histde/ddb-newspaper-corpus

📰 DDB Newspaper Corpus

A corpus of 11,551,703 pages of historical German newspapers in the public domain, harvested from the Deutsche Digitale Bibliothek (DDB) and its Zeitungsportal.

It covers 1,607,744 issues from 796 newspapers published between 1638 and 1964, totalling 25.8 billion whitespace tokens (157.9 billion characters) of OCR fulltext. Every page carries an explicit per-page license (Public Domain Mark or CC0) and links back to its full-resolution scan (IIIF) and its coordinate-level OCR (ALTO XML).

📊 Key Statistics

Metric	Value
Pages	11,551,703
Issues	1,607,744
Newspapers (ZDB IDs)	796
Providers	19
Characters	157,907,966,609
Tokens (whitespace)	25,766,037,329
Mean characters per page	13,670
Median characters per page	12,437
Temporal coverage	1638-1964

The corpus peaks in the late 19th century: the decades from 1870 to 1919 alone contribute about 59% of all pages, with 1890-1899 as the largest bin (2,090,315 pages, 18.1%). Half of all pages hold 2,048 or more tokens, which makes the corpus comparatively long-context: pages with at least 2,048 tokens carry 75% of all tokens. The full statistics (pages and tokens per 10-year bin, length percentiles, per-provider and per-newspaper breakdowns) are available in the GitHub repository.

🗂️ Data Fields

Column	Description
id	DDB page identifier; the part before the first - is the issue id
zdb_id	Newspaper identifier in the Zeitschriftendatenbank
paper_title	Newspaper title
provider	Name of the holding institution
provider_ddb_id	DDB identifier of the holding institution
pagenumber	Page number within the issue
pagename	Internal page name within the issue
publication_date	Issue date (ISO 8601)
place_of_distribution	Place(s) of publication, may be null
language	Language code(s), ger
text	OCR plain text of the page
license	Per-page license URL (CC0 or Public Domain Mark)
iiif_url	IIIF Image API URL of the page scan
preview_reference	ALTO XML of the page (word coordinates)
thumbnail	DDB binary UUID of the scan

🚀 Usage

Streaming is the recommended starting point, no full download required:

from datasets import load_dataset

dataset = load_dataset("histde/ddb-newspaper-corpus", split="train", streaming=True)

page = next(iter(dataset))
print(page["paper_title"], page["publication_date"])
print(page["text"][:200])

Full download:

dataset = load_dataset("histde/ddb-newspaper-corpus", split="train")

🖼️ Linked Scans and OCR Coordinates

Each page links to its source materials, both publicly accessible without an API key:

• iiif_url points to the full-resolution page scan on the DDB IIIF Image API server. All standard IIIF region, size and rotation parameters work, e.g. replace /full/max/ with /full/667,/ for a smaller derivative.
• preview_reference points to the ALTO XML of the page with word-level coordinates, which enables OCR ground-truth work, layout analysis and text-image alignment.

🏛️ Source and Provenance

The corpus is a snapshot of the public-domain, German-language part of the DDB Zeitungsportal, harvested on August 7-8, 2026 via the DDB API. The digitized newspapers originate from 19 institutions; the largest contributors are the Sächsische Landesbibliothek - Staats- und Universitätsbibliothek Dresden (21.1% of pages), the Universitäts- und Landesbibliothek Bonn (14.3%), the Universitäts- und Landesbibliothek Sachsen-Anhalt in Halle (13.6%), the Staatsbibliothek zu Berlin (10.8%) and the Staats- und Universitätsbibliothek Hamburg (8.2%).

This dataset is an updated version of the DDB newspaper collection presented in the German Commons paper, whose unfiltered version contains 10.382 million newspaper pages. The present snapshot adds roughly 1.17 million pages (about 11% more) that were digitized or newly licensed as public domain since then, and additionally links every page to its IIIF scan and ALTO XML.

🛠️ Corpus Construction

The harvesting pipeline is open source: github.com/histde/ddb-corpus. In short:

1. All public-domain, German-language newspaper issues are collected from the DDB newspaper-issues Solr index together with their per-issue license.
2. All pages of these issues are downloaded via issue-id prefix queries, so only public-domain pages are transferred. Completeness is verified per batch and incomplete server answers are rejected and retried.
3. The pages are converted to Parquet shards (10,000 pages per shard, zstd) with a fixed, verified schema.

All requests identified themselves with a descriptive User-Agent including a contact address, and the pipeline throttles itself collectively as soon as the API shows signs of distress.

⚖️ Licensing

The license of every page is stated in its license column, as assigned by the digitizing institution in the DDB:

License	Pages	Share
Public Domain Mark 1.0	11,521,541	99.74%
CC0 1.0	30,162	0.26%

The descriptive metadata is provided by the DDB under CC0.

⚠️ Limitations and Biases

• OCR quality varies considerably. Most newspapers are set in Fraktur (blackletter); depending on scan quality and OCR engine, error rates range from near-perfect to noisy. Historical typography (long s "ſ", ligatures) and historical orthography are preserved.
• Page-level granularity. A row is a newspaper page, not an article. Articles are not segmented and text columns may interleave in reading order.
• Hyphenation. Line-break hyphenation is generally not merged; words split across lines appear as two tokens.
• Temporal skew. Coverage concentrates in 1850-1940; the 17th and 18th centuries are sparse.
• Historical content. The corpus reflects the full range of its era's press, including nationalist, antisemitic, colonialist and otherwise discriminatory language, as well as historical events described from contemporary perspectives. It is provided as a historical source; the content does not represent the views of the dataset creator.

📖 Citation

@misc{schweter-2026-ddb-newspaper-corpus,
  author    = {Schweter, Stefan},
  title     = {{DDB Newspaper Corpus: 11.5M Public-Domain Historical German Newspaper Pages}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/histde/ddb-newspaper-corpus}
}

🙏 Acknowledgements

This corpus exists thanks to the Deutsche Digitale Bibliothek and its Zeitungsportal. Additionally, we also thank the 19 libraries and archives that digitized their newspaper collections and released them under public domain!