TAPUSCORPUS

Description

Ground Truth dataset for French typewritten OCR

Content

154 pairs of images and PAGE XML files divided into 9 sub-corpus.

#	name	nb of images	Date	GT for segmenter?	GT for recognizer?	link(s) to source images
1	12148-btv1b52502601r	(15)	X	y	y	https://gallica.bnf.fr/ark:/12148/btv1b52502601r
2	20130043_87-1962-catalogue-dactylographié-des-enregistrements	(14)	1962	n	y	https://www.siv.archives-nationales.culture.gouv.fr/siv/UD/FRAN_IR_050603/c-2cfmj84sh--fdjfx0b91tk6
3	bljd-DSN390-DSN355-DSN113_(27)-DSN169	(14)	1922-1945	y	y	http://bljd.sorbonne.fr/ark:/naan/a011441804309XrO1fa/e3ff656f45 , http://bljd.sorbonne.fr/ark:/naan/a011441804309Dy4ooR/8f9020bc6e , http://bljd.sorbonne.fr/ark:/naan/a01144180430809P21l/036952bd71 , http://bljd.sorbonne.fr/ark:/naan/a0114418043093CzACj/ae2ff446cd
4	12148-btv1b10581912k	(12)	19??	y	y	https://gallica.bnf.fr/ark:/12148/btv1b10581912k
5	AN-96AP_1-dossier_1-pièces_1_à_202-début-du-journal-dactylographié	(9)	1914-1917	y	y	https://www.siv.archives-nationales.culture.gouv.fr/siv/UD/FRAN_IR_050082/c-65dxt4cy4--yqhpaasluvz1
6	12148-btv1b525062185	(4)	X	y	y	https://gallica.bnf.fr/ark:/12148/btv1b525062185
7	12148-btv1b10583038c	(6)	1896	y	y	https://gallica.bnf.fr/ark:/12148/btv1b10583038c
8	12148-btv1b10583138r	(2)	1962	y	y	https://gallica.bnf.fr/ark:/12148/btv1b10583138r
9	12148-btv1b53097347b	(10)	1984	y	y	https://gallica.bnf.fr/ark:/12148/btv1b53097347b
10	europeana-random-selection	(64)	X	n	y	Europeana

(3): images are blurred (low quality), but readable none the less.

Unfortunately, I did not keep track of the URL related to the images taken at random from Europeana...

Annotation system

Undescored words are preceded with _such as "This is an example" will be transcribed as: "_This _is _an example".

Portions of text that are superscripted are preceded with ^ such as "1^er" will be transcribed as "1^er". If several words are superscripted, each word starts with a "^".

Crossed out words are not rendered:

words that be read under the stroke are transcribed as if not crossed out;
words that cannot be read under the stroke are transcribed like any portion of text that is not type-written.

Any portion of text that is not type-written is transcribed as a series of ~~~ (always 3). There are as many repetition of ~~~ (with a space between each instance) as there are words.

How to cite

This dataset was built and is maintained by Alix Chagué (@alix-tz). The original works and their digitization are all copyright-free, but properly annotating a corpus takes time and is a task that should be recognized. If you use any item from this corpus of ground truth, cite the dataset using the following information:

name: 'Tapuscorpus'
url: 'https://github.com/HTR-United/tapuscorpus'
author: 'Alix Chagué'
month: 'janv'
year: '2021'
version: '{any version}'
description: 'Ground Truth dataset for French typewritten OCR (20th century documents)'
language: 'French'
time: '1900-1999'
hands: '30'
license:
    - {name: 'CC-BY 4.0', url: 'https://creativecommons.org/licenses/by/4.0/'}
format: PAGE-XML
volume:
    - {count: "150", metric: pages}

This work is licensed under a Creative Commons Attribution 4.0 International License.