Research And Application Of Multi-Hop Question Answering Method Based On Knowledge Graph Embedding

With the development of network technology,more and more information appears in the internet,which contains a lot of useless knowledge interfering people’s access to key information and reducing work efficiency.Knowledge graph question answering system is a new technology emerging from the wave of artificial intelligence that meets the demand of users for easy access to information,solves user problems timely and improves the efficiency.With the rapid development of deep learning technology,significant breakthroughs have been made in various research tasks in the field of natural language processing.Existing knowledge graph question answering systems have been able to answer simple questions accurately and directly to meet some simple requirements in people’s daily lives.But users prefer to express specific information in complex questions when consulting them.Existing methods still perform poorly in dealing with such complex interrogatives.How to reason effectively about complex questions in knowledge graphs is the main challenge for knowledge graph question answering systems.This thesis improves the existing knowledge graph question answering methods to address the above issues,enhance the reasoning ability of the methods and improve their Q&A accuracy.The main work and contributions of this thesis are as follows:(1)An inference model-NRQA incorporating neighbor interaction networks and relation recognition module is proposed for multi-hop question answering tasks.The model focuses on solving complex multi-hop problems based on the conditions of sparse knowledge graph by selectively capturing the hidden information in the knowledge graph and overcoming the limitation of answer scope.In addition,NRQA proposes a relation recognition module to make full use of the triadic relational information contained in the question to further filter the candidate entities.Experimental results on two publicly available datasets show that the model can effectively capture the semantic information of the knowledge graph for inference and achieve better results than the baseline model.(2)A graph contrastive learning knowledge graph embedding model-GCL-KGE is proposed.The model uses both graph attention networks and contrastive learning algorithms to aggregate multi-order neighbor information to optimize entity representations.The model consists of two main components,an encoder-decoder framework and a contrastive learning algorithm.The encoder-decoder framework uses the graph attention network and a triple scoring function to evaluate the knowledge graph embedding in the link prediction task.To avoid interfering with the entity representation by noisy information in the stack of the graph attention networks,the model uses the contrastive learning algorithm to provide auxiliary supervised signals to optimize the embedding representation of the knowledge graph.Finally,results on four publicly available datasets demonstrate the effectiveness of the GCL-KGE in the knowledge graph embedding task and show that the contrastive learning algorithm is able to attenuate the noise interference introduced during the deepening of the graph neural network.(3)By analyzing the actual demand scenario,this thesis designs and implements a question answering system based on knowledge graphs in the film field.The two models proposed in this thesis are applied to the film question answering system to assist the system to complete the query task.The system is built using flask framework and the whole system is mainly divided into three parts: data module,knowledge question answering module and front-end display module.The test results of all aspects of the system show that the accuracy and stability of the answers returned by the system meet the actual needs of users.