Research On Chinese Question Answering Technology Based On Knowledge Graph

In recent years,with the rapid development of Natural Language Processing technology,question answering system has become an indispensable function of the mainstream search engines.Based on knowledge graph,KBQA converts the user's natural language problem into structured query language such as SPARQL to query knowledge graph and returns the result to users.However,the current Chinese knowledge graph resources are relatively scarce,and there are some problems in them,such as the unlabeled data type,the triple object not linked by entity and the incomplete classification system,so it is difficult to support complex SPARQL queries,which bring difficulties for the application of question answering system based on Chinese knowledge graph.In order to improve this situation,this paper has carried out research from the following aspects:1.Analyze the existing Chinese knowledge graph resources,and focus on their deficiencies in supporting complex SPARQL queries.Then,based on the analysis results,a method is proposed to support the knowledge graph to support complex SPARQL queries through the fusion of Chinese knowledge graphs.Finally,the knowledge graph is evaluated.Experimental results show that the fusion knowledge graph can support complex SPARQL queries,and effectively improve the problems in the current mainstream Chinese knowledge graph.2.A kind of basic problem of question answering in KBQA system--single entity relation question is studied.Aim at this kind of question,a method of Central entity recognition based on conditional random fields and predicate mapping method combining word vector and string similarity calculation is proposed.The experimental results show that the central entity recognition accuracy and F1 scores of single entity relationship questions are 89.14% and 88.81% respectively.The accuracy and MRR of the question predicate mapping are 93.41% and 95.97% respectively.3.This paper develops a Chinese KBQA prototype system which can answer the question of single entity relationship based on the knowledge graph obtained from the fusion.The sampling test results on NLPCC 2016 KBQA data sets show that the overall accuracy rate of Chinese KBQA constructed by this paper is 84.10%,and the MRR is 86.29%.Through the two important steps of central entity identification and question predicate mapping,the KBQA prototype system constructed in this paper can answer the single entity relation problem of Chinese with higher accuracy.However,due to the lack of relevant data sets,the current system cannot answer complex factoid questions,and the analysis of such problems will be the next step of research work.