chcaa/eno-newspapers-enriched

Danish Historical Newspaper Articles Dataset (enriched)

This dataset contains approximately 4.9 million Danish historical newspaper articles (1666–1850) with document embeddings and assigned fictionality tags, providing a comprehensive resource for studying Danish language, culture, and history through primary journalistic sources.

Dataset Details

Dataset Description

This dataset comprises digitized newspaper articles from Danish newspapers, featuring full-text content along with metadata including publication dates, newspaper titles, and automated category classifications. The collection spans 28 Danish historical newspapers and periodicals, published between 1666 and 1850 in Denmark-Norway.

The dataset includes document embeddings (pooled representations) created with the Old_News_Segmentation_SBERT_V0.1 model.

Key statistics:

• Total articles: 4,898,084
• Dataset size: ~33.7 GB (uncompressed), ~25.1 GB (download)
• Language: Danish (da)
• Format: Parquet files

Features:

• text: Full article text content
• date: Publication date
• id: Unique article identifier
• pwa: Probability-weighted attribute (float)
• newspaper: Newspaper title
• pooled: Pre-computed embeddings (list of floats) for semantic analysis
• year: Publication year (derived from date)
• decade: Publication decade (derived from year)
• predicted_category: Automatically assigned content category: National News, International News, Advertisements, Paratext - result of a classification task (see Paper 1)
• fictionality_tag: Automatically assigned ficitonality tag: fiction/nonfiction - result of classification task (see Paper 2)
• fiction_prob/non_fiction_prob: probability accompanying the fictionality tags

Source details:

• Curated by: Alie Lassche (a.w.lassche@cas.au.dk), Pascale Feldkamp (pascale.feldkamp@cas.au.dk) & Johan Heinsen (heinsen@dps.aau.dk)
• Source digitized by: Johan Heinsen & Camilla Bøgeskov, The ENO Project
• Language(s): Danish (da)

Dataset Sources

• Repository: GitHub
• Paper 1: Evaluating Embedding Models on Danish Historical Newspapers DOI
• Paper 2: A Growing Literature of the Public Sphere: Fiction in Danish Newspapers (1666-1850) link

Source Data

Data Collection and Processing

The newspapers are kept in extraordinarily well-preserved collections housed by the national libraries of Denmark and Norway. The processing pipeline is described in Paper 1.

Known Limitations

• OCR errors: Text extracted from historical newspapers may contain errors due to print quality, paper degradation, or OCR limitations
• Segmentation issues: Article boundaries may be inaccurately identified, leading to merged or split articles
• Historical bias: Content reflects the perspectives, biases, and social attitudes of the time period, which may include prejudiced views
• Coverage gaps: Digitization may not be comprehensive across all time periods or newspapers
• Language evolution: Danish language, spelling, and orthography have changed over time
• Selection bias: Available newspapers may not represent all regions, political viewpoints, or social classes equally

Details on the automatic annotations

Predicted Category Tags

The dataset includes, for each article, predicted category tags – category_tag (Advertisement, National news, International news, or Miscellaneous) – and corresponding probabilities.
Predictions were generated using a Logistic Regression classifier trained on embeddings from the Old_News model. The model was evaluated using a 50-iteration 80/20 train-test split and achieved:

• High Precision, recall, and F1-scores across all categories (F1-score: 0.93–0.97 for Advertisement, National news, and International news).

For methodological context, see Paper 1. Citation:

@inproceedings{lassche2026evaluating,
  title={Evaluating Embedding Models on Danish Historical Newspapers: A Corpus and Benchmark Resource},
  author={Lassche, Alie and Feldkamp, Pascale and Bizzoni, Yuri and Baunvig, Katrine and Nielbo, Kristoffer and Heinsen, Johan},
  booktitle={Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
  pages={3577--3589},
  year={2026},
  address={Palma, Mallorca, Spain},
  date={11--16 May 2026},
  organization={ELRA Language Resources Association (ELRA)}
}

Predicted fictionality tags

The dataset includes, for each article, predicted fictionality – fictionality_tag (fiction or non-fiction) - and corresponding probabilities (fiction_prob, non_fiction_prob). Predictions were generated using a Logistic Regression classifier trained on embeddings with the Old_News model, with performance validated via stratified group k-fold cross-validation. The model achieves:

• Precision/Recall (fiction & non-fiction): 0.88–0.89 ± <0.05
• Well-calibrated probabilities (Brier score: 0.024), making the confidence scores reliable for filtering (e.g., fiction_prob > 0.99). Original probabilities were kept. We tested calibrated scores but found no improvement over the original probabilities.
• 77,513 articles classified as fiction (ca. 1.58% of the dataset)

For methodological context, see Paper 2. Citation:

@inproceedings{feldkamp2026growing,
  title={A Growing Literature of the Public Sphere: Fiction in Danish Newspapers (1666-1850)},
  author={Feldkamp, Pascale and Lassche, Alie and Eriksen, Rie and Morgenstjerne, Kit and Heinsen, Johan and Nielbo, Kristoffer and Bizzoni, Yuri},
  booktitle={Proceedings of the First Workshop on Creating Interoperable Corpora of Historical Newspapers (PressMint-LREC2026)},
  year={2026},
  address={Marseille, France},
  date={16 May 2026},
  organization={ELRA/LREC}
}

How to use it

Available Splits

• **train**: Full dataset (all years).
• **decade_{YYYY}s**: Articles from a specific decade (e.g., decade_1900s, decade_1910s).

Decade splits are provided for efficient loading and temporal sampling. The full dataset remains the canonical form; splits are a convenience layer rather than separate datasets.

Load the Full Dataset

from datasets import load_dataset

dataset = load_dataset("chcaa/eno-newspapers-enriched", split="train")

Load a Single Decade

from datasets import load_dataset

# Example: Load articles from 1700-1709
from datasets import load_dataset

dataset_1700s = load_dataset(
    "chcaa/eno-newspapers-enriched",
    split="decade_1700s")

Load a Range of Years

from datasets import load_dataset, concatenate_datasets

splits = ["decade_1700s", "decade_1710s", "decade_1720s"]

datasets = [load_dataset("chcaa/eno-newspapers-enriched", split=s) for s in splits]

dataset = concatenate_datasets(datasets)

Dataset Card Author

Alie Lassche & Pascale Feldkamp, Center for Humanities Computing, Aarhus University