DenisaB/htr_leibniz_dataset_v1

Dataset Card for Leibniz's Manuscripts (HTR Ground Truth)

This dataset is composed of trascribed and manually corrected folios of Gottfried Wilhelm Leibniz's manuscripts, together with automatically aligned ground truth in order to train or fine-tune Handwritten Text Recognition (HTR) models.

Dataset Details

This ground truth was produced to fine-tune existing HTR models for the recognition of Leibniz's handwriting, with the aim of assisting scholars in the production of critical editions and providing larger access to the text for research purposes.

The corpus was created in two phases. In the first phase, 275 pages (20,139 lines) were manually corrected. This subset is designated "clean" within the corpus. Part of this data was later used to train FoNDUE-GD-v2, giving a first robust HTR model.

Following this initial phase, a different strategy was adopted: aligning existing ground truth with the output of the FoNDUE-GD-v2 model, yielding 43,372 lines across 735 pages. This step relied on the availability of a sufficiently accurate HTR model. The resulting data, designated "noisy" may contain minor alignment errors, such as words from previous or following line being misassigned to the current line, or fragments of marginal additions appearing at line boundaries due to imperfect text extraction. Only lines with a similarity score of ≥ 0.7 (Levenshtein distance) and a length between 80% and 100% of the corresponding original HTR line were retained. HTR lines without an equivalent aligned ground truth line were left empty.

The validation set is composed of 27 files selected from the "clean" dataset.

Dataset Sources [optional]

• All images have been extracted from the Digitale Sammlungen der Gottfried Wilhelm Leibniz Bibliothek in the IIIF format.
• The ground truth used for the alignments was extracted from the Series VI, Volume 4 of Leibniz's philosophical writings: AVI-4

Uses

Direct Use

• Handwritten Text Recognition

Dataset Structure

All training data is present in the "data" folder. Both training data (clean and noisy) were split into training and validation sets using a 90/10 ratio.

• train_clean_.* and train_noisy_.* files compose the training set.
• val_clean_.* represents the validation set.
• no test set was created for this training, in order to maximize the amount of data available for training.

The original formats (image + XML Page files) can be found directly in the train and val folders.

Dataset Creation

Data Collection and Processing

We selected manuscript pages with few or no mathematical expressions, since HTR models struggle to recognize multi-level expressions. The corpus is composed of philosophical writings, physics manuscripts, correspondance, drafts and clean copies, pages with different degrees of layout complexity.

Who are the source data producers?

ERC Philiumm, Laboratoire SPHERE, Univ. Paris Cité - CNRS, France

Project Manager : Rabouin, David

Annotations [optional]

Annotation process

Text annotation guidelines Layout annotation guidelines

Who are the annotators?

We are specially grateful to the contributors listed below, as well as to the Berlin team, for their contribution to the creation of the ground truth data used in the development of the FoNDUE-GD_v2 model and of this fine-tuned version of it.

• Bumba, Denisa-Florina
• Giovannangeli, Vincent
• Rabouin, David
• Jaffrezic, Audrey
• Garnier, Mathias
• Michel, Florian
• Cai, Jiarui
• Giraud-Hua, Sébastien

Bias, Risks, and Limitations

This dataset is specialized on one's author's writings and over a small period of time (17th-early 18th c.).

Funding

This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme, Advanced Grant ADG No. 101020985.