Research map / Large language models / Retrieval and knowledge
KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge
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 KG-FIT, which builds an LLM-refined entity hierarchy from LLM-generated descriptions and text embeddings, then fine-tunes structure-based KG embeddings with hierarchy, semantic anchoring, and link prediction losses.
Problem
Structure-only KG embeddings ignore open-world knowledge about entities, while PLM-based KGE methods are costly and limited to small models.
Approach
Construct a semantically coherent entity hierarchy via agglomerative clustering over LLM-generated descriptions and text embeddings, refine it with GPT-4o (split, bottom-up merge/include), and use it with text-embedding initialization and constraints to fine-tune any KGE model (global LLM knowledge guiding local KG structure).
Key results
- KG-FIT with LHR improves Hits@10 over base structure models by 12.6% on FB15K-237, 6.7% on YAGO3-10, and 17.7% on PrimeKG (averaged over reported ranges).
- KG-FIT_HAKE reaches H@10 .572 on FB15K-237 vs CSProm-KG .538, and KG-FIT_RotatE reaches MRR .563 on YAGO3-10.
- Training time per epoch on PrimeKG is 1.2m for KG-FIT versus 15m for CSProm-KG, the fastest PLM-based method (about 12x faster).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply KG-FIT with HAKE and the LHR hierarchy instead of standalone HAKE | HAKE · FB15K-237 | Hits@10: 6.1% | Helped |
| Apply KG-FIT with RotatE and the LHR hierarchy instead of standalone TuckER | RotatE · YAGO3-10 | Hits@10: 4.6% | Helped |
| Apply KG-FIT with ComplEx and the LHR hierarchy instead of standalone ComplEx | ComplEx · PrimeKG | Hits@10: 5.8% | Helped |
| Replace the seed hierarchy with LHR (ablation) | KG-FIT-TransE · FB15K-237 | Hits@10: .547 -> .568 | Helped |
| Replace the seed hierarchy with LHR (ablation) | KG-FIT-RotatE · YAGO3-10 | Hits@10: .708 -> .722 | Helped |
| Replace the seed hierarchy with LHR (ablation) | KG-FIT-ComplEx · PrimeKG | Hits@10: .823 -> .835 | Helped |
| Add Hierarchical Distance Maintenance (ablation) | KG-FIT-TransE · FB15K-237 | MRR: .345 -> .362 | Helped |
| Add Hierarchical Distance Maintenance (ablation) | KG-FIT-HAKE · YAGO3-10 | MRR: .558 -> .568 | Helped |
| Add Semantic Anchoring alongside HDM (ablation) | KG-FIT-TransE · FB15K-237 | MRR: .335 -> .345 | Helped |
| Add Semantic Anchoring alongside HDM (ablation) | KG-FIT-HAKE · YAGO3-10 | MRR: .545 -> .558 | Helped |
| Add Inter-level Cluster Separation alongside HDM (ablation) | KG-FIT-TransE · FB15K-237 | MRR: .343 -> .345 | Helped |
| Add Inter-level Cluster Separation alongside HDM (ablation) | KG-FIT-HAKE · YAGO3-10 | MRR: .544 -> .558 | Helped |
Surprising findings
- KG-FIT does not outperform pure PLM-based methods (C-LMKE, SimKGC) on the lexical WN18RR dataset; combining with PKGC as a recall model was needed to be competitive.
- Zero-shot LLM text-embedding link prediction is very weak (e.g., MRR .023-.030 on FB15K-237), showing raw LLM embeddings alone are not useful for link prediction.
Limitations
- Performance depends on quality of the seed hierarchy; suboptimal clustering can lower quality.
- LHR applies only a single operation per parent-child triple, trading off performance for efficiency.
- Does not outperform strong PLM-based contrastive methods on lexical WN18RR.
Experimental setup
Datasets: FB15K-237 YAGO3-10 PrimeKG WN18RR
Benchmarks: FB15K-237 YAGO3-10 PrimeKG WN18RR
Metrics: MR MRR Hits@1 Hits@5 Hits@10
Baselines: KG-BERT StAR PKGC C-LMKE KGT5 KG-S2S SimKGC CSProm-KG TransE DistMult ComplEx ConvE
Builds on: TransE RotatE HAKE PKGC SimKGC
Closest papers by meaning
- KGDM: A Diffusion Model to Capture Multiple Relation Semantics for Knowledge Graph Embedding (AAAI 2024 · similarity 0.90)
- MKGL: Mastery of a Three-Word Language (NeurIPS 2024 · similarity 0.89)
- K-ON: Stacking Knowledge on the Head Layer of Large Language Model (AAAI 2025 · similarity 0.89)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.89)
- Select and Augment: Enhanced Dense Retrieval Knowledge Graph Augmentation (Abstract Reprint) (AAAI 2024 · similarity 0.89)
- Editing Language Model-Based Knowledge Graph Embeddings (AAAI 2024 · similarity 0.88)
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.