ahmedHamdi/CzechHistoricalNewspapers4AS

Czech Historical Newspapers

Scanned pages of Czech historical newspapers annotated with layout regions and ground-truth article cluster labels. The dataset is designed for unsupervised clustering experiments: the task is to group text regions on each page into the articles they belong to, using only text content or visual features.

Dataset structure

Two configurations are available:

pages (30 rows — one per newspaper page)

Column	Type	Description
image_id	int	Page ID (1–30)
image_name	str	Filename, e.g. Brnenske_48_D.jpg
series	str	Newspaper series (Brnenske or Lidove)
issue	str	Issue number within the series
page	str	Page letter within the issue
texts	list[str]	OCR text of each text region, in scan order
category_ids	list[int]	Category ID for each text region
category_names	list[str]	Category name for each text region
cluster_labels	list[int]	Ground-truth article cluster index for each region
num_texts	int	Number of text regions on the page
num_clusters	int	Number of distinct articles on the page

from datasets import load_dataset
ds = load_dataset("your-username/czech-historical-newspapers", "pages")
row = ds["train"][0]
# row["texts"]          -> list of OCR strings
# row["cluster_labels"] -> ground-truth article assignment per text

annotations (881 rows — one per text region)

Column	Type	Description
annotation_id	int	Original annotation ID
image_id	int	Page ID
image_name	str	Filename
series	str	Brnenske or Lidove
issue	str	Issue number
page	str	Page letter
category_id	int	Region type ID
category_name	str	Region type name
text	str	OCR text content
bbox_x/y/w/h	int	Bounding box in pixels [x, y, width, height]
cluster_label	int	Ground-truth article cluster index

ds = load_dataset("your-username/czech-historical-newspapers", "annotations")
# Filter to a single page
page = ds["train"].filter(lambda x: x["image_name"] == "Brnenske_48_D.jpg")

Category types

Annotations belong to one of the following layout region types. Categories 1, 3, 6, and 7 are excluded from both configs (no text content).

ID	Name	Kept	Count
1	SeparatorRegion	No	297
2	Heading	Yes	132
3	Caption	No	67
4	Enumeration	Yes	17
5	TextRegion	Yes	572
6	ImageRegion	No	83
7	Article	No	0
8	TableRegion	Yes	1
9	AdvertRegion	Yes	159

Statistics

• 30 pages from 2 Czech historical newspaper series
• 12 Brnenske pages (issues 09, 44, 48) and 18 Lidove pages (issues 04, 06, 07)
• 881 text annotations with ground-truth labels
• 1–27 articles per page (median ~7)
• Text is in Czech; embeddings from multilingual models are recommended

Intended use

The ground-truth cluster_labels are the reference for evaluating clustering algorithms (K-Means, Spectral Clustering, Affinity Propagation, etc.) using metrics such as Adjusted Rand Index and Normalized Mutual Information.

Images are not included. The original scans are high-resolution JPEGs (~3500×5100 px).

Annotation

Annotated by "honza" using bounding-box tools. Bounding boxes follow the COCO convention [x, y, width, height] in pixels.

license: cc-by-4.0