Datasets:
language: la
license: cc-by-sa-4.0
multilinguality: monolingual
size_categories:
- 10K<n<100K
source_datasets:
- original
task_categories:
- translation
task_ids:
- text2text-generation
pretty_name: Burchard's Decretum Dataset Abbreviations Augmented
annotations_creators:
- expert-generated
Dataset Card for bdd-abbreviations-augmented
Dataset Description
Dataset Summary
- Burchards Dekret Digital (BDD): Website
- Source data: BDD Gitlab
- Trainingscripts: Abbreviationes GitHub
- Zenodo: Zenodo
- Contact: Michael Schonhardt (michael.schonhardt@tu-darmstadt.de, ORCID)
The BDD Abbreviations Augmented dataset is a parallel corpus for automated abbreviation expansion in transcriptions from 11th-century Latin manuscripts. It provides paired examples of abbreviated text (source) and its corresponding full expansion (target).
The dataset was generated from the five key manuscripts of the Decretum Burchardi, a major work of medieval canon law, transcribed by the Burchards Dekret Digital (BDD) project. It is specifically designed for fine-tuning byte-level sequence-to-sequence models like ByT5 to assist in the creation of digital critical editions.
The core task is to translate a "graphemic" transcription, which preserves scribal abbreviations using special Unicode characters (MUFI), into a fully readable, expanded Latin text.
Supported Tasks
abbreviation-expansion: The dataset is intended for training models on a sequence-to-sequence basis. The model takes thesource_textas input and generate thetarget_text.
Languages
The text is in Medieval Latin (la).
Dataset Structure
Data Instances
A typical data instance consists of a source string with abbreviations and a target string with the expansions.
{
"source_text": "om̅s posteri eorū cuncta sibi uendicarent sed semꝑ maiores causę sicut s̅ ep̅oꝝ et potiorū curę negotiorū ad unā beati principis apꝉoꝝ petri sedē confluerent ut inde",
"target_text": "omnes posteri eorum cuncta sibi uendicarent sed semper maiores causę sicut sunt episcoporum et potiorum curę negotiorum ad unam beati principis apostolorum petri sedem confluerent ut inde"
}
Data Fields
source_text: A string containing the abbreviated Latin text from the manuscript, represented using standard and special Unicode characters.
target_text: A string containing the corresponding fully expanded Latin text.
Data Splits
The dataset contains a single train split.
Source Data
The data originates from the final, manually corrected TEI-XML transcriptions of five key 11th-century manuscripts of Burchard's Decree, produced in the scriptorium of Worms:
B: Bamberg, Staatsbibliothek, Msc. Can. 6
F: Frankfurt, Universitätsbibliothek, Ms. Barth. 50
K: Köln, Erzbischöfliche Diözesan- und Dombibliothek, Cod. 119
V2: Vatican, Biblioteca Apostolica Vaticana, Pal. lat. 585
V3: Vatican, Biblioteca Apostolica Vaticana, Pal. lat. 586
Annotations
The dataset was generated automatically from tei source files. The process was as follows:
Annotation: Editors from the BDD project transcribed and expanded all abbreviations in the source manuscripts, encoding the results in TEI-XML using the element with (abbreviation) and (expansion) children.
Pair Generation: A script processed the final TEI files. It used a 3-line sliding window to provide context. For each window, it extracted the text from the tags to create the source_text and the text from the tags to create the target_text. Rare abbreviations were automatically multiplied to represent them well in the final model.
Considerations for Using the Data
The dataset is highly specialized. It reflects the scribal practices of a single scriptorium (Worms) from a specific time period (early 11th century). The system of abbreviations is consistent within the corpus but may not be representative of other medieval scripts, regions, or time periods. A model trained on this data will perform best on similar manuscripts and will likely require further fine-tuning to be used on texts from other traditions. As the dataset is augmented to better represent rare brevigraphs, the trained model might fail in instances where distribution of brevigraphs differs.
The model trained on this dataset is designed to expand a fixed set of brevigraphs and special characters recommended by the Medieval Unicode Font Initiative (MUFI). It will not be able to expand abbreviations that are not represented in the training data. The main characters are listed below:
Dataset Curators
The dataset was generated by Michael Schonhardt (TU Darmstadt / Akademie der Wissenschaften und der Literatur | Mainz), using transcriptions prepared by the Burchards Dekret Digital Project Team. Individual contributions are encoded in TEI files.
Licensing Information
Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).
Citation Information
If you use this dataset in your research, please cite the original project and the dataset.
@misc{schonhardt_bdd_abbreviations_2025,
author = {Schonhardt, Michael},
title = {BDD Abbreviations Augmented: A Parallel Corpus for Abbreviation Expansion in Medieval Latin},
year = {2025},
institution = {Burchards Dekret Digital},
doi={https://doi.org/10.5281/zenodo.16735498},
howpublished = {\url{[https://huggingface.co/datasets/mschonhardt/bdd-abbreviations-augmented](https://huggingface.co/datasets/mschonhardt/bdd-abbreviations-augmented)}}
}