biglam/yalta_ai_tabular_dataset

YALTAi Tabular Dataset

353 page images of historical documents with tabular layouts, mostly notarial registers, with bounding-box annotations for four zone types: Col, Header, Marginal and text. Created by Thibault Clérice and deposited on Zenodo alongside the paper You Actually Look Twice At it (YALTAi) (Journal of Data Mining and Digital Humanities, 2022), which treats page layout recognition on historical documents as an object detection problem rather than a pixel classification one.

The annotations support layout analysis before transcription. Telling the columns of a register apart from its header and its marginalia is what decides whether a downstream OCR/HTR engine emits usable structured text or one merged block.

The sibling dataset. biglam/yalta_ai_segmonto_manuscript_dataset is the other half of the same experiment: manuscripts and early printed books, labelled with the SegmOnto zone vocabulary. This dataset does not use SegmOnto — tables were given their own four-class scheme.

Splits

	train	validation	test
images	196	22	135

Object instances per class, from Table 3 of the paper:

	Train	Dev	Test	Total	Average area	Median area
Col	724	105	829	1658	9.32	6.33
Header	103	15	42	160	6.78	7.10
Marginal	60	8	0	68	0.70	0.71
text	13	5	0	18	0.01	0.00

Only Col and Header are annotated in the test split. The paper's reason: "marginal and text were not clearly defined". Marginal and text have zero test instances and cannot be evaluated on held-out data — treat them as training signal only, and expect a model trained on all four classes to be scored on two.

The test split is out of domain by design. Train and validation come from a single source; the test set is unrelated material from the 17th to the early 20th century, including one soldier's war report, with column borders that are neither drawn nor printed and layouts the paper calls "masonry". Test scores here measure transfer, not in-domain accuracy. In the paper the best model reaches 4.77% mAP on this test set, against 47.75% on the manuscripts dataset.

Column labels were collapsed before release. The source annotations used "around 16 different ways to describe columns, from Col1 to Col7, the case-different col1-col7 and finally ColPair and ColOdd" — a convention that works around Kraken's tendency to merge adjacent regions of the same type. All of them were reduced to a single Col.

Source data

Train and validation come from the Lectaurep Repertoires dataset (Rostaing et al., 2021). LECTAUREP (LECTure Automatique de REPertoires), begun in 2018, is a joint project of the Minutier central des notaires de Paris at the Archives nationales, the ALMAnaCH team at Inria and the EPHE, in partnership with the French Ministry of Culture. The material is notaries' repertories: ruled, tabular, handwritten registers.

Loading the data

This repo still ships a legacy loading script, which datasets no longer runs:

>>> load_dataset("biglam/yalta_ai_tabular_dataset", "YOLO")
RuntimeError: Dataset scripts are no longer supported, but found yalta_ai_tabular_dataset.py

(checked against datasets 5.0.0). Until this repo is converted, take the data from Zenodo — a 359 MB zip laid out as yaltai-table/{train,val,test}/{images,labels}, one YOLO-format .txt per image (class index, then normalised centre-x, centre-y, width, height):

curl -L -o yaltai-table.zip "https://zenodo.org/record/6827706/files/yaltai-table.zip?download=1"

The loading script defined two configurations over the same annotations: YOLO, the original format, and COCO, converted for the Transformers object-detection processors. The COCO conversion left segmentation empty — these are boxes, not polygons.

Licence

Creative Commons Attribution 4.0 International.

Citation

@dataset{clerice_thibault_2022_6827706,
  author       = {Clérice, Thibault},
  title        = {YALTAi: Tabular Dataset},
  month        = jul,
  year         = 2022,
  publisher    = {Zenodo},
  version      = {1.0.0},
  doi          = {10.5281/zenodo.6827706},
  url          = {https://doi.org/10.5281/zenodo.6827706}
}

The accompanying paper: Thibault Clérice, "You Actually Look Twice At it (YALTAi): using an object detection approach instead of region segmentation within the Kraken engine", Journal of Data Mining and Digital Humanities, 2022. arXiv:2207.11230.

Thanks to @davanstrien for adding this dataset to the Hub.