Research On The Construction Of Classification System And Knowledge Reasoning Of Large-scale Knowledge Graph

With the increase of internet data,knowledge graph has become a powerful tool for knowledge organization,management,and application.The taxonomy,as the backbone of the knowledge graph,has played a crucial role and has achieved significant progress in semantic search,intelligent Q&A,etc.The increasing amount of unstructured text data generated on the internet provides a rich data foundation for taxonomy construction tasks,but also poses challenges to traditional methods due to their low efficiency and accuracy in handling largescale data.Therefore,there is a growing interest in developing faster and more accurate methods for constructing taxonomy.In this thesis,we explore how to efficiently and accurately construct taxonomies,starting from the basic element of the taxonomy,the hypernym relationship,in order to further improve its usability and better serve related tasks in the knowledge graph field.The main contents of this thesis are as follows:(1)This thesis presents a parallel hypernym relationship extraction method.Existing methods for extracting hypernym relationships from text often suffer from low efficiency,especially when dealing with large-scale unstructured text data.In addition,the strict hierarchical relationship of hypernym relationships is often ignored in the extraction process,leading to a large number of erroneous results.To address these issues,we investigate the acceleration effect on large-scale data of a hypernym relationship extraction task based on the Spark distributed memory computing framework.To further improve the accuracy of the extraction results,we analyze and correct erroneous results,including unrecognized inverse hypernym relationships,to improve the accuracy of the extraction.To verify the feasibility and effectiveness of the proposed method,we conducted experiments on Chinese Wikipedia data,tested the accuracy and running time of the proposed method.(2)This thesis presents a novel hypernym relationship reasoning method.Existing relation representation learning methods cannot deeply represent the semantic information in hypernym relationships.Therefore,we propose a hyper-sphere semantic modeling-based hypernym relationship reasoning method using knowledge graph embeddings,which models the hierarchy and transitivity of hypernym relationships.By learning semantic features of hypernym relationships in the taxonomy,we can deeply represent the differences between hyponyms and their hypernyms and use the results for semantic reasoning,thus further improving the robustness of taxonomy.We validate the effectiveness of our method in link prediction and triple classification tasks for hypernym relationships.In summary,the proposed methods for extracting and reasoning over hypernym relationship in this thesis achieve good performance in constructing large-scale taxonomy,and the effectiveness and feasibility of the proposed methods have been proven through experimental analysis.