Hugging Face's logo

Datasets:
CATMuS
/
modern

Tasks:
Image-to-Text
Modalities:
Image
Text
Formats:
parquet
Languages:
French
German
English
Size:
100K - 1M
Tags:
optical-character-recognition
humanities
handwritten-text-recognition
modern documents
contemporary documents
good quality
Libraries:
Datasets
Dask
License:
Dataset card Viewer Files Community
1
modern / README.md
alix-tz's picture
alix-tz
Update README.md
a23ea51 verified
preview code
|
raw
history blame contribute delete
14.1 kB
metadata
license: cc-by-4.0
task_categories:
  - image-to-text
version: 1.0.0
language:
  - fr
  - de
  - en
  - it
  - es
  - oc
  - la
pretty_name: CATMuS modern (or McCATMuS)
size_categories:
  - 100K<n<1M
tags:
  - optical-character-recognition
  - humanities
  - handwritten-text-recognition
  - modern documents
  - contemporary documents
  - good quality

CATMuS Modern Banner

Dataset Card for CATMuS Modern and Contemporary (McCATMuS)

Join our Discord to ask questions about the dataset: Join the Discord

Dataset Details

Handwritten Text Recognition (HTR) has emerged as a crucial tool for converting manuscripts images into machine-readable formats, enabling researchers and scholars to analyze vast collections efficiently. Despite significant technological progress, establishing consistent ground truth across projects for HTR tasks, particularly for complex and heterogeneous historical sources, remains nonetheless challenging.

We introduce the Consistent Approaches to Transcribing Manuscripts (CATMuS) dataset for modern and contemporary manuscripts (McCATMuS), which offers:

  • a uniform framework framework for annotating modern and contemporary manuscripts;
  • a benchmarking environment for evaluating automatic text recognition models across multiple dimensions, enriched with metadata such as century, language, and writing technique;
  • a benchmarking environment for additional tasks like script classification and dating approaches;
  • a benchmarking environment for exploratory work in computer vision and digital paleography, particularly for line-based tasks including generative approaches.

Built upon datasets from institutions and projects committed to Open Science, McCATMuS provides an interoperable dataset encompassing over 180 manuscripts in 8 different languages. It includes more than 118,000 lines of text and nearly 4 million characters, covering a period from the early 16th century to the present day.

All the datasets were automatically or, when precised, manually corrected to correspond to the CATMuS guidelines, available here: https://catmus-guidelines.github.io/

Dataset Description

  • Curated by: Alix Chagué
  • Language(s) (NLP): French and Middle French, Spanish, Italian, English, Latin, German, Occitan
  • License: CC-BY 4.0

Train

Writing Type Total Count Languages and Counts
Handwritten 71296 French: 65844, Spanish: 2864, German: 1940, English: 390, Italian: 258
Printed 34684 French: 30376, Middle French: 1873, Latin: 1592, Italian: 266, Occitan: 258, German: 174, English: 85, Spanish: 60
Typewritten 298 English: 298

Validation

Writing Type Total Count Languages and Counts
Handwritten 3833 French: 3662, Spanish: 149, English: 21, German: 1
Printed 1825 French: 1608, Middle French: 115, Latin: 82, Occitan: 12, Spanish: 4, English: 3, German: 1
Typewritten 18 English: 18

Test

Writing Type Total Count Languages and Counts
Handwritten 3898 French: 3724, Spanish: 152, English: 21, German: 1
Printed 1760 French: 1546, Middle French: 115, Latin: 82, Occitan: 12, English: 3, Spanish: 1, German: 1
Typewritten 18 English: 18

Uses

Direct Use

  • Handwritten Text Recognition
  • Date classification
  • Script classification

Out-of-Scope Use

  • Text-To-Image

Dataset Structure

  • Data contains the main split that can be loaded through load_dataset("CATMuS/modern")
  • Data can also be split with each manuscript represented in train, val and test using the gen_split columns which roughly results in a 90/5/5 split
  • The image is in the im column, and the text in the text column
  • Each text line is dated with the combination of not_before and not_after, the precision of the dating can very greatly depending on the available metadata
  • Each text line is associated to a genre, a writing_type (printed, handwritten or typewritten), a region_type and a line_type following SegmOnto's vocabulary, a shelfmark identifying the documents from which the text line is extracted, and a project identifying the project having produced the original dataset.
  • When shelfmark contains "nobs", it means that the documents are not associated to any known shelfmark, the shelfmark in these cells was thus created for the purpose of this metadataset.

