Reproduce “Contrastive Knowledge Graph Error Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qinggang Zhang, Junnan Dong, Keyu Duan, Xiao Huang, Yezi Liu, Linchuan Xu
Researchers from The Hong Kong Polytechnic University and UC Irvine introduced CAGED, a contrastive learning framework for Knowledge Graph error detection that models triples as nodes in augmented hyper-views and incorporates an error-aware graph encoder. The method consistently surpassed existing state-of-the-art baselines, demonstrating up to 7.3% higher precision/recall on the WN18RR dataset and effectively identifying subtle, semantically relevant yet mismatched errors.
Key insights
- Treating KG triples as nodes and generating specific 'hyper-views' enables effective contrastive learning at a granular, triple-level, which is more suitable for error detection than general graph augmentations.
- An error-aware gated attention mechanism within the EaGNN is crucial for learning robust triple representations by actively mitigating the propagation of noisy information from potentially erroneous neighbors.
- Combining both local (self-consistency via KG embedding) and global (cross-view consistency via contrastive learning) signals is essential for comprehensively identifying a wide range of KG errors, including subtle, complex ones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.