DE4KG: A Comprehensive Framework for Detecting and Eliminating Low-Quality Issues in Domain-Specific Knowledge Graphs
Kai Zhang, Yuchen Li, Zhiying Tu, Wenlong Meng, Hongliang Sun, Yongchao Xing, Bohai Zhao, Dianhui Chu, Yuhao Zhang · Data Intelligence · 2025
With the substantial accumulation of data from the modern internet and the rapid development of artificial intelligence technologies, knowledge graphs (KGs), as a structured method of knowledge representation, have been widely applied across various fields. They play a particularly critical role in the representation and reasoning of domain-specific knowledge within vertical industries. However, KGs may suffer from issues such as logical errors, data inconsistencies, and a lack of domain context, which severely constrain the accuracy of reasoning and decision-making processes. To this end, we propose a framework for detecting and eliminating low-quality issues in domain-specific KGs (DE4KG). Firstly, data logic problems such as structural rule problems and link errors are solved through solutions based on statistical and graph algorithms and multiview comparative learning. Secondly, for data consistency issues, such as entity co-reference and relationship co-reference, we propose solutions that combine graph neural networks with semantic matrices and structural semantic clustering. These methods progressively perform comprehensive detection and elimination of both types of issues in low-quality graphs, ultimately producing high-quality domain KGs. Moreover, we introduce retrieval augmented generation (RAG) to support the detection and correction of low-quality KGs. Finally, to facilitate the visualization of these methods, we design a full-process system for detecting and eliminating low-quality issues in domain KGs, which helps builders quickly identify and correct low-quality problems.