Annotations

Annotation process

The annotations in this dataset result:

  • for layout extraction, line extraction, typing and transcription, from the original creators of the dataset in most cases, or from automatic or manual corrections by the curator of the CATMuS modern dataset,
  • for the rest of the metadata, from automatic or manual collection of the metadata by the curator of the CATMuS modern dataset.

The metadata where set generally set at document level.

The values inregion_type and line_type are, as much as possible, conformant with the SegmOnto vocabulary.

The values in writing_type and genre follow a vocabulary set for this dataset:

  • possible values in writing_type are: handwritten, printed or typewritten.
  • possible values in genre are: document of practice, drama, narratives, epistolary, treaties, poetry. There can be multiple values, in which case they are separated by semi-colons.

The detail of the annotation rules applied for the transcription of the text can be found at https://catmus-guidelines.github.io/.

Who are the annotators?

This list includes all the annotators identified by the producers of the datasets gathered in McCATMuS dataset.

Collecting the names of the authors and annotators of datasets can be combersome. If you think your name was mistakenly added to the list below or if your name is missing, please accept our apologies and do get in touch!

  • Chagué, Alix
  • Clérice, Thibault
  • Gabay, Simon
  • Pinche, Ariane
  • Carrow, Jennifer
  • Albert, Anaïs
  • Bey, Laura
  • Champougny, Kevin
  • Charbonnier, Pauline
  • Chiaretti, Alessandro
  • Christensen, Kelly
  • Cicchini, Marco
  • Clavaud, Florence
  • Davoury, Baudoin
  • de Champs, Emmanuelle
  • Dechavanne, Sylvie
  • Denis, Nathalie
  • Doat, Soline
  • Dubourg Glatigny, Pascal
  • Durand, Marc
  • Elsa, Falcoz
  • Fabert, Eliott
  • Faure, Margaux
  • Genero, Jean-Damien
  • Guimarães, Ingrid
  • Humeau, Maxime
  • Jacsont, Pauline Jahan, Claire
  • Jaureguy, Yvan
  • Le Fourner, Victoria
  • Limon-Bonnet, Marie-Françoise
  • Martini, Manuela
  • Maurel, Perrine
  • Mazoue, Anais
  • Meissel, Nina
  • Mikhalchuk, Anna
  • Nahon, Peter
  • Norindr, Jade
  • Nougaret, Christine
  • Ozturk, Yagmur
  • Paupe, Elodie
  • Pérez, Gilles
  • Rebetez, Jean-Claude
  • Riondet, Charles
  • Rostaing, Aurélia
  • Skilbeck-Gaborit, Eden
  • Van Kote, Elsa
  • Vanneau, Laurie
  • Vlachou-Efstathiou, Malamatenia
  • Weddigen, Tristan
  • Wojszvzyk, Elise
  • ALemoine
  • ASJPeronneau
  • Alcofrybas
  • BeaLct
  • CLbt
  • Chloelsa
  • DMichel
  • Desauthieux
  • EPerrin
  • EdChamps
  • GBMireille
  • GPINET
  • Genea78
  • JMGoux
  • Jideuxhemme
  • LBIsabelle
  • Lamotte
  • MFGarreau
  • MIna
  • Maniet
  • MarionJo
  • PGambette
  • PPocard
  • PROMBAUT
  • PaulineTest
  • SCayeux
  • SL.
  • SLespinasse
  • Silver08
  • TPellé
  • Valérie
  • alp
  • jmorvan
  • lelia
  • majubama
  • mickael.lefevr
  • sgauthier

Software

The software used to generate this version of the dataset was built by Thibault Clérice and Alix Chagué.

Reused datasets

All the datasets reused to create the CATMuS Modern and Contemporary dataset are listed below along with the version we used. They can also be found in the Zotero group dedicated to this metadataset.

Bias, Risks, and Limitations

The data is skewed toward French which is overly represented in the current version of the dataset.

No language is represented over all centuries and all writing type, but French has the better coverage for handwritten text lines.

Only one document is available in Spanish. Occitan is only represented in printed lines.

Since the metadata were set at document level, some lines may incorrectly be set to handwritten when they are in fact printed, and inversely. Further versions of the dataset will aim to reduce this phenomenon.

Citation

TBD

Information on the creation process for this dataset can be found in several blog posts: https://alix-tz.github.io/phd/categories/catmus/

Dataset Card Contact

Alix Chagué (first.last@inria.fr)