README.md

metadata

language:
  - da
license: cc-by-4.0
size_categories:
  - n<1K
task_categories:
  - image-to-text
pretty_name: Modern Danish Handwriting
tags:
  - OCR
  - HTR
  - handwriting
  - modern
dataset_info:
  features:
    - name: image
      dtype: image
    - name: doc_id
      dtype: int64
    - name: sequence
      dtype: int16
    - name: page
      dtype: string

Dataset Card for Modern Danish Handwriting

The Modern Danish Handwriting dataset is a Danish-language dataset containing more than 200 pages of transcribed and proofread handwritten text.

Dataset Details

Dataset Description

The Modern Danish Handwriting dataset currently consists of handwritten samples of text from the ePAROLE dataset. The samples were created by volunteers at the Danish National Archives and guests at the festival Historiske Dage in 2025.

The ePAROLE dataset is a 2015-version of the PAROLE, which consists primarily of publically available newspaper and magazine articles from the early 1990s.

The Modern Danish Handwriting dataset consists of 223 transcribed images, containing a total of 977 lines, 5632 words, and 33026 characters (including spaces).

• Curated by: Joen Rommedahl
• Language(s): Danish
• License: Creative Commons Attribution 4.0 International

Uses

Direct Use

The dataset has been created as training data for Handwritten Text Recognition (HTR) of modern danish handwriting, but can also be used as training data for baseline detection and polygon extraction models, as these have been carefully created and quality checked as well.

Out-of-Scope Use

The dataset is not particularly suited as training data for text region detection models, as all text on all samples are contained within a single text region, and thus not representative of how documents would look 'in the wild'.

The line segmentations of the Modern Danish Handwriting dataset is not subdivided into individual words, but operate on textlines as the lowest granularity, and thus cannot be used for training word detection and isolation.

Dataset Structure

Each data instance represents a single scanned, segmented and transcribed physical page with handwritten text. The dataset contains the following fields:

• image: a jpeg image containing a scan of the original physical page
• doc_id: internal id of parent document (collection of single pages) to which the page belongs
• sequence: the specific pagenumber of the page within the parent document
• page: an xml-encoded string containing layout and content information of the physical page, stored in PAGE XML, version 2013-07-15

To uniquely identify a single page within the dataset, one can use the doc_id integer in combination with the sequence integer.

Dataset Creation

Curation Rationale

The dataset was created as part of an initiative to let the guests of the Historiske Dage festival explore the proces of handwritten text recognition and training data creation. It also seeks to draw attention to the fact, that no large-scale dataset of modern, danish handwriting exists as a HTR-resource.

Source Data

The source data consists of written samples of the ePAROLE dataset. The original text is primarily news articles from the 1990s and the handwritten samples were created in 2025.

Data Collection and Processing

The original physical samples were set up by the Data Science department at the Danish National Archives. The samples were selected randomly from the ePAROLE dataset, curated by The Society for Danish Language and Literature, Ole Norling-Christensen, Britt-Katrin Keson, Jørg Asmussen and more. See the documentation for ePAROLE for further information on how it was created.

Who are the source data producers?

The digitized and transcribed sentences were originally written by the journalists of various newspaper outlets and magazines in Denmark, and are publically available.

Text samples of varying length were printed out and handed out to volunteers at the Danish National Archives and guests at the festival Historiske Dage in 2025 with the instruction of writing the text by hand underneath the printed sample while keeping line breaks and formating exactly as the printed text.

After re-collecting the samples, the papers were scanned and the original ePAROLE sample id used to infer what text the volunteer had written. It has not been verified that the volunteers actually wrote the text and preserved line breaks as instructed.

Layout and line segmentation was performed on each page with the Loghi/Laypa toolkit, developed by Rutger van Koert and Stefan Klut at KNAW Humanities Cluster, and afterwards manually corrected by the Data Science departments at the Danish National Archives using the Transkribus interface.

Personal and Sensitive Information

The ePAROLE dataset is publically available through The Society for Danish Language and Literature. The Data Science department at the Danish National Archives does not take responsibility for any data therein that might be considered personal, sensitive or private. No efforts have been made to anonymize the data.

Bias, Risks, and Limitations

Due to its limited size we do not expect any model trained only on this dataset to achieve great HTR-results. As the text corpus is primarily based on magazine resources, the language represented by the dataset is defined thereby. Thus, there is probably little language from specialized fields nor informal conversation.

More Information

Thank you to the guests at Historiske Dage who contributed to the dataset! Also thank you to all the people involved with creating and maintaining the ePAROLE dataset at The Society for Danish Language and Literature, as one of the few publically available license-free resources of danish language!

Dataset Card Contact

Point of Contact: Joen Rommedahl