Trustworthiness and path regularity based contrastive learning for noisy knowledge graph error assertion detection

In large-scale Knowledge Graphs (KGs), the vast volume of data and continuous updates often lead to the accumulation of numerous erroneous information, thereby undermining the trustworthiness of these KGs. With the rapid advancement of AI, the demand for high-quality KGs has increased substantially. This paper proposes a method named TP-contrast for detecting erroneous assertions in KGs which may not be completely trustworthy (i.e., containing noise). TP-contrast employs a multi-module framework that combines trustworthiness assessment, path resource planning, and contrastive learning techniques. Our approach has been rigorously validated on a wide range of datasets including three real-world KGs, demonstrating significant effectiveness and outperforming existing state-of-the-art baselines.