| With the popularization of the Internet and the enhancement of people’s awareness of their own health care,more and more people obtain medical information through the Internet.However,the information on the current search engines is mixed,and there is a lot of irrelevant information;although the online consultation platform is highly authoritative,there is a problem that the feedback is not timely.Therefore,the medical question answering system based on knowledge graph constructed in this thesis is of practical significance.For the input question,the medical question answering system firstly performs intent classification and key entity extraction,and then returns the most accurate result after querying on the knowledge graph according to the user’s query intent.Specific studies include:（1）Semantic parsing of question sentences based on ALBERT.In this thesis,the semantic parsing of question sentences is divided into two sub-tasks,entity recognition and intent recognition,and the identification of entity keywords in questions and the classification of question intent are completed respectively.For the named entity recognition of medical texts,this thesis introduces the pre-training model ALBERT,gives the model a relatively complete initialization parameter,extracts contextual feature information in combination with the BiLSTM model,and uses the CRF model to introduce label constraints to correctly classify entity categories,and implements a medical entity recognition ALBERT-BiLSTM-CRF model.Experiments show that the F1 value of the model under the MED dataset and the CCKS electronic medical record dataset is 80.8% and 89.2%,respectively,and the effect is significantly higher than the classical model in the field of entity recognition.For the intent recognition of medical questions,this thesis introduces the ALBERT pre-training model to improve the overall understanding accuracy of the language,and then uses TextCNN to convolve and pool the semantic features to output the intent classification of the question,which is used to classify the intent of the question,and implements the ALBERT-TextCNN model for question sentence intent classification.The F1 value of this model under the intent recognition dataset MQA in this paper is 87.6%,and the effect is improved compared with the classic models in the field of text classification.（2）Design of answer retrieval method based on knowledge graph.Based on two situations that lead to inaccurate answer retrieval in the answer retrieval process,this thesis studies the entity preprocessing based on negative words and the synonym and near-synonym matching method based on semantic similarity measurement,and applies them to the process of answer retrieval.A more precise search result was obtained.At the same time,a medical knowledge graph with about 40,000 entities and about 250,000 entity relationships was constructed according to the processes of data collection,schema construction,and data storage,which provided the knowledge base for the question answering system.（3）Implementation and testing of medical question answering system.This thesis integrates various algorithm modules based on the Django framework,and builds an intelligent medical question answering system with functions such as entity recognition,entity search,relationship search,and medical question answering through the visual interface of Echars and Layui frameworks.The system can provide accurate answers according to the user’s query intention and question entity,and provide users with a good question-and-answer experience,which has high practical value. |
