Research On Chinese Medical Answer Selection Based On Deep Learning

In recent years,the online medical community has become more and more active in China.On the one hand,it is because the traditional outpatient service has the problem of congestion caused by uneven distribution of medical resources and patients are limited in distance and time.On the other hand,under the background of COVID-19,online inquiry can relieve the pressure of the hospital under the line,provide professional medical services to patients and reduce the cross infection of COVID-19.However,the existing medical question answering（QA）is more manual.Therefore,the development of an automatic medical question answering system can effectively reduce the workload of doctors and alleviate the imbalance of medical resources.Answer selection is one of the key components of the automatic QA system,and the research on the answer selection algorithm determines the final answer quality of the automatic QA system.This article first explains the research background and significance of answer selection algorithms in the Chinese medical community.From three aspects: traditional algorithms based on feature engineering,general deep neural network algorithms,and deep neural network algorithms based on attention mechanism,a detailed analysis of the current research status of answer selection tasks is carried out.Then,existing problems of the existing answer selection model are pointed out,and the corresponding solution ideas and methods are designed.The main contributions of this article are:（1）For the Chinese medical QA data containing a large number of professional terms,how to deal with and represent these medical terms,so that the model can make full use of the rich semantic information in the Chinese text.A semantic feature information extraction method for Chinese text is proposed in this thesis.This method first uses the Chinese character embedding pre-trained by Word2 Vec to represent the text of the question and answer.The Chinese character embedding contains certain semantic information,and reduces the probability of unregistered words and reduces memory Consumption.Then Bi GRU is used for performing context encoding on the Chinese character embedding of the question and answer respectively,and modeling the context around Chinese characters.Next,a multi-scale convolutional neural network is used for extracting the local semantic feature information of the question and the answer at different scales.By sliding the window on the text by the multi-scale convolution kernel,semantic feature mining can be performed from multiple different scales such as characters,words,and phrases.Through comparative experiments and ablation experiments,it is proved that above modules can extract useful semantic information from Chinese texts for answer selection tasks,and improve the performance of the model.（2）For Chinese medical QA scenarios in a specific field,when the answer pool is composed of samples from the same medical topic such as epilepsy.The existing models have “over pooling” problem in the processing of questions and answers.Multi-scale co-attention fusion network（MCFN）proposed in this thesis first uses co-attention mechanism in the co-attention fusion module to model the interactive relationship between the question and the answer to generate attention information,so that the model can focus on meaningful text content in questions and answers simultaneously.Then semantic residual fusion mechanism is used to fuse local semantic feature information and attention information according to three different comparison operations.Under the action of the co-attention fusion module,MCFN will pay attention to the important information in the question and answer based on the interactive information between the question and the answer when selecting the answer.And for preventing co-attention mechanism from focusing on the information related to the topic,semantic residual fusion mechanism introduces the previously extracted local semantic information,so that the model can focus on some important text content that is not related to the topic.Through comparison experiments and ablation experiments,it is proved that co-attention fusion module enables MCFN to alleviate “over pooling” problem.（3）For the problem that Word2 Vec can not adjust the word vector according to the specific context,this thesis proposes a BERT answer selection algorithm based on co-attention fusion mechanism（BERT-CF）.BERT-CF uses BERT to dynamically adjust the word vector,and uses the prior knowledge and strong representation ability obtained by pre-training BERT.And co-attention fusion mechanism enables BERT-CF to pay attention to other important text information in addition to itself or non topic related in the answer selection task in a specific field.Comparison experiments show the outstanding performance of BERT-CF in answer selection task.In order to prove the effectiveness of the model,in addition to using the public Chinese medical QA datasets c Med QA v1.0 and v2.0,this thesis also creates a Chinese epilepsy QA dataset.From the experimental results,the performance of MCFN and BERT-CF proposed in this thesis exceeds the benchmark methods on the three datasets,and also outperforms the state-of-the-art model on cMedQA v1.0 and v2.0.