A Research On Question Answering System Based On The Knowledge Graph

The development of the Internet has enriched the channels for people to obtain information,but the efficiency of obtaining effective information has also become lower.In order to solve this problem,the question and answering system based on the knowledge graph(KBQA)has become one of the hotspots of research.Compared with traditional search engines,it can give users accurate and concise answers in real time.A total of four steps are required to construct a question and answer system based on the knowledge graph: named entity recognition,entity linking,attribute selection and answering.Among them,named entity recognition and attribute selection are two key technologies for studying KBQA.After studying the related technologies of KBQA,the thesis firstly conducted a lot of comparative experiments in the named entity recognition module,and selected a good-performance entity recognition model.Secondly,thesis proposed a new attribute selection model MAF?CNN.At the same time,the thesis also proposed a matrix fusion algorithm in the model.Finally,the thesis connected the four major steps and realized a complete KBQA.The specific works of the thesis are as follows:(1)Named entity recognition is an indispensable part of the question and answer system.Because the experimental results of the entity recognition model on the Chinese corpus are not sufficient,several groups of comparative experiments are conducted on the same dataset for the current common models.Finally,the thesis chose the named entity recognition model based on BERT-Bi LSTM-CRF.(2)In the attribute selection module,the thesis proposed an attribute selection model MAF?CNN based on convolutional neural network.In order to solve the problem that a single Chinese word segmentation tool cannot fully extract the input information,the model uses a multi-granular input method.At the same time,a multi-head self-attention mechanism is added between the embedding layer and the convolutional layer to capture the long-distance context information of the sentence,and then match and fuse the question feature vector and the candidate attribute feature vector which are extracted by the convolutional layer and the k-max pooling layer.(3)In order to solve the problem of insufficient similarity feature extraction in the MAF?CNN model,the thesis proposed a matrix fusion algorithm,which can capture similar features of text sequences effectively.And after interactively fusing the matrix,the deeper similarity features are extracted through the convolutional layer and the maximum pooling layer..(4)The thesis implemented the four modules of KBQA and connected them to implement a question and answer system based on Chinese knowledge graphs.The thesis also performed a functional test on the systemThe Q&A system was tested on the NLPCC2016 contest Q&A dataset and compared with the top five results of the contest.The experimental results show that the average F1 value of the Q&A system exceeds the winner solution of the competition,which verifies the effectiveness of the model studied in the thesis.