--- license: mit task_categories: - object-detection - visual-document-retrieval language: - en - fr tags: - table - extraction - detection - structure - recognition pretty_name: Table Extraction Scientific Datasets size_categories: - 10K

Benchmarking Table Extraction from Heterogeneous Scientific PDF Documents

## Dataset Sources - **Repository:** [GitHub](https://github.com/soricm/table-extraction-benchmark) - **Code archive:** [Zenodo](https://doi.org/10.5281/zenodo.20486311) (DOI `10.5281/zenodo.20486311`) - **Paper:** [10.1145/3770855.3817462](https://doi.org/10.1145/3770855.3817462) - **Extended version:** [arXiv:2511.16134](https://arxiv.org/abs/2511.16134) - **Interactive demo:** [table-extraction-benchmark-explorer.streamlit.app](https://table-extraction-benchmark-explorer.streamlit.app/) ([source](https://github.com/soricm/table-extraction-benchmark-explorer)) ## Dataset Details - PubTables (`pubtables/*.tar.gz`) - Subset of PubTables-Test dataset, enriched with HTML table ground truth - `/images`: 46,942 JPG files; one file for each page image - `/test`: 46,942 XML files containing bounding boxes in PASCAL VOC format - `/pdfs`: 23,567 PDF files from which images are taken from - `/html`: 55,991 HTML files containing extracted tables with in HTML mark-up - `/word`: 46,942 JSON files containing bounding boxes of tokens in images - Table-arXiv (`table-arxiv/`, **generation recipe only, see below**) - Dataset synthetically generated from LaTeX source files from arXiv preprints for ground truth generation. - The arXiv license does not let us redistribute the source preprints or anything derived from them, so **the images, annotations and PDFs are not published here**. We ship the list of papers we used plus the code that rebuilds the dataset from them. See [Dataset Creation](#dataset-creation). - `table-arxiv-pdf-list.csv`: the 2,488 arXiv identifiers, PDF URLs and license fields the dataset was built from - Regenerating from that list reproduces 36,869 page images, 5,327 PASCAL VOC XML annotations, 6,441 table HTML files and 40,261 word-box JSON files - Table-BRGM (`table-brgm.tar.gz`) - Dataset manually annotated, domain-specific and comprises French and English geological reports sourced from a selection of BRGM - `/images`: 2003 JPG files; one file for each page image - `/test`: 251 XML files containing bounding boxes in PASCAL VOC format - `/pdfs`: 26 PDF files from which images are taken from - `/html`: 346 HTML files containing extracted tables with in HTML mark-up - `/word`: 2003 JSON files containing bounding boxes of tokens in images - `/grid`: 2003 pickle files holding the tokenizer grid input (`input_ids` and per-token bounding boxes) consumed by the VGT model - ICDAR-2013 (`icdar-2013.tar.gz`) - Dataset manually corrected from the ICDAR-2013 competition, contains European Union and US Government PDF reports. - `/images`: 238 JPG files; one file for each page image - `/test`: 135 XML files containing bounding boxes in PASCAL VOC format - `/pdfs`: 67 PDF files from which images are taken from - `/html`: 163 HTML files containing extracted tables with in HTML mark-up - `/word`: 238 JSON files containing bounding boxes of tokens in images - `/grid`: 238 pickle files holding the tokenizer grid input (`input_ids` and per-token bounding boxes) consumed by the VGT model ## Dataset Creation ### Table-arXiv is not redistributed Table-arXiv is built from arXiv preprints, which are covered by the [arXiv.org non-exclusive license to distribute](https://arxiv.org/licenses/assumed-1991-2003/license.html). That license lets arXiv distribute the papers but does not give us the right to redistribute them, and the same restriction carries over to the page images, PDFs and table crops derived from them. We therefore publish **the recipe, not the data**: - `table-arxiv/table-arxiv-pdf-list.csv`: the exact 2,488 papers used, as `arxiv_id, pdf_url, license`. Fetch the PDFs and LaTeX sources from arXiv yourself, under arXiv's terms. - The generation code lives in [table-arxiv-generate](https://github.com/soricm/table-extraction-benchmark) (`scripts/table-arxiv-generate/`). It compiles each paper's `.tex` with anchor macros injected around every `\tabular` environment, reads the resulting named PDF destinations to recover table bounding boxes, and emits page images, PASCAL VOC XML, per-table HTML and word-box JSON, the same layout as the other three datasets in this repo. Running that pipeline over the published paper list reproduces the Table-arXiv split used in the paper. The other three datasets (PubTables, Table-BRGM, ICDAR-2013) are distributed here in full. ## Benchmark Results `experience-data/` holds the raw model output and the evaluation scores behind the paper, for all four datasets and 15 systems (14 scored under each metric, see below). - `-predictions.tar.gz`: one `predictions_.json` per system, the raw extraction output before scoring. Table-BRGM and ICDAR-2013 additionally carry `predictions_vgt_td.json`, the table-detection-only VGT variant. - `-results.tar.gz`: the evaluation scores, as `results__final_.json` where `` is `bbox` or `token`, plus a `manifest.json` recording the provenance and sha256 of every file. The [interactive demo](https://table-extraction-benchmark-explorer.streamlit.app/) browses a 62-page sample of these files page by page, if you would rather look than unpack. The two variants differ in how a table counts as detected before its structure is scored. `bbox` (called **bbox TE** in the paper) requires IoU with the ground truth above 0.5. `token` (**txt TE**) replaces IoU with the Jaccard index over 2-grams of the table's text content, so that methods emitting no bounding boxes can still be evaluated. | Variant | Systems evaluated | | --- | --- | | `bbox` | `cam doc gemini gpt grobid mathpix monkey pedia plum pymu qwen tatr vgt xy` | | `token` | `cam doc gemini got gpt grobid mathpix monkey plum pymu qwen tatr vgt xy` | `got` (GOT-OCR 2.0) appears only under `token`, since it outputs no bounding boxes. The aggregate tables are in `experience-data/benchmark-results-bbox.csv` and `experience-data/benchmark-results-token.csv`. Their headline column is `TEDS`, which is the end-to-end **`F1_TEDS`** score of the paper: a table extraction F1 whose per-table contribution is weighted by the TEDS structure score, so it is *not* a plain TEDS value. The columns map to the paper's notation as `Raw` = `F1_bbox` (resp. `F1_txt`) for table detection, then `Topology` = `F1_Top`, `Content` = `F1_Con` and `TEDS` = `F1_TEDS` for extraction. See Section 5.3 and Tables 4 and 8 of the [extended version](https://arxiv.org/abs/2511.16134) for the full definitions. `tatr`, `xy` and `vgt` are the exception. They attach a confidence score to each detection, so instead of a single operating point they are scored with **`AP_TEDS`**, the area under the TEDS-weighted precision–recall curve swept over that threshold (paper Table 3, where they appear as `TATR`, `XY+TATR` and `VGT+TATR-struct`). For these three rows the same four CSV columns therefore hold `AP`, `AP_Top`, `AP_Con` and `AP_TEDS`: | Model | Variant | PubTables | Table-arXiv | Table-BRGM | ICDAR-2013 | | --- | --- | --- | --- | --- | --- | | `tatr` | bbox | 0.80 | 0.52 | 0.64 | 0.76 | | `tatr` | token | 0.79 | 0.36 | 0.61 | 0.78 | | `xy` | bbox | 0.78 | 0.51 | 0.73 | 0.71 | | `xy` | token | 0.77 | 0.37 | 0.68 | 0.73 | | `vgt` | bbox | 0.70 | 0.40 | 0.14 | 0.35 | | `vgt` | token | 0.64 | 0.14 | 0.04 | 0.15 | The paper's Table 3 covers the bbox variant only. Its values match on PubTables, Table-arXiv and Table-BRGM, and differ on ICDAR-2013 (0.66, 0.63 and 0.34 in the paper), because that evaluation was rerun afterwards. ## Citation ``` @misc{soric2025benchmarkingtableextractionheterogeneous, title={Benchmarking Table Extraction from Heterogeneous Scientific Extraction Documents}, author={Marijan Soric and Cécile Gracianne and Ioana Manolescu and Pierre Senellart}, year={2025}, eprint={2511.16134}, archivePrefix={arXiv}, primaryClass={cs.DB}, url={https://arxiv.org/abs/2511.16134}, } ```