| In medical question-answer systems,the answers that users need often come from multiple data sources with different structures,and are also very different from each other.Knowledge mapping helps us present multi-source and complex data in a more intuitive and humanized way.Existing medical guidance systems have the following two problems: First,knowledge graph lacks medical data or medical data is incomplete;Secondly,low intelligence level caused by traditional deep learning models.To overcome the shortcomings mentioned above,the following researches are carried out: First,In order to build a better medical knowledge mapping,some customized extractors are designed to capture data,filter and screen non-sentence content according to different medical website structures.For the data filtered and screened,select a bidirectional recurrent network based on pre-trained models that uses less in the medical question and answer field,and combine conditional random fields for knowledge extraction.The extracted and captured data are proceeded and integrated by synonyms averaging algorithms,such as overlapped word similarity,edit distance,and cosine similarity.Neo4 j database is used for data storage.Secondly,Focusing on the improvement of the "intelligence level" of the system,some modules are designed and realized attentively,such as the feature extraction module and the intention recognition module.For the feature extraction module,we compare and evaluate the accuracies of the model integrated with BERT,BiLSTM,and CRF,respectively,not only by deeply benchmarking and integrating characteristic of other model,but also combining medical corpus data set.According to the characteristic that the length of input sentences in question-answer systems is usually short,we design and realize an intent recognition module based on TEXTCNN model,which selects TEXTCNN model for short texts to roughly judge the intention of the text,then filters and screens the output data with keywords.Thirdly,Relying on We Chat applet and Flask.A Chinese medical question-answer service platform is established with functions such as self-diagnosing,medical history recording,and medical messages pushing.Experiment results show that the BERT-BiLSTM-CRF fusion model performs better than the traditional BiLSTM-CRF in entity-name recognition.The F1 value is increased to 91% with 3% improvement.In short text intent recognition,F1 of TEXTCNN and BERT are both 94%.However,for single sentence prediction,TEXTCNN is faster than BERT.TEXTCNN takes 2.53 ms,while BERT takes 7.45 ms.The average time of each module in the system is abot 10 ms,and the response time of the service platform is about 0.35 s,which meets most application scenarios. |
