README.md

metadata

dataset_info:
  features:
    - name: File
      dtype: string
    - name: Date
      dtype: int64
    - name: Region_OCR
      dtype: string
    - name: Region_OCR_aligned
      dtype: string
    - name: Region_GT_aligned
      dtype: string
    - name: Sentence_OCR_aligned
      dtype: string
    - name: Sentence_GT_aligned
      dtype: string
    - name: Sentence_OCR
      dtype: string
    - name: Sentence_GT
      dtype: string
    - name: Distance
      dtype: int64
    - name: CER
      dtype: float64
    - name: WER
      dtype: float64
    - name: Sentence_OCR_corrupted
      dtype: string
    - name: corrupted_cer
      dtype: float64
    - name: corrupted_wer
      dtype: float64
  splits:
    - name: train
      num_bytes: 18654472
      num_examples: 2632
    - name: dev
      num_bytes: 2542628
      num_examples: 336
    - name: test
      num_bytes: 2031987
      num_examples: 301
  download_size: 6813801
  dataset_size: 23229087
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev
        path: data/dev-*
      - split: test
        path: data/test-*

This dataset is a filtered version of the ICDAR2017 Competition on Handwritten Text Recognition, focusing on monograph texts written between 1800 and 1900. It consists of a total of 957 documents, divided into training, validation, and testing sets, and is designed for post-correction of OCR (Optical Character Recognition) text.

• Training Set: 2.63k phrases
• Validation Set: 336
• Test Set: 301

Purpose

The dataset aims to improve the accuracy of digitized texts by providing a reliable Ground Truth for comparison and correction, specifically addressing the challenges of French text of 19th century.

Dataset Columns Description

This dataset contains several columns with detailed information about OCR (Optical Character Recognition) outputs and their corresponding ground truths (GT). Below is a description of each column:

Column Name	Data Type	Description
File	string	Name or identifier of the file associated with the OCR and GT data.
Date	int64	Date of the document, encoded as an integer.
Region_OCR	string	OCR-recognized region of text.
Region_OCR_aligned	string	Aligned OCR-recognized region for better comparison with GT data.
Region_GT_aligned	string	Aligned ground truth region for validation against OCR results.
Sentence_OCR_aligned	string	Aligned OCR-recognized sentences for improved accuracy in comparisons.
Sentence_GT_aligned	string	Aligned ground truth sentences corresponding to OCR results.
Sentence_OCR	string	Original OCR-recognized sentences, unaligned.
Sentence_GT	string	Original ground truth sentences, unaligned.
Distance	int64	The computed edit distance between OCR sentences and ground truth.
CER	float64	Character Error Rate (CER) for OCR sentences compared to the ground truth. Values range from 0 to 0.29.
WER	float64	Word Error Rate (WER) for OCR sentences compared to the ground truth. Values range from 0 to 1.5.
Sentence_OCR_corrupted	string	OCR sentences with additional corruption introduced for synthetic error analysis.
corrupted_cer	float64	Character Error Rate for corrupted OCR sentences compared to the ground truth. Values range from 0 to 0.35.
corrupted_wer	float64	Word Error Rate for corrupted OCR sentences compared to the ground truth.

This structured dataset is designed to analyze and evaluate OCR post-correction performance with both real and synthetic data.

Author Information

Prepared by Mikhail Biriuchinskii, an engineer in Natural Language Processing at Sorbonne University.

Original Dataset Reference

For more information, visit the original dataset source: ICDAR2017 Competition on Post-OCR Text Correction.

Copyright

The original corpus is publicly accessible, and I do not hold any rights to this deployment of the corpus.