Research map / Graph learning / Knowledge and temporal graphs
Graph-based Document Structure Analysis
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 graph-based document structure analysis (gDSA), the GraphDoc dataset (80K DocLayNet pages with 4.13M spatial/logical relation annotations), and DRGG, a relation head added to DETR-style layout detectors.
Problem
Document layout analysis only detects elements and misses spatial and logical relations between them, so models cannot reach deeper, human-like structural understanding of documents.
Approach
Reframe document understanding as generating a relation graph over detected layout elements, with a plug-and-play relation head (DRGG) that pools decoder object queries into pairwise relation features, plus threshold-based mR_g/mAP_g metrics that tie relation scores to detection quality.
Key results
- DRGG with InternImage backbone and RoDLA detector reaches mAP_g@0.5 of 57.6% and mR_g@0.5 of 30.7% on GraphDoc gDSA.
- The same configuration reaches 81.5% DLA mAP@50:5:95, above the 80.5% RoDLA baseline without DRGG.
- Reference relations remain hard, with best AP of 18.8% (ResNeXt with RoDLA) and 16.8% for the InternImage configuration.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RoDLA instead of DINO with InternImage and DRGG | DRGG · GraphDoc | mAP_g@0.5: 25.2 -> 57.6 | Helped |
| Use RoDLA instead of DINO with InternImage and DRGG | DRGG · GraphDoc | mR_g@0.5: 19.2 -> 30.7 | Helped |
| Use InternImage instead of ResNeXt with RoDLA and DRGG (ablation) | RoDLA + DRGG · GraphDoc | mAP@50:5:95: 77.9 -> 81.5 | Helped |
| Use InternImage instead of Swin with RoDLA and DRGG (ablation) | RoDLA + DRGG · GraphDoc | mAP@50:5:95: 73.7 -> 81.5 | Helped |
| Use InternImage instead of ResNet with RoDLA and DRGG (ablation) | RoDLA + DRGG · GraphDoc | mAP@50:5:95: 71.0 -> 81.5 | Helped |
| Use RoDLA instead of DINO with InternImage and DRGG (ablation) | InternImage + DRGG · GraphDoc | mAP@50:5:95: 79.5 -> 81.5 | Helped |
| Add DRGG relation head to InternImage + RoDLA | InternImage + RoDLA · GraphDoc | mAP@50:5:95: 80.5 -> 81.5 | Helped |
| Use RoDLA instead of DETR with InternImage and DRGG (ablation) | InternImage + DRGG · GraphDoc | mAP_g@0.5: 19.8 -> 57.6 | Helped |
Surprising findings
- Reference relation AP stays below 20% across all configurations, so the rarest logical relation remains largely unsolved.
Limitations
- Model uses only visual input, without text or multimodal input.
- Dataset and approach target single-page documents; multi-page relations are not covered.
- Reference relation performance is low, and the dataset is strongly imbalanced across relation types.
Experimental setup
Datasets: GraphDoc DocLayNet
Benchmarks: GraphDoc
Metrics: mAP@50:5:95 mR_g@0.5 mAP_g@0.5 mAP_g@0.75 mAP_g@0.95
Baselines: DETR Deformable DETR DINO RoDLA
Builds on: DETR RoDLA DocLayNet Recursive X-Y Cut
Closest papers by meaning
- Relation-Rich Visual Document Generator for Visual Information Extraction (CVPR 2025 · similarity 0.86)
- A Simple yet Effective Layout Token in Large Language Models for Document Understanding (CVPR 2025 · similarity 0.86)
- SRFUND: A Multi-Granularity Hierarchical Structure Reconstruction Benchmark in Form Understanding (NeurIPS 2024 · similarity 0.85)
- TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning (NeurIPS 2026 · similarity 0.85)
- Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation (NeurIPS 2026 · similarity 0.85)
- MMLONGBENCH-DOC: Benchmarking Long-context Document Understanding with Visualizations (NeurIPS 2024 · similarity 0.85)
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.