SeSICL: Semantic and Structural Integrated Contrastive Learning for Knowledge Graph Error Detection

As knowledge graphs (KGs) become more widely used in various applications, error detection for KGs has received more attention, which can reduce quality issues such as errors and inconsistencies. With the development of representation learning, embedding-based methods have significantly improved error detection performance. The recent error detection algorithm uses KG structural embedding loss and constructs a reasonable score function, ranking the confidence scores for each triplet. However, these methods ignore the factual semantics of the triplet itself, which primarily exist in the entities and relations descriptions text. Therefore, we propose Semantic and Structural Integrated Contrastive Learning(SeSICL) to simultaneously capture graph structural patterns and deep semantic features from descriptions text. Our method is based on contrastive learning without data augmentation, which utilizes encoder perturbations to generate contrasting views, making SeSICL highly suitable for complex error detection tasks and robust against real-world noise. We evaluate SeSICL on three baseline datasets with abnormal data and fluctuations. SeSICL outperforms the previous state-of-the-art methods, demonstrating our method’s performance and robustness in more complex scenarios.