KGER: Knowledge Graph Error Detection and Refinement with Reinforcement Learning

作者：Aishan Maoliniyazi, Chaohong Ma, Xiaofeng Meng · 年份：2026 · DOI：10.1109/icassp55912.2026.11463012 · 研究领域：Advanced Graph Neural Networks、Graph Theory and Algorithms、Multimodal Machine Learning Applications

Knowledge Graphs (KGs) are essential for AI applications. However, pervasive KG errors substantially impact their performance. Existing error correction methods often distinguish detection from refinement, underutilize KG embeddings, and largely depend on scarce labeled data for generalization and scalability. This paper introduces KGER, a novel approach for knowledge graph error detection and refinement with reinforcement learning. KGER frames KG detection and refinement as a unified reinforcement learning process. It employs a Deep Q-Network (DQN) agent that integrates KG embeddings and uses a staged training approach. This allows it to learn robust correction policies applicable across varying levels of supervision, from supervised to unsupervised settings. Evaluations on various datasets show that KGER effectively identifies and refines erroneous KG triples, enhances knowledge graph quality, and outperforms existing methods. The code and data are available at https://github.com/kazgu/KGER.