Research On Multi-Round Q&A In Medical Field Based On Knowledge Graph

With the deepening research on Question-Answering(QA)systems,simple singleturn QA has become relatively mature,but there is still a lack of research in multi-turn QA.In real-world scenarios,due to the non-rigorous sentence structure in daily Chinese communication,such as the frequent omission of the subject,the effectiveness of singleturn QA is not satisfactory in the usage process.Therefore,in order to make the QA system more user-friendly and accurately meet user needs,research on multi-turn QA is necessary.Currently,research on multi-turn QA is mainly focused on daily conversations and performs poorly when applied to the medical field.This is because the medical field has specialized terminology that cannot be recognized by existing systems,and there are many challenges in researching multi-turn conversations in the medical field.For example,there are few relevant medical datasets,medical data contains specialized terminology and nested entities,and mainstream models perform poorly on medical datasets.The thesis combines knowledge graphs and multi-turn interaction technology to implement knowledge graph-based multi-turn QA in the medical field.The research content of the thesis includes the following aspects:(1)In response to the problem of difficult recognition of specialized medical terminology and nested entities in the medical field,the thesis explores the inclusion of external feature vectors formed by medical professional terms in mainstream named entity recognition models.The thesis proposes the method of external feature vectors +BERT-Bi LSTM-CRF.Medical data is obtained through web crawlers to form a dictionary,and then medical professional terms are converted into vectors and embedded into the pre-trained language model.Experimental results show that introducing external feature vectors can effectively enhance the named entity recognition effect of the model in the medical field.(2)In response to the problem of a small number of medical-related QA datasets and insufficient training samples,the thesis studies methods of external data augmentation and models for processing small sample data.The paper proposes the Sim BERT data augmentation method combined with BERT-Text CNN for medical intent recognition.The Sim BERT model is used to incrementally process the small sample data in the dataset,and then BERT-Text CNN is used to train the augmented dataset.Experimental results show that data-augmented medical intent recognition performs better.(3)The thesis analyzes and studies methods of knowledge extraction,designs the Schema structure of the medical knowledge graph,and converts the extracted entity relationships into triple form according to the designed structure.A medical knowledge graph is built based on the Neo4 j graph database.Based on the constructed knowledge graph,a multi-turn QA system is researched,and a knowledge graph-based multi-turn QA system in the medical field is ultimately implemented.