Research On Question-answering Model In Traditional Chinese Medicine Field Based On Recurrent Neural Network

The rapid development and wide popularization of the Internet promote the rapidity of information acquisition.It has become a development trend to obtain effective information through man-machine question-and-answer.Traditional Chinese medicine is the summary of the knowledge and experience of the Chinese people in the long-term struggle against diseases,and plays a vital role in the medical and health care of the Chinese people.At present,a large amount of knowledge in the field of traditional Chinese medicine exists in unstructured medical record texts,which can not be effectively used;When learning the existing diagnosis and treatment experience,doctors can only obtain diagnosis and treatment ideas by analyzing the medical cases,and can not accurately obtain the corresponding diagnosis and treatment knowledge.In view of the above problems,this thesis integrates and unifies a variety of knowledge sources in the field of traditional Chinese medicine through the knowledge graph,and use the way of man-machine question-and-answer to accurately obtain the desired knowledge.This thesis carries out key technology research from the aspects of the construction of traditional Chinese medicine knowledge graph,question generation and the classification of question intention.The main work is as follows:1.In view of the lack of unified system description of traditional Chinese medicine knowledge,this thesis analyzes the knowledge architecture of traditional Chinese medicine,constructs the knowledge ontology layer of traditional Chinese medicine,and constructs a knowledge graph of traditional Chinese medicine by using a variety of data sources,including structured knowledge base data,semi-structured web page data and unstructured medical case text data.For the problem that the existing entity recognition models fail to make full use of sentence features,a named entity recognition model NERCW（Named Entity Recognition by Fusing Character and Word Information）integrating word information is proposed.NER-CW introduces the word information of each character through the attention mechanism with context,and then combines the character and word information as the input feature for entity recognition.The experimental results on the constructed traditional Chinese medicine medical case dataset show that the F1 score of NEW-CW is 97.16%.2.Aiming at the lack of corpus in the field of traditional Chinese medicine questionand-answer,based on the constructed knowledge graph of traditional Chinese medicine,the question generation model QG-TC（Question Generation Model based on Triplet Description Features and Copy Mechanism）is proposed.In the QG-TC model,the description features of triples are introduced into the encoder to enrich the input feature,and the copy mechanism is introduced into the decoder to generate words that do not appear in the training data.The experimental results on NLPCC-2018-KBQG,Simple Question and TCM-KBQG datasets show that the three evaluation indexes BLUE-4,METEOR and ROUGE-L of QG-TC are better than the Original Encoder-Decoder architecture.3.Aiming at the problem of semantic sparsity caused by the short length of questions,an intention classification model CF-GAT（Intention Classification Model by Fusing Corpus Features and Graph Attention Networks）is proposed.Cf-GAT calculates the weight of words and the edge weight between words in the corpus respectively through gravity model and point mutual information,and then integrates them into the graph attention network to guide node update.The experimental results show that the average accuracy of CF-GAT is 75.48% on six short text classification datasets and the accuracy is 92.41% on the constructed TCM-QA dataset.4.The question answering system based on TCM knowledge graph is designed and built.The system takes the Springboot framework as the server architecture,apply flask as the deployment architecture of the model algorithm,and uses the Neo4 j to store triples.The main functions include user authority management,entity recognition,traditional Chinese medicine knowledge graph management,traditional Chinese medicine knowledge map visualization,intention classification,intelligent question answering,etc.The system extracts the knowledge of unstructured text through the entity recognition function,and then stores the extracted knowledge into the knowledge graph.At the same time,it can visually display the subgraph of the knowledge graph,and answer the questions entered by the user.