Research map / Graph learning / Knowledge and temporal graphs
Editing Language Model-Based Knowledge Graph Embeddings
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.
Introduces the task of editing language-model-based KG embeddings (EDIT and ADD), builds four datasets from FB15k-237 and WN18RR, and proposes KGEditor, a hypernetwork that adds parametric FFN layers to edit facts.
Problem
Language-model-based KG embeddings are static after deployment, so correcting wrong facts or adding new facts requires costly re-training; no method existed for editing them efficiently while preserving other knowledge.
Approach
Treat KG embedding updates as a learning-to-update problem and add an extra FFN-shaped layer whose weight shift is generated by a BiLSTM hypernetwork from the (x, y, a) edit, so only a small editor is trained while the rest of the model is untouched.
Key results
- On E-FB15k237, KGEditor reaches Succ@1 0.866 versus 0.828 for MEND and 0.702 for KE, with RK@3 0.874 (KE 0.912 is higher).
- On A-FB15k237, KGEditor achieves Succ@1 0.796 versus 0.714 for CALINET and 0.648 for KE, with RK_roc 0.920.
- KGEditor uses 38.9M params on E-FB15k237 versus 88.9M for KE and 59.1M for MEND, with editing time 0.226 vs 0.368 (KE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KGEditor instead of MEND | PT-KGE · E-FB15k237 | Succ@1: 0.828 -> 0.866 | Helped |
| Use KGEditor instead of KE | PT-KGE · E-FB15k237 | RK@3: 0.912 -> 0.874 | Hurt |
| Use KGEditor instead of MEND | PT-KGE · E-WN18RR | Succ@1: 0.815 -> 0.833 | Helped |
| Use KGEditor instead of MEND | PT-KGE · E-WN18RR | RK@3: 0.957 -> 0.956 | Hurt |
| Use KGEditor instead of CALINET | PT-KGE · A-FB15k237 | Succ@1: 0.714 -> 0.796 | Helped |
| Use KGEditor instead of KE | PT-KGE · A-FB15k237 | RK@3: 0.926 -> 0.899 | Hurt |
| Use KGEditor instead of MEND | PT-KGE · A-WN18RR | Succ@1: 0.999 -> 0.998 | Hurt |
| Use KGEditor instead of KE | PT-KGE · A-WN18RR | RK@3: 0.975 -> 0.956 | Hurt |
Surprising findings
- Fine-tuning all parameters (KGE_FT) or a subset (K-Adapter) maximizes reliability on ADD but sharply degrades previously acquired knowledge.
- Prompt-based KGE (PT-KGE) yields better editing performance than fine-tuning-based KGE (FT-KGE).
Limitations
- Experiments use only small BERT-based KGE models; no evaluation on large LLMs.
- Handling many-to-many relations remains difficult; edits can bias the model toward the edited entity.
- KE and MEND are adapted to the task by removing equivalence sets, so baseline fidelity is uncertain.
Experimental setup
Datasets: FB15k-237 WN18RR
Benchmarks: E-FB15k237 A-FB15k237 E-WN18RR A-WN18RR
Metrics: Succ@1 Succ@3 ER_roc RK@3 RK_roc Time Params
Baselines: KE MEND CALINET KGE_FT KGE_ZSL K-Adapter
Builds on: Knowledge Editor (KE) MEND CALINET KG-BERT PKGC
Closest papers by meaning
- KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge (NeurIPS 2024 · similarity 0.88)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.88)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Should We Really Edit Language Models? On the Evaluation of Edited Language Models (NeurIPS 2024 · similarity 0.88)
- Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis (NeurIPS 2026 · similarity 0.87)
- SWEA: Updating Factual Knowledge in Large Language Models via Subject Word Embedding Altering (AAAI 2025 · similarity 0.87)
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.