Question Understanding Based On Graph Matching In Question Answering Over Knowledge Base

The rapid development of knowledge graph technology has led to the massive storage of knowledge and information that can be queried.This makes people's path to obtain information no longer depend on traditional search engines based on character and keyword matching,and knowledge base Q&A technology provides new solutions for this.However,the natural language input by people cannot be directly understood by the computer,and natural language need to be understood as a form that can be executed by a computer.The task of understanding natural language questions is divided into two subtasks that information extraction and information disambiguation,the graph matching method studied in this article is to match the similarity between the knowledge subgraph and the question.And for the information extraction task,a joint entity and relation extraction model based on table convolutional neural network is proposed.For the information disambiguation task,combining the idea of granular computing,a Q&A technology based on super semantic graph matching and disambiguation is proposed.A combination of these two methods is proposed to match complex natural language questions containing multiple relationships with knowledge graphs,and turn them into correct question understanding query graph to query the correct answer from the knowledge graph to complete the question answer task.Main contributions of this thesis are follows:1.A joint entity relation extraction model ETC based on table convolutional neural network is proposed,which can jointly extract entities and relations in natural language input.First,the basic feature map of the sentence is generated through neural network and conditional random field.Then,the table convolutional neural network layer proposed in this paper is used to extract the dependency information between word pairs to generate a new feature map.Finally predict the entity type of the word and the relation type between the word pairs by table filling,which avoiding extensive search on the score map.In order to verify the validity of the model,experiments were performed on the human-annotated dataset Co NLL04 and distant supervision dataset NYT,and compared with the current popular methods,ETC has achieved relatively good performance in named entity recognition,relation classification and joint extraction tasks.2.On the task of information disambiguation,a question answering technique based on super semantic graph to matching disambiguation is proposed.This method first extracts information from the question and fuses knowledge graph information to construct a super semantic graph.Then it uses prior knowledge to construct multigranular context features to disambiguate entities in the super semantic graph.Finally,the rank strategy based on semantic matching,literal matching and relation matching gradually performs inference disambiguation and obtains the final understanding result of complex multi-relation questions,which can effectively improve the efficiency and accuracy of the question answering system.And experimental verification on the Chinese Q&A data set CCKS2019-CKBQA and NLPCC-ICCPOL 2016,compared with the optimal methods,without using artificially designed rules and templates or multi-model fusion,only use the simplest model to obtain a better experimental results.