Jonnob/the-spiritualist-enriched

The Spiritualist Enriched

Ground truth OCR annotations for The Spiritualist, a 19th-century British newspaper. An enriched version of the original derived from Transkribus. Aadditional manually labelled with, classes, column types, reading order and Semantic Structural Units (SSUs).

The page images and original ground truth are available from the companion dataset: NationalLibraryOfScotland/Spiritualist_Newspaper

Contents

File	Description
annotations/gt_ssu_bboxes.parquet	425 SSU regions across 49 pages: bounding boxes, polygon points, SSU label, ground truth text
annotations/characters_inferred.parquet	~827k character-level annotations with per-character bounding boxes and SSU label
alto_xml/ocr_gt_labelled.zip	Raw ALTO XML files (ALTO v4, as exported from Transkribus)

Schema — gt_ssu_bboxes.parquet

Column	Type	Description
filename	string	Image filename (matches the companion image dataset)
page_id	string	Page identifier
image_width	int	Page image width in pixels
image_height	int	Page image height in pixels
x	int	Bounding box left edge
y	int	Bounding box top edge
width	int	Bounding box width
height	int	Bounding box height
polygon_points	string	Space-separated polygon vertices (x,y pairs)
ssu_id	string	Semantic Structural Unit label (e.g. ssu_masthead, ssu_1_col_1)
gt_text	string	Ground truth transcription for the region

Schema — characters_inferred.parquet

Column	Type	Description
char_id	string	Unique character identifier
page_id	string	Page identifier
char_text	string	Character
x	float	Character bounding box left edge
y	float	Character bounding box top edge
w	float	Character bounding box width
h	float	Character bounding box height
ssu_id	string	SSU label for the containing region

Citation

If you use this dataset please cite

@misc{bourne2026charactererrorvector,
      title={The Character Error Vector: Decomposable errors for page-level OCR evaluation},
      author={Jonathan Bourne and Mwiza Simbeye and Joseph Nockels},
      year={2026},
      eprint={2604.06160},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://doi.org/10.48550/arXiv.2604.06160}
}