Research On Question Answering Over Knowledge Bases With External Texts Based On Graph Attention Network

With the rapid development of Internet technology and the popularization of various electronic products,people are uploading data to the Internet all the time,so that there is a huge amount of data in the Internet.Facing the massive information in the Internet,traditional search engines based on keyword matching and document ranking have been unable to meet people's demand for obtaining information quickly.The question answering system over knowledge base has become the development direction of the next generation search engine because it can understand the user's natural language question and find the answer to the question directly from the knowledge base.Thanks to the development of deep learning and artificial intelligence technology,the question answering system over knowledge has been widely used in automatic question answering,search and other fields and has achieved great successThe question answering system over knowledge base relies on a complete and accurate knowledge base,but the construction of the knowledge base requires a lot of effort.On the other hand,the knowledge base cannot always contain all the knowledge needed to answer user's questions,especially the open-domain questions.However,there are a large number of texts on the Internet,which contain rich content.These text contents can provide additional information for the knowledge base,thereby improving the accuracy of the question answering system.However,the knowledge base is structured data while the text is unstructured data.In order to extract the answer to the questions from knowledge base and external texts,the model should be able to fuse the information from both.In response to this problem,this paper mainly conducted the following research(1)This paper proposes a Graph Neural Network with Question Attention(GQAT)GQAT introduces the problem attention mechanism into the Graph Attention Network.It can determine the importance of the information neighbor node according to the matching degree between the neighbor nodes and the problem.So that GQAT can aggregate answer evidence from the knowledge base and improve the accuracy of the question answering system in the knowledge base(2)Based on the GQAT,this paper presents a model that can integrate knowledge from knowledge base and supplementary texts.supplemented.By adding new entities from supplementary texts to the question subgraph,the content of the knowledge base can be dynamically supplemented.In addition,the semantic information in texts is integrated into the entity representation vector,so that the entity representation vector can merge the knowledge from the knowledge base and the text(3)Multiple experiments is conducted to verify the models proposed in this paper.The experimental results show that the GQAT model can learn knowledge related to the problem from the knowledge base,and the fusion model can make full use of knowledge from both the knowledge base and the text.Compared with other models,models proposed in this paper performance better.