Research map / Graph learning / Knowledge and temporal graphs
Select and Augment: Enhanced Dense Retrieval Knowledge Graph Augmentation (Abstract Reprint)
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 DRKA, a multi-task framework where a dense retriever (SBERT-based) selects top-k relevant text descriptions for KG triples and aligns them with KG embeddings, improving link and relation prediction on FB15K.
Problem
Text-enhanced knowledge graph embeddings typically rely on a single entity description, which cannot sufficiently represent an entity's context and lexical ambiguity; the paper asks whether selecting multiple relevant descriptions improves KG completion.
Approach
Jointly train a dense retriever that selects top-k text descriptions for a triple (query formed by concatenating triple embeddings) and an attention-based Text-KG aligner that fuses SBERT-encoded descriptions into KG embeddings, with a joint retrieval plus margin-based alignment loss.
Key results
- DRKA + RotatE (k=5) reaches 42.7 MRR and 55.7 Hits@10 overall on FB15K link prediction, versus 38.8/53.1 for RotatE alone and 40.6/54.8 for DKRL(CNN)+RotatE.
- Relation prediction: DRKA + RotatE achieves mean rank 1.1 and Hits@1 93.4, versus 3.5 and 89.0 for RotatE alone.
- Varying k shows gains rise up to about k=16 and decline beyond, per the paper's analysis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dense SBERT retriever with joint retrieval and alignment loss (DRKA) over DKRL-style CNN text | RotatE · FB15K link prediction, Overall | MRR: 40.6 -> 42.7 | Helped |
| remove retriever (random relevant descriptions) from DRKA (ablation) | DRKA + TransE k=4 · FB15K link prediction, Overall | Hits@10: 54.8 -> 51.5 | Hurt |
| remove retriever from DRKA (ablation) | DRKA + TransE k=4 · FB15K link prediction, Overall | MRR: 40.7 -> 39.8 | Hurt |
| decouple DPR retriever from alignment (DRKA(DPR)) (ablation) | TransE and DistMult · FB15K link prediction, Overall | MRR: 41.2 -> 38.2 (TransE) | Hurt |
| increase number of augmenting descriptions k from 1 toward 16 (ablation) | DRKA · FB15K link prediction, Hits@10 | Hits@10 | Helped |
| increase k beyond 16 (up to 20) (ablation) | DRKA · FB15K link prediction, Hits@10 | Hits@10 | Hurt |
| text-enhanced augmentation with dense retrieved descriptions vs. KG-only embeddings | TransE, DistMult, CompIEx, RotatE · FB15K relation prediction | MR and Hits@1: RotatE MR 3.5 -> DRKA + RotatE 1.1 | Helped |
Surprising findings
- DRKA(DPR) decoupled from alignment performs worse than the KG-only baselines in several settings, which the authors attribute to error propagation.
- Replacing the retriever with random relevant descriptions lowers link prediction performance, especially Hits@10.
Limitations
- Performance declines for k above about 16, likely from overlapping descriptions confusing the model.
- Decoupled DRKA(DPR) variant suffers error propagation between retrieval and alignment.
- Abstract-only reprint; full text contains inconsistent numbers (e.g., 5.5% vs 6% MRR gain).
Experimental setup
Datasets: FB15K BabelNet Google News Wikipedia articles
Benchmarks: FB15K
Metrics: MRR Hits@10 Mean Rank Hits@1 triplet classification accuracy
Baselines: TransE DistMult ComplEx RotatE DKRL(CNN) DRKA(DPR)
Builds on: TransE DPR DKRL SBERT
Closest papers by meaning
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.91)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.89)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.89)
- KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge (NeurIPS 2024 · similarity 0.89)
- K-ON: Stacking Knowledge on the Head Layer of Large Language Model (AAAI 2025 · similarity 0.88)
- MKGL: Mastery of a Three-Word Language (NeurIPS 2024 · 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.