README.md

metadata

license: cc-by-4.0
task_categories:
  - image-to-text
language:
  - fr
tags:
  - synthetic
  - print
  - automatic-text-recognition
size_categories:
  - 1K<n<10K
pretty_name: Noisy Ground Truth - Missing Words

Noisy Ground Truth - Missing Words

Dataset of synthetic data for experimentation with noisy ground truth. The text in the dataset is based on Colette's Sido and Les Vignes, also the data was processed prior to generating images with the TextRecognitionDataGenerator.

In Noisy Ground Truth - Missing Words, each variation column is affected by the noise, without considering the split between train, validation and test.

Dataset Description

• created by: Alix Chagué
• license: CC BY 4.0
• languages: French

Data structure

The dataset is composed of the following columns:

• gold: the original text
• source: identifies the original work from which the text was extracted
  • ex: "sido-colette" means that the text was extracted from Colette's Sido (or Les Vignes)
• MW-F: the first word is missing
• MW-L: the last word is missing
• MW-50: half of the words are missing, chosen randomly
• MW-50-L: half of the words are missing, chosen from the end of the text
• MW-1: one word is missing, chosen randomly
• MW-1-POS: position of the missing word in MW-1
• n_line: number identifying the line in page # n_page
• n_page: number identifying the page (the page changes every 25 lines)
• split: split between train, validation and test (applied at page level)
• im: synthetic text line image

Dataset Card Contact

Alix Chagué (first.last@inria.fr)