tuandunghcmut/document-ai-paper-collection

Document AI Paper Collection

Curated repository of Document AI research papers with machine-readable catalog metadata for exploration, filtering, and downstream tooling.

Dataset Summary

• Purpose: support literature navigation and benchmark tracking for Document AI topics.
• Unit of data: one PDF paper file plus one catalog row in CSV/JSON.
• Primary use cases:
  • build paper browsers and search interfaces
  • benchmark-oriented reading lists (DLA, parsing, OCR, structure)
  • reproducible internal survey workflows

Repository Contents

• Papers (.pdf): 246
• Python scripts (.py): helper scripts for catalog generation and maintenance
• Catalog files:
  • catalog/catalog.csv
  • catalog/catalog.json
  • catalog/dla/DLA_research_catalog.csv
  • catalog/dla/DLA_research_catalog.json
  • catalog/dla/DLA_research_catalog_enriched.csv
  • catalog/dla/DLA_research_catalog_cleaned.csv
  • catalog/dla/DLA_papers.bib
  • catalog/dla/DLA_research_links.md

Topic Coverage

Current topic folders include:

• Document Layout Analysis (DLA)
• Document Parsing
• Document Segmentation
• Document Structure Analysis
• Document Understanding
• OCR
• Table Understanding / Recognition
• General MLLMs
• MLLMs for Document
• Pretrained for Document
• Document Extraction
• plus root-level Document AI papers

Benchmark Highlights

Key benchmark papers covered in the collection include:

• PubLayNet (ICDAR 2019)
• DocLayNet (KDD 2022)
• DocBank (COLING 2020)
• M6Doc (CVPR 2023)
• D4LA (ICCV 2023)
• PRImA/RDCL competition references

Scope

• This repo keeps papers and Python scripts only.
• Raw dataset archives (.zip, .tar*, .7z, .rar) are intentionally excluded by design.
• This is a literature collection, not a training-ready benchmark package.

Data Provenance

• Files were aggregated from public paper sources (conference proceedings, journals, arXiv, project pages).
• Metadata fields are parsed from filenames and directory structure, then normalized via scripts.
• No OCR text extraction or annotation labels are distributed in this dataset card.

Data Schema

Catalog columns:

• file_name: original PDF file name
• title: parsed paper title
• year: parsed publication/preprint year
• venue: parsed venue/source token (e.g., CVPR, ICCV, arXiv)
• topic: folder/topic name
• relative_path: path within repository
• file_size_bytes: file size in bytes
• file_size_mb: file size in MB (rounded)

Intended Uses

Recommended:

• literature review support
• benchmark reference indexing
• repository/Space demo backends

Not recommended:

• direct model training claims from this repo alone
• legal/compliance decisions without checking original publisher terms

Licensing Notes

• Repository metadata and organization are provided under CC-BY-4.0.
• Individual papers remain under their original publishers/authors' licenses and terms.
• Users must respect redistribution and usage terms of each original paper source.

Rebuild Catalog

python scripts/generate_catalog.py

DLA Enrichment Pipeline

The DLA topic includes a dedicated metadata enrichment pipeline:

python scripts/build_dla_research_catalog.py
python scripts/postprocess_dla_catalog.py

Outputs are written under catalog/dla/ and include arXiv abstract fields, reference links, and publication-ready BibTeX.

Validation and QA

Current checks performed:

• verify catalog row count matches number of PDFs
• ensure all relative_path entries resolve to existing files
• ensure archive file types are excluded from publish repo

Limitations

• metadata quality depends on filename consistency
• venue/year parsing may contain edge-case noise for non-standard file names
• no full-text extraction, citations graph, or abstract-level normalization yet

Maintenance

• Last major refresh: 2026-03-01
• Maintainer: tuandunghcmut
• Contributions: add papers/scripts via pull request with updated catalog

Citation

If you use this collection, cite as:

@misc{vo_document_ai_paper_collection_2026,
  title        = {Document AI Paper Collection},
  author       = {Dung Vo},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/tuandunghcmut/document-ai-paper-collection}},
  note         = {Curated repository of Document AI papers and metadata catalog}
}