Research On Medical Intelligent Question Answering Algorithm Based On LSTM&Topic-CNN Model

In recent years,more and more disease patients seek medical advice from doctors through online medical websites and medical apps.At present,the main way of online diagnosis and treatment is still that patients leave messages to doctors on the network platform,and the doctors give advice after browsing the information.However,this method is often not time-sensitive and cannot achieve the desired effect of online diagnosis and treatment.The intelligent question-and-answer system that can reply in real time and reply with high quality can not only help patients get more scientific and healthy medical advice,but also save a lot of medical resources,which is becoming the focus of current medical work.In view of this,based on the in-depth study of the question and answer system and the characteristics of online doctor-patient dialogue text,starting from the two aspects of community question and answer and knowledge graph,this paper constructs a set of medical question and answer system oriented to short text dialogue.The main work of this paper is as follows:(1)Based on the idea of feature fusion,the LSTM&Topic-CNN model is established by combining the semantic features of the text context with the topic features.On the one hand,the word2 vec method is used to obtain distributed word vectors,and the long short-term memory(LSTM)model is used to extract the semantic features of the text.On the other hand,in view of the fact that the number of topics in the implicit Dirichlet distribution(LDA)model is difficult to determine,the number of topics with the best performance is obtained through the sensitivity analysis experiment of the sub-sample set,and then the topic characteristics based on LDA are obtained.In order to avoid topic feature redundancy under different topic models,a one-dimensional convolutional neural network(CNN)is used to further feature filtering of joint topic features.Finally,the similarity analysis of patient questions is carried out by organically fusing the semantic features of the text and the topic features.Use python software to collect online doctor-patient dialogue texts from 39Q&A(http://ask.39.net/)network and 120 ask network(http://120 ask.net/)to construct experimental data sets.Combine the open source Chinese medical subject vocabulary,and use common drug dictionary mapping methods for text preprocessing.On this basis,the effectiveness of the LSTM&Topic-CNN model is verified through two-class and multi-class experiments.The results show that compared with other machine learning models(SVM,RF,RNN,LSTM),the model in this paper has higher recall,accuracy and F1 value,which further reflects that the model has a better text representation effect and indirect verification The feasibility and effectiveness of question answering system based on this model are discussed.(2)Based on the knowledge graph theory,this chapter proposes an improved intention recognition method and applies it to the intelligent question answering system for cardiovascular diseases.Relevant data is crawled from the cardiovascular disease module of Xunyiwenyao.com,and the neo4 j graph database is used to construct and store the cardiovascular disease knowledge graph.On this basis,the patient’s question is parsed through two methods:entity extraction and intention recognition.First,combined with commonly used drug dictionaries,a string matching method is used to map and extract the entity characteristics of patient questions.Secondly,based on the intention trigger words and the LSTM&Topic-CNN model,the patient’s question intention is identified from two aspects: subjective rule judgment and objective quantitative analysis.For patient questions containing intention trigger words,the patient’s intention can be directly obtained through subjective rules.For questions that do not contain intention trigger words,after obtaining similar questions through the LSTM&Topic-CNN model,the patient’s intention is determined according to subjective rules.Finally,based on the results of entity extraction and intent recognition,query the response template in the cardiovascular knowledge graph and get the most appropriate answer.