# Graph-based Document Structure Analysis

**Authors:** Yufan Chen, Ruiping Liu, Junwei Zheng, Di Wen, Kunyu Peng, Jiaming Zhang, Rainer Stiefelhagen  
**arXiv:** [2502.02501](https://arxiv.org/abs/2502.02501)  
**Date:** 2025-02-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2502.02501v1/

## Key points

- Authors address limitations of traditional document layout analysis methods
- Introduce graph-based Document Structure Analysis (gDSA) as a novel task
- gDSA generates spatial and logical relations in the form of a graph structure
- Constructed dataset called GraphDoc with 80K document images and 4.13M relation annotations
- Dataset supports tasks like analyzing reading order, hierarchical structures, and complex inter-element relations
- Proposed document relation graph generator (DRGG) achieves 57.6% mAP$_g$@0.5 performance compared to baseline
- Authors believe graphical representation of document structure is an innovative advancement in analysis
- Plan to make dataset and code publicly available at https://yufanchen96.github.io/projects/GraphDoc

## Summary

In their paper titled "Graph-based Document Structure Analysis," authors Yufan Chen, Ruiping Liu, Junwei Zheng, Di Wen, Kunyu Peng, Jiaming Zhang, and Rainer Stiefelhagen address the limitations of traditional document layout analysis (DLA) methods in capturing the nuanced spatial and logical relations between document elements. They introduce a novel task called graph-based Document Structure Analysis (gDSA), which goes beyond basic instance detection to generate spatial and logical relations in the form of a graph structure. This approach aims to enable a deeper comprehension of documents similar to human reading. To support the gDSA task, the authors construct a dataset called GraphDoc, consisting of 80K document images with 4.13M relation annotations. This dataset allows training models to perform tasks such as analyzing reading order, hierarchical structures, and complex inter-element relations. Additionally, they propose a document relation graph generator (DRGG) specifically designed for the gDSA task. The DRGG achieves a performance of 57.6% at mAP$_g$@0.5 when compared to a strong benchmark baseline on this new task and dataset. The authors believe that their graphical representation of document structure marks an innovative advancement in document structure analysis and understanding. They plan to make the new dataset and code publicly available at https://yufanchen96.github.io/projects/GraphDoc. This work was accepted by ICLR 2025 and showcases how graph-based approaches can enhance traditional DLA methods for improved document analysis capabilities.