Research map / Graph learning / Knowledge and temporal graphs
Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction
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.
Reframes document-level relation extraction as knowledge-graph link prediction, combining Wikidata/WordNet entity context and context paths with document reasoning (DRN-style) and path-based beam search for explanations.
Problem
Document-level relation extraction requires inferring relations between entities dispersed across sentences, where document evidence alone is often latent or imprecise and existing DocRE datasets have missing triples.
Approach
Treat DocRE as link prediction over a KG built from extracted triples, WordNet entity synonyms/types, and Wikidata N-hop context paths, scored by an R-GCN encoder with DistMult decoder, then aggregate with a multi-path reasoning module (intra-sentence, logical, co-reference) and explain predictions via path-based beam search.
Key results
- On DocRED test, DocRE-CLiP reaches 68.51 F1 (66.31 Ign F1), about 1 point above DREEAM (67.53) and KD-DocRE (67.28).
- On ReDocRED test, DocRE-CLiP achieves 81.55 F1 versus DREEAM's 81.44.
- On DWIE test, DocRE-CLiP reaches 67.10 F1 versus RECON's 66.94.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Wikidata and WordNet context and link prediction to the reasoning framework | DocRE-CLiP · DocRED test | F1: 67.53 -> 68.51 | Helped |
| Add Wikidata and WordNet context and link prediction to the reasoning framework | DocRE-CLiP · ReDocRED test | F1: 81.44 -> 81.55$_{\pm 0.14}$ | Helped |
| Add Wikidata and WordNet context and link prediction to the reasoning framework | DocRE-CLiP · DWIE test | F1: 66.94 -> 67.10$_{\pm 0.11}$ | Helped |
| Add context to R-GCN link prediction (ablation) | R-GCN · DocRED | Hits@1: 0.06 -> 0.11 | Helped |
| Add context to R-GCN link prediction (ablation) | R-GCN · ReDocRED | Hits@1: 0.033 -> 0.51 | Helped |
| Add context to R-GCN link prediction (ablation) | R-GCN · DWIE | Hits@1: 0.38 -> 0.67 | Helped |
| Incorporate context, including document triplets, entity context, and context paths (ablation) | DocRE-CLiP · DocRED, ReDocRED, DWIE | F1 | Helped |
| Add entity context such as aliases, descriptions, and song information (ablation) | DocRE-CLiP · Case study, TY.O–Troublemaker relation | relation prediction accuracy | No clear effect |
Surprising findings
- The ReDocRED gain over DREEAM is marginal (about 0.1 F1), which the authors attribute to ReDocRED already correcting dataset incompleteness.
Limitations
- Gains over the strongest baselines are small (about 1 F1 on DocRED, 0.1 on ReDocRED).
- Entity context from WordNet can add little when it is not informative, as in the failure case of entity context.
- Some link prediction MRR/Hits values in the reported tables are inconsistent or exceed 1, so the link prediction evidence is weakly supported.
Experimental setup
Datasets: DocRED ReDocRED DWIE Wikidata WordNet
Benchmarks: DocRED ReDocRED DWIE
Metrics: F1 Ign F1 Hits@1 Hits@3 Hits@10 MRR
Baselines: SIRE HeterGSAN-Rec ATLOP DRN DocuNet KD-DocRE DREEAM EIDER SAIS KIRE RESIDE RECON
Builds on: DRN R-GCN DistMult ATLOP
Closest papers by meaning
- DocKS-RAG: Optimizing Document-Level Relation Extraction through LLM-Enhanced Hybrid Prompt Tuning (ICML 2025 · similarity 0.89)
- COMM: Concentrated Margin Maximization for Robust Document-Level Relation Extraction (AAAI 2025 · similarity 0.89)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.88)
- An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction (AAAI 2024 · similarity 0.87)
- A Hierarchical Network for Multimodal Document-Level Relation Extraction (AAAI 2024 · similarity 0.86)
- Anchoring Path for Inductive Relation Prediction in Knowledge Graphs (AAAI 2024 · similarity 0.86)
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.