Enevaeldens Nyheder Online (ENO)

• Enevaeldens Nyheder Online (ENO) is a historical newspaper dataset that reconstructs 565,000 pages and 474 million words from Danish-Norwegian publications dating from 1660 to 1849.
• It employs neural-network-based OCR combined with custom layout segmentation and BERT-enhanced classification to extract stand-alone text units from complex, multi-text newspaper editions.
• The dataset supports advanced research in historical linguistics and print culture analysis by providing quality-estimated text units, specialized language models, and rigorous metadata.

Enevaeldens Nyheder Online (ENO), translated in the source as “News during Absolutism Online,” is a historical newspaper dataset that reconstructs the contents of major newspapers in Denmark and Norway during the period 1660–1849, corresponding to the Danish-Norwegian “Age of Absolutism.” It is presented as a Data Descriptor and comprises approximately 474 million words derived from about 565,000 pages across 28 major newspapers, using neural-network-based processing of digitised microfilm from Danish newspapers together with a smaller selection of Norwegian publications that had previously been illegible for computers. The resource is defined not merely as OCR output, but as a corpus of stand-alone text units extracted from historically complex newspaper editions, with an accompanying BERT model trained on a beta-version of the dataset (Heinsen et al., 2 Sep 2025).

1. Historical scope and corpus composition

ENO covers the period 1660–1849 and spans almost 200 years of newspaper production. The dataset is dominated by Danish in historical forms, includes a smaller selection of Norwegian publications, and contains some German texts, especially in advertisements and certain segments. Geographically, it covers major national and provincial newspapers in Denmark as well as major Norwegian titles, particularly from Oslo, Bergen, and Trondheim; in this sense, the newspapers are described as reflecting the print culture of the entire Denmark-Norway state (Heinsen et al., 2 Sep 2025).

The scale of the corpus is one of its defining characteristics. The dataset contains approximately 565,000 pages and about 474 million words, drawn from 28 major newspapers. Many series are complete, while some are partial because of damaged or lost originals. This combination of temporal depth, linguistic heterogeneity, and broad regional coverage makes ENO relevant both for historical linguistics and for the study of newspaper culture, information circulation, and print publics under absolutism (Heinsen et al., 2 Sep 2025).

A central design premise is that a newspaper edition is not treated as a single monolithic text. Instead, ENO models the fact that one edition may comprise multiple texts and that a text may cross page boundaries. This treatment is methodologically important because the newspaper is approached as a structured aggregation of heterogeneous textual units rather than as a flat sequence of OCR-recognized lines. A plausible implication is that the corpus is better aligned with downstream research tasks such as genre-sensitive retrieval, discourse analysis, and study of advertisements or social communications.

2. Source materials and digitization constraints

The source materials differ substantially across the Danish and Norwegian portions of the corpus. For Denmark, the input consisted of digitized microfilm from Denmark’s Royal Library, originally scanned in the 1950s and later digitized. For Norway, the input consisted of modern digital photographs from Norwegian national collections. Both sources suffered from major limitations, including poor microfilm quality, old commercial OCR, and the prevalence of fraktur typeface (Heinsen et al., 2 Sep 2025).

The paper identifies several factors that caused previous OCR attempts to fail: poor microfilm quality, fraktur typeface, complex page layouts with narrow column separators, poor page contrast, small print, and physical damage. The reported predicted word accuracy using the library’s OCR was approximately 47–50% median, with Aarhus Stifts-Tidende given as an example at 48.5%. These figures establish the baseline difficulty of the task and explain why conventional OCR pipelines trained primarily on antiqua and modern print were inadequate (Heinsen et al., 2 Sep 2025).

The source conditions also shaped the project’s technical priorities. A common assumption in historical OCR projects is that character recognition is the dominant obstacle; ENO explicitly complicates that assumption by showing that layout segmentation, reading order, and article extraction were equally central. The paper’s example of narrow column separators, referenced in Figure 1, illustrates that the problem was not simply one of recognizing glyphs but of reconstructing a document topology usable for scholarship.

