Research map / Graph learning / Knowledge and temporal graphs
Knowledge Graph Error Detection with Contrastive Confidence Adaption
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CCA, an end-to-end KG error detector that reconstructs masked triplet entities with a BERT text encoder and a Transformer structure encoder, aligns them via interactive contrastive learning, and fuses scores into adaptive pseudo-label confidences; also builds semantically-similar and adversarial noise benchmarks on FB15K-237 and WN18RR.
Problem
Detecting errors in knowledge graphs is hard because existing structure-based detectors struggle to distinguish realistic noise, especially semantically-similar triplets, from correct ones; they under-use textual information.
Approach
Combine a PLM text encoder and a Transformer graph encoder that both reconstruct masked head/tail entities, align their projections with interactive contrastive learning (with error-triplet negatives), and use rank-fused scores as adaptive triplet confidences that reweight training losses.
Key results
- On FB15K-237 with 5% noise, CCA reaches precision@top-1% of 0.969 and precision@top-5% of 0.534, the best among the compared models.
- On WN18RR, CCA reaches precision@top-5% of 0.733, versus 0.680 for CSProm-KG.
- On FB15K-237 semantically-similar noise, CCA attains precision@top-5% of 0.453 versus 0.419 for the next best (CSProm-KG).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine textual and graph structural information with reconstruction and interactive contrastive learning | CCA · FB15K-237, semantically-similar noise, top-5% | precision@top-5%: 0.419 -> 0.453 | Helped |
| combine textual and graph structural information with reconstruction and interactive contrastive learning | CCA · FB15K-237, adversarial noise, top-5% | precision@top-5%: 0.211 -> 0.240 | Helped |
| combine textual and graph structural information with reconstruction and interactive contrastive learning | CCA · WN18RR, adversarial noise, top-5% | precision@top-5%: 0.599 -> 0.545 | Hurt |
| add adaptive confidence (ablation) | CCA · FB15K-237, top-1% | precision@top-K: .951 -> .969 | Helped |
| add adaptive confidence (ablation) | CCA · WN18RR, top-1% | precision@top-K: .971 -> .986 | Helped |
| add interactive contrastive learning (ablation) | CCA · FB15K-237, top-1% | precision@top-K: .961 -> .969 | Helped |
| add interactive contrastive learning (ablation) | CCA · WN18RR, top-1% | precision@top-K: .979 -> .986 | Helped |
| add structure reconstruction (ablation) | CCA · FB15K-237, top-1% | precision@top-K: .797 -> .969 | Helped |
| add structure reconstruction (ablation) | CCA · WN18RR, top-1% | precision@top-K: .974 -> .986 | Helped |
| add text reconstruction (ablation) | CCA · FB15K-237, top-1% | precision@top-K: .778 -> .969 | Helped |
| add text reconstruction (ablation) | CCA · WN18RR, top-1% | precision@top-K: .577 -> .986 | Helped |
Surprising findings
- On WN18RR adversarial noise, KG-BERT (0.599) outperforms CCA (0.545).
- Removing structure reconstruction barely changes WN18RR precision (0.974 vs 0.986 full at top-1%), suggesting the sparse graph contributes little there.
Limitations
- Gains on WN18RR are small because the graph is sparse, limiting the structural encoder's contribution.
- Adversarial noise is generated only with TransE, so it may favor structure-based construction errors.
Experimental setup
Datasets: FB15K-237 WN18RR
Benchmarks: FB15K-237 WN18RR
Metrics: precision@top-K recall@top-K
Baselines: TransE DistMult ComplEx KGTtm CAGED KG-BERT StAR CSProm-KG
Builds on: KG-BERT CAGED TransE InfoNCE
Closest papers by meaning
- S²DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.87)
- Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning (NeurIPS 2026 · similarity 0.85)
- Knowledge Graphs Can be Learned with Just Intersection Features (ICML 2024 · similarity 0.85)
- Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment (ICML 2025 · similarity 0.85)
- GraphVis: Boosting LLMs with Visual Knowledge Graph Integration (NeurIPS 2024 · similarity 0.84)
- Improved Graph Contrastive Learning for Short Text Classification (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.