Medical Question Answering System Based On Knowledge Graph

More and more people have been consulting medical and health problems via the Internet.But information overload and false information problems are increasingly exposed,and it is difficult for people to filter out really useful information from a large amount of results.Question answering systems based on knowledge graphs have been widely used in information consultation in many fields.Differing from search engines,the question answering system employs the professional knowledge graph as the knowledge source,thus users can get accurate answers to effectively avoid information overload and false information.The thesis focuses on the entity extraction method and user intent recognition algorithm in the medical question answering system based on knowledge graph.The main contents include:(1)Aiming at the problem of sample imbalance in the Chinese medical named entity recognition dataset,a FT-BERT+Soft-Lexicon_BiLSTM_CRF model fused with radicals is proposed.The model uses the FT-BERT obtained through medical data training to extract word vectors,uses the Soft-Lexicon method to incorporate lexical information into the word vectors,and fuses radical features,which solves the problem of poor model performance caused by unbalanced sample distribution.Compared with BiLSTM_CRF,the F1 value of the model reaches 93.12%,an increase of 2.15%.(2)Aiming at the problem that the medical user intent recognition dataset has a small amount of data and not rich label types,this paper integrates and re-labels the public datasets KUAKE-QIC,CHIP and CMID to obtain a new dataset.Aiming at the problem of more noise in the dataset,a hybrid neural network text classification method is proposed,which uses the weighted CLS vector of ERNIE as the input of TextCNN,and the word vector generated by ERNIE as the input of BiLSTM.Mixing TextCNN and BiLSTM improves the robustness of the model.Compared with the TextCNN and ERNIE_BiLSTM algorithms,the F1 value on the dataset is increased by 17.45%and 2.16%respectively.(3)A medical question answering system based on knowledge graph was designed and implemented.After knowledge acquisition,a medical knowledge graph with about 44,000 entities and 290,000 entity relationships was obtained.The two algorithms based on the research realize the medical question answering system,and find the corresponding answer from the graph database according to the acquired entity,relationship or attribute according to the rule template,which has the functions of entity recognition,user intent recognition,and medical consultation question answering.