3. Neural OCR and layout reconstruction

ENO employed the Transkribus platform because of its success with historical handwriting and print. Model training began with 100,000 words of 18th-century Danish and was progressively expanded across multiple newspapers and years, including course and workshop contributions by students. The final model was trained on approximately 420,000 words spanning the whole period. On validation data, the paper reports a character error rate of 0.6%, corresponding to approximately 95%+ word accuracy; elsewhere in the summary, this is described as “<1% character error rate.” The transcription strategy was diplomatic, preserving original orthography and idiosyncrasies such as interchangeable $i/j$, and the model was trained mainly on fraktur while also including some antiqua types (Heinsen et al., 2 Sep 2025).

Layout segmentation was treated as critical for both reading order and separation of articles. The initial approach used custom line-recognition models for each layout variant. This was later improved through a fields-model architecture in Transkribus, which enhanced column segmentation, especially on pages with narrow column separators. Manual corrections were applied where automatic segmentation remained unreliable, and the paper explicitly notes that layout was more labour intensive than pure text recognition (Heinsen et al., 2 Sep 2025).

This distinction between text recognition and layout reconstruction is significant. ENO presents historical newspaper digitization as a joint computer-vision and document-structuring problem rather than as OCR in the narrow sense. The reported success of neural text recognition does not eliminate the need for labor-intensive structural intervention; instead, it shifts the bottleneck toward page segmentation, reading-order recovery, and the derivation of analytically meaningful text units.

4. Stand-alone text extraction and segmentation methodology

A defining feature of ENO is its conversion of editions into stand-alone texts. The paper states that this was critical to the dataset design because newspapers contain multiple, often mixed-type texts per edition, and meaningful analysis requires extraction of cohesive text units. The core problem was therefore how to split continuous text into meaningful articles, advertisements, or news items (Heinsen et al., 2 Sep 2025).

The initial solution relied on machine learning with hand-crafted structural features. These features included grammar, such as the presence of a full stop, line length, capitalization, and regex patterns for headlines. A random forest classifier was then used to predict whether a line marked the start of a text. This approach was subsequently enhanced with BERT-based classification. The project trained a Danish-language BERT on 260 million words and used a fine-tuned BERT, together with approximately 30,000 manually tagged lines as gold data and line context including first line, header, and last line, to improve classification. Features from the hand-crafted and neural approaches were then combined in a random forest in order to obtain robustness against formatting variance (Heinsen et al., 2 Sep 2025).

The reported result for this segmentation pipeline is an $F_1$-score of 98.9% on a validation set of approximately 7,000 tagged lines. The paper also notes that some edge cases remained problematic, especially long poems and essays. This combination of feature engineering, line-level gold annotation, contextual BERT representations, and ensemble classification situates ENO’s article segmentation as a hybrid statistical-neural workflow rather than a purely end-to-end neural pipeline. A plausible implication is that formatting heterogeneity in historical newspapers still rewards explicit structural modeling even when LLMs are available.

5. Quality estimation, language modeling, and retrieval infrastructure

ENO supplements OCR and segmentation with a quality-estimation layer based on dictionary-matched predicted word accuracy per text. The dictionary was built from period writings, census data, and manually curated unrecognized words. Every recognized word was matched against this dictionary to provide a quick aggregate performance check, and the resulting pwa scores were used to flag low-quality segments, including texts in German and texts derived from blurry scans. Figures 4 and 5 are reported as showing, respectively, the quartile distribution of pwa values per title and the full distribution of pwa scores across the corpus (Heinsen et al., 2 Sep 2025).

