient’s medical questions,it returns the answers to the standard medical questions in the medical text database.These answers are the results of real communication between patients and doctors,so they are highly scientific.Based on the above analysis,this paper proposes a semantic matching model based on Chinese medical text and a medical question answering system based on medical text matching technology.The main work of this paper is as follows:(1)The twin Siamese-Uni LM model is constructed by combining the twin network framework and the pre-training model.The self-coding Uni LM model and the autoregressive language model can capture the deep information of the text.The combination of the twin network can not only effectively understand the deep meaning of the medical text association,but also obtain the similarity information of the text.Compared with the traditional neural network model and the text matching method based on the pre-training model,In the open data set of chip-2019 medical text data set,the accuracy rate and F1 value increased by 3.95% and 3.96%,respectively;in the open data set of Tianchi Competition data set,the accuracy rate and F1 value increased by1.33% and 1.53%,respectively.This proves that the Siamese-Uni LM model can indeed capture the deep information of medical texts to a certain extent.(2)A knowledge distillation model based on the Siamese-Uni LM model is proposed.After knowledge distillation,the student model has fewer parameter layers and parameters than the teacher model.Compared with other large-scale pre-training models,the loss is very small.It achieves the same inference effect as other large-scale pre-trained models with less performance,making it a reality for large-scale models to be deployed online for inference.It can be seen from the experiment that after knowledge distillation,the model loses 4.16% of the accuracy rate in terms of key indicators such as accuracy and reasoning time,but saves 80.43% in model reasoning time compared with the teacher model.Compared with the teacher model,the time consumption is reduced.At the same time,the parameters of the model are effectively reduced.(3)By deploying the medical text semantic matching model proposed in this paper to online,a medical question answering system is constructed to provide a specific realization of the medical question answering function.