Research On The Techniques Of Knowledge Graph Based Chinese Multi-Hop Question Answer

With the increasing of data in various fields on the Internet,how to use these data or knowledge to provide users with convenient automatic question answering system has become a major research topic.Knowledge map can effectively connect the disordered knowledge of the Internet and facilitate query and reasoning.Because of its simplicity and rapidity,knowledge map is becoming the main knowledge source of automatic question answering.With the establishment of more and more high-quality knowledge map,the knowledge base of automatic question answering technology has been expanding.Now the problem is how to build a good supervision corpus,and how to train a good automatic question answering system based on knowledge map.After the popularity of deep neural network,it has achieved good results in the field of knowledge map Q & A.Now the mainstream knowledge base question answering basically adopts deep learning method,and the process mainly includes entity recognition,entity link,attribute recognition and query statement generation.However,as the question of Chinese knowledge base Q & a data set becomes more and more complex,the question involves more and more vertical fields,and the simple entity recognition model and attribute recognition are still limited.Based on the above questions,this paper carries out specific research from the following aspects:(1)Aiming at the problem of entity reference recognition,this paper uses the method of combining the segmentation model based on domain entity dictionary and the entity recognition model based on deep learning to recognize the entity reference,which can more effectively identify the entity reference in domain problems.(2)For the problem of entity linking,this paper recalls the candidate entities in the knowledge base as comprehensively as possible,and then mines some features of the candidate entities mentioned and candidate entities,and uses these features to train a logistic regression model,and uses the model to score each candidate entity,so as to select the top n candidate entities.(3)Aiming at the complex multi hop problem,based on the statistical analysis of the data set,this paper proposes some problem templates,trains three problem classification models according to the templates,and generates candidate query paths according to the classification of problems,which can greatly reduce the number of generated query paths,and improve the efficiency and accuracy of the model.(4)In order to solve the problem of path matching,a path attention siamese network based on the pre-training model is constructed to solve the difference between the candidate query path and the expression of the problem.