The accompanying model, DA-BERT_Old_News_V1, is described as a BERT fill-mask model trained on 260 million words from segmented ENO texts. Its purpose is to capture Danish as used in newspapers from 1660–1849, including orthographic and historical-specific features. The model was used to power segmentation, to generate text embeddings for navigation and similarity search in the public interface, and to support experiments in OCR error correction that are characterized as promising but early-stage. For similarity search, the paper describes embeddings from a fine-tuned BERT combined with FAISS for real-time similar-text suggestions, with up to 25 similar results shown per text; it also notes that similarity scores were powered by Sentence-BERT or equivalent fine-tuned versions and mentions SetFit models for downstream tasks such as text boundary detection (Heinsen et al., 2 Sep 2025).

These components situate ENO simultaneously as a corpus and as an infrastructure for computational access. The public interface allows browsing, full-text and regex search, and similarity navigation via text embeddings, with each entry linked to the original library scan when possible. Full CSV and Parquet versions are available on Hugging Face and contain stand-alone text units with fields including date, title, and pwa score, while excluding hyperlinks and similarity indices to limit file size and memory usage. The dataset is also included in DynaWord as a major contribution to the Danish Dynaword corpus for language modeling (Heinsen et al., 2 Sep 2025).

6. Research significance, limitations, and usage conditions

The paper presents two major findings. First, neural text recognition substantially outperformed legacy OCR for complex fraktur historical print, even under poor microfilm conditions. Second, it demonstrated the viability of a historical newspaper LLM trained on Danish from this period, enabling both language research and in-corpus tasks. At the same time, the paper identifies several challenges: layout complexity often exceeded OCR difficulty; some material, such as late Kjøbenhavns Adresseavis, was so degraded that it had to be sampled rather than included wholesale; Norwegian scans sometimes fared worse despite higher image quality because models had learned to expect degraded images; and article segmentation required a combination of statistical and neural methods rather than generic rules alone (Heinsen et al., 2 Sep 2025).

The dataset is explicitly not presented as a neutral corpus. The paper acknowledges historical bias and states that coverage reflects what has survived and what image quality made processable, rather than an exhaustive representation. It also warns that newspaper language, values, topics, and sentiments are of their time, including attitudes related to colonialism and violence. For this reason, pwa scores are provided per text and aggressive filtering is recommended where high accuracy is required (Heinsen et al., 2 Sep 2025).

A common misconception would be to treat ENO as a simple OCR dump of scanned newspapers. The description instead emphasizes diplomatic transcription, careful layout modeling, extraction of stand-alone texts, and integration with language-model-based navigation. Another misconception would be to assume that greater image clarity automatically yields better recognition performance; the reported Norwegian case indicates that model expectations about degradation can affect transfer performance. This suggests that historical-document pipelines remain sensitive to domain shift not only in typography and language but also in image-formation characteristics.

7. Access, formats, and place within historical newspaper corpora

ENO is distributed through multiple access channels. A custom site per publication is available at https://hislab.quarto.pub/eno/, the full dataset in CSV and Parquet format is hosted at https://huggingface.co/datasets/JohanHeinsen/ENO, and the LLM DA-BERT_Old_News_V1 is hosted at https://huggingface.co/CALDISS-AAU/DA-BERT_Old_News_V1. The paper also notes that code is available through the project website and a GitHub repository linked in the paper (Heinsen et al., 2 Sep 2025).

The dataset’s format reflects its research orientation. By distributing stand-alone text units rather than only page-level or edition-level text, ENO supports analytical workflows that require coherent document boundaries. The inclusion of date, title, and pwa score provides immediate metadata for filtering and temporal analysis, while exclusion of hyperlinks and similarity indices from the main downloadable files reflects an optimization for file size and memory use. In practical terms, ENO is designed for both human-readable exploration through the web interface and machine-readable ingestion through structured files (Heinsen et al., 2 Sep 2025).

Within the broader landscape of historical corpora, ENO’s significance lies in the combination of large-scale neural OCR, explicit layout reconstruction, article-level segmentation, and a historically specialized BERT model built from the corpus itself. This suggests a model for newspaper digitization in which corpus construction, document engineering, and language modeling are mutually constitutive rather than sequentially isolated tasks.