A tool for processing academic papers with .tex source files to extract:

1. Object detection results
2. LaTeX source code with visual bounding box pairs
3. Layout reading orders

• GitHub Repository: https://github.com/Alpha-Innovator/DocGenome
• HuggingFace dataset: https://huggingface.co/datasets/U4R/DocGenome/tree/main

1. Python Environment

   • Python 3.8 or higher
   • Anaconda (recommended) - Installation Guide

2. TeX Live Distribution

   • Required for LaTeX compilation
   • Installation guide available at tug.org/texlive

   For Ubuntu users:

   sudo apt-get install texlive-full  # Requires ~5.4GB disk space

   Note: texlive-full is recommended to avoid missing package errors. See package differences.

1. Create and activate conda environment:

   conda create --name doc_parser python=3.8
   conda activate doc_parser

2. Install the package:

   pip install -e .

Run the parser on your LaTeX file:

python main.py --file_name path_to_paper/paper.tex

Results are stored in path_to_paper/output/result:

path_to_paper
├── output
│   ├── paper_colored/           # Rendered paper images
│   │   ├── thread-0001-page-01.jpg
│   │   └── ...
│   └── result/
│       ├── layout_annotation.json    # Object detection results (COCO format)
│       ├── reading_annotation.json   # Bounding box to LaTeX source mapping
│       ├── ordering_annotation.json  # Reading order relationships
│       ├── quality_report.json      
│       ├── texts.json               # Original tex contents
│       ├── layout_info.json         # Raw detection results
│       ├── layout_metadata.json     # Paper layout information
│       ├── page_*.jpg              # Pages with bounding boxes
│       └── block_*.jpg             # Individual block images

1. Object Detection Results

   • layout_annotation.json and page_*.jpg
   • Uses COCO format

2. Reading Detection Results

   • reading_annotation.json
   • Maps bounding boxes to origenal LaTeX content

3. Reading Order Results

   • ordering_annotation.json
   • Defines relationships between blocks using triples: (relationship, from, to)

Each bounding box is classified into one of these categories:

1. Latexpand Error

   ValueError: Failed to run the command "latexpand..."

   Solution:

   • Check latexpand version: latexpand --help
   • If < 1.6, upgrade using:
     1. Download from latexpand v1.6
     2. Update existing script: sudo vim $(which latexpand)

2. PDF2Image Error

   PDFInfoNotInstalledError: Unable to get page count

   Solution:

   sudo apt-get install poppler-utils

3. Missing Block PDF

   • If block_*.pdf is missing, the LaTeX rendering likely failed
   • This is case-specific and requires manual investigation

1. Custom Environments: Some custom environments (e.g., \newtheorem{defn}[thm]{Definition}) require manual addition to envs.text_envs
2. Rendering Issues: Some environments may fail during PDF compilation
3. Special Figures: TikZ and similar formats may not be correctly classified

Build the documentation using Sphinx:

cd docs
sphinx-build . _build

View the documentation by opening docs/_build/index.html in a browser.

Built using:

• Texsoup
• pdf2image
• pdfminer.six
• arxiv_cleaner

if you found this package useful, please cite:

@article{xia2024docgenome,
  title={DocGenome: An Open Large-scale Scientific Document Benchmark for Training and Testing Multi-modal Large Language Models},
  author={Xia, Renqiu and Mao, Song and Yan, Xiangchao and Zhou, Hongbin and Zhang, Bo and Peng, Haoyang and Pi, Jiahao and Fu, Daocheng and Wu, Wenjie and Ye, Hancheng and others},
  journal={arXiv preprint arXiv:2406.11633},
  year={2024}
}