Research On Multi-source Knowledge Quality Evaluation In Knowledge Graphs

In recent years,applications based on knowledge graphs(KGs)have been extensively researched and received considerable attention in the computer science domain.KGs typically organize the concepts,entities,and their relationships from the real world into a graphical structure as triples,facilitating tasks such as knowledge storage,reasoning,and computation.Currently,they are broadly applied in areas like semantic search,personalized recommendation,and social networking.As application demands grow,obtaining knowledge from a single source might not meet the needs due to the limited amount and coverage of knowledge within a single source.Consequently,acquiring knowledge from a diverse array of sources has become a primary method for knowledge acquisition.However,the quality of knowledge sourced from multiple origins can vary significantly due to factors like the presence of low-quality sources and the limitations of knowledge extraction algorithms.This variance in quality can lead to conflicts or the presence of erroneous information,severely affecting the performance of downstream applications.Despite numerous research efforts proposing methods for verifying knowledge and controlling its quality at different stages of knowledge graph construction,these methods often struggle to effectively validate vast amounts of knowledge from diverse sources,especially in complex knowledge source scenarios.Furthermore,many existing methods usually rely on specific domain information or rules to verify or evaluate knowledge,making it difficult to extend these approaches to other fields.Therefore,exploring a method for the validation and evaluation of multi-source knowledge that doesn’t depend on specific domain information or expert rules holds significant importance.This thesis aims to build a knowledge quality evaluation scheme suitable for multi-source scenarios,grounded on an in-depth analysis of current methods for controlling knowledge graph quality.The approach starts from perspectives such as knowledge sources,graph patterns,and cross-modal validation,addressing the main challenges and problems in this research area by leveraging the consistency of factual statements among multi-source knowledge.Firstly,this thesis constructs generalized quality evaluation methods for unverified knowledge in multi-source scenarios,which uses the consistency of knowledge in different sources to evaluate knowledge quality.It aims to simplify the process of multi-source knowledge quality evaluation,making the process independent of domain-specific information.Secondly,by combining the common information of multi-source knowledge,such as the prior quality of sources,the structure of KGs,etc.,the quality of knowledge could be evaluated quantitatively and accurately.Specifically,a knowledge quality evaluation method based on the fusion of prior and consistency is proposed.Based on the consistency of multi-source knowledge,the quality of different knowledge sources is integrated into the quality evaluation process of knowledge elements,so that knowledge quality can be updated with a balance of a prior and consistency.In addition,for triples from different KGs,this thesis incorporates the internal consistency of triples from different KGs into the knowledge quality evaluation process.A knowledge quality evaluation method based on the fusion of internal and external consistency is designed.Furthermore,this research explores a knowledge quality evaluation method based on cross-media knowledge verification,constructing a data validation-based knowledge quality evaluation scheme.This method refines the correlation and causal information between attributes in statistical data corresponding to knowledge,aiding in the automated assessment of knowledge quality.Finally,the core idea of knowledge quality evaluation based on the consistency of multi-source information is integrated into various application tasks and scenarios to demonstrate its utility.Tailored methods are designed for different task requirements and scenarios,including integrating the multi-source knowledge quality evaluation method with blockchain architecture to design a shared knowledge graph management framework for collaborative construction by multiple contributors,ensuring data security and quality during the collaborative building and sharing process of knowledge graphs.Additionally,beyond assessing the quality of triples in knowledge graphs,the proposed concept is applied to scenarios involving the integration of multi-label knowledge,designing a high-quality annotation knowledge filtering framework for multisource labels.This framework selects and filters high-quality data labels among multisource annotated knowledge,thereby ensuring the performance of downstream learning models.Experimental results in each chapter indicate that the proposed methods exhibit good application effects in various multi-source knowledge scenarios.In the applications of the two different tasks proposed in this thesis,the designed knowledge quality evaluation methods can serve as core algorithms for selecting trustworthy knowledge from multiple sources and filtering out low-quality information.