metadata
license: mit
task_categories:
- text2text-generation
language:
- fr
Description
The dataset is sourced from the French Senate's records, available at this link. After segmenting the text into sentences, 200,000 sentences were extracted(160,000 for train, 40,000 for test).
OCR errors are generated using the following script, designed to simulate errors found in certain documents from the French National Library caused by physical constraints (e.g., page curvature).
import random
import re
from typing import Dict, List
class FrenchOCRNoiseGenerator:
def __init__(self, error_density: float = 0.45):
self.error_density = max(0.0, min(1.0, error_density))
# Adjust error type weight distribution
self.error_config = {
'diacritics': 0.25, # Diacritic errors
'case_swap': 0.15, # Case swapping
'symbols': 0.18, # Symbol substitution
'number_letter': 0.12, # Number-letter confusion
'split_words': 0.08, # Word splitting errors
'random_space': 0.10, # Random spacing
'hyphen_error': 0.07, # Hyphen-related errors
'repeat_char': 0.05 # Character repetition
}
# Extended French-specific replacement mapping
self.french_replace_map: Dict[str, List[str]] = {
# Diacritic characters
'é': ['e', '3', '€', 'è', 'ê', 'ë', ' '],
'è': ['e', '€', 'é', 'ë', '¡'],
'ê': ['e', 'è', 'ë', 'â'],
'à': ['a', '@', 'â', 'á', 'ä', 'æ'],
'ù': ['u', 'û', 'ü', '%'],
'ç': ['c', '¢', '(', '[', ' '],
'â': ['a', 'à', 'ä', 'å'],
'î': ['i', 'ï', '!', '1'],
'ô': ['o', 'ö', '0', '*'],
'û': ['u', 'ù', 'ü', 'v'],
# Uppercase letters
'É': ['E', '3', '€', 'È'],
'È': ['E', 'Ê', 'Ë'],
'À': ['A', '@', 'Â'],
'Ç': ['C', '¢', ' '],
# Number-letter confusion
'0': ['o', 'O', 'Ø', ' '],
'1': ['l', 'I', '!', '|'],
'2': ['z', 'Z', 'è'],
'5': ['s', 'S', '$'],
'7': ['?', ' ', '~'],
'9': ['q', 'g', ' '],
# Special symbols
'-': ['~', '—', '_', ' ', '.', ''],
"'": ['`', '’', ',', ' '],
'’': ["'", '`', ' '],
',': ['.', ';', ' '],
';': [',', ':', ' ']
}
# Common OCR error patterns
self.ocr_patterns = [
(r'(.)\1{2,}', lambda m: m.group(1)*random.randint(1,3)), # Handle repeated characters
(r'(?i)tion', lambda m: 't10n' if random.random() < 0.3 else m.group()),
(r'(?i)ment\b', lambda m: 'm'+random.choice(['&', '@', '3'])+'nt')
]
def _apply_diacritic_errors(self, char: str) -> str:
"""Process diacritic-related errors"""
if char in {'é', 'è', 'ê', 'à', 'ù', 'ç', 'â', 'î', 'ô', 'û'}:
# 30% chance to remove diacritics completely
if random.random() < 0.3:
return char.encode('ascii', 'ignore').decode() or char
# 70% chance to replace with another diacritic character
return random.choice(self.french_replace_map.get(char, [char]))
return char
def _hyphen_errors(self, text: str) -> str:
"""Handle hyphen-related errors"""
return re.sub(r'[-—]', lambda m: random.choice(['~', '_', ' ', '.', ''] + ['']*3), text)
def _apply_ocr_patterns(self, text: str) -> str:
"""Apply predefined OCR error patterns"""
for pattern, repl in self.ocr_patterns:
text = re.sub(pattern, repl, text)
return text
def _generate_errors(self, char: str) -> str:
"""Core error generation logic"""
if random.random() > self.error_density:
return char
# French-specific processing
if char in self.french_replace_map:
return random.choice(self.french_replace_map[char])
error_type = random.choices(
list(self.error_config.keys()),
weights=list(self.error_config.values())
)[0]
# Apply different types of errors
if error_type == 'diacritics':
return self._apply_diacritic_errors(char)
elif error_type == 'case_swap' and char.isalpha():
return char.swapcase() if random.random() < 0.6 else char.lower()
elif error_type == 'symbols':
return random.choice(['~', '*', '^', '¡', '¦']) if random.random() < 0.4 else char
elif error_type == 'number_letter':
return random.choice(self.french_replace_map.get(char.lower(), [char]))
elif error_type == 'repeat_char' and char.isalnum():
return char * random.randint(1,3)
return char
def _space_errors(self, text: str) -> str:
"""Optimized space-related errors: more deletions, fewer insertions"""
# Randomly delete spaces (25% chance to remove any space)
text = re.sub(r' +', lambda m: ' ' if random.random() > 0.25 else '', text)
# Randomly add spaces (10% chance to insert a space before a character)
return re.sub(
r'(?<! )(?<!^)',
lambda m: ' ' + m.group() if random.random() < 0.1 else m.group(),
text
)
def add_noise(self, clean_text: str) -> str:
# Preprocessing: Apply OCR patterns
noisy_text = self._apply_ocr_patterns(clean_text)
# Inject character-level errors
noisy_chars = [self._generate_errors(c) for c in noisy_text]
# Post-processing steps
noisy_text = ''.join(noisy_chars)
noisy_text = self._hyphen_errors(noisy_text)
noisy_text = self._space_errors(noisy_text)
# Final cleanup
return re.sub(r'\s{2,}', ' ', noisy_text).strip()
ocr_generator = FrenchOCRNoiseGenerator(error_density=0.4)
df['OCR Text'] = df['Ground Truth'].apply(ocr_generator.add_noise)