Research And Application Of COVID-19 Question Answering System Based On Knowledge Graph

With the rapid development of software and hardware technologies,the construction and storage of large-scale knowledge graphs have become possible,and it has provided a knowledge base for artificial intelligence applications such as question answering systems and drug discovery.As a promising application in the field of artificial intelligence,the question answering system has been widely concerned.Compared with acquiring knowledge through search engine,question answering system can give exact answers more intelligently and efficiently.The knowledge based question answering system(KBQA)combines the advantages of both of them,parses the user’s query into logical form,and then retrieves and returns the answer in the knowledge graph.This thesis focuses on the key technologies and applications of KBQA.The main contents are as follows:Entity Mention Recognition,Entity Disambiguation and Relation Detection are the key technologies in KBQA system.First,aiming at the problem that there are a large number of entities and relations in the open domain knowledge graph,and there are differences in the form of expression between Chinese natural language questions and fact triples,a pipelined KBQA system called BERT-CKBQA based on feature-enhanced BERT is proposed to solve the single relation fact-type questions.The BERT-CRF model is used to identify the entities mentioned in the question;the candidate entity relation features are introduced to enhance the BERT-CNN model for Entity Disambiguation;the BERT-Bi LSTM-CNN model that introduces the answer entity relation features through the attention mechanism is proposed for Relation Detection.This method combines the pre-trained model and relation features to effectively improve the performance of the subtasks,and obtains an averaged F1-score of 88.75% on the NLPCC-2016-KBQA dataset,which improves the accuracy of question answering.Secondly,to solve the problem that the number of candidate paths recalled increases exponentially with the number of relation hops in the Relation Detection of complex question answering,a hop-by-hop Relation Detection framework is proposed which divides Relation Detection task into two main subtasks: stop decision task and path similarity task.Extension and aggregation operations are used to solve the chain type questions and multi-entity type questions respectively.Experimental results on the CCKS-2019-KBQA dataset shows the effectiveness of this method.While achieving competitive performance compared with existing methods,the hop-by-hop mechanism can significantly reduce the candidate size and improve the overall performance and efficiency of the question answering system.Finally,at the time of the outbreak of the COVID-19 in 2020,in order to allow medical personnel and non-computer professionals to easily and quickly obtain a large amount of valuable information about the COVID-19 contained in the knowledge graph,and to reduce the threshold for obtaining medical expertise,a question answering system based on the COVID-19 open knowledge graph is constructed by combining multiple COVID-19 open knowledge graphs released by Open KG and open domain KBQA key technologies.In this thesis,the system is divided into Question Classification module,Entity Linking module,Relation Detection module and FAQ module based on question similarity.And it is encapsulated based on Web technology to provide an interface for practical application.