alakxender/dhivehi-noisy-sentences

Dhivehi Noisy Sentences Dataset

This dataset contains parallel examples of clean text and text with introduced errors across three categories: spelling, grammar, and punctuation.

Dataset Description

This dataset is designed to train models that can correct errors in Dhivehi text. Each example consists of:

• clean_text: The correct, error-free Dhivehi text
• noisy_text: The same text with introduced errors
• error_type: The category of error (spelling, grammar, or punctuation)

Dataset Statistics

• Total samples: ~3x the number of source sentences
• Error types: 3 categories (spelling, grammar, punctuation)
• Minimum sentence length: 4 words
• Script: Thaana (Unicode range U+0780-07BF)

Data Splits

• Train: 80%
• Validation: 10%
• Test: 10%

Error Types

1. Spelling Errors (spelling)

Multiple random errors (1-4 per sentence) including:

• Repeated vowels: ެ or ެެެ
• Missing vowels: Random deletion of vowel diacritics
• Wrong vowels: Substitution with different vowels
• Character deletion: Random character removal
• Character insertion: Extra characters added
• Character transposition: Swapping adjacent characters
• Character replacement: Consonant substituted with another
• Vowel replacement: Vowel substituted with another

Severity levels: Light (1-2 errors), Medium (2-3 errors), Heavy (3-4 errors)

2. Grammar Errors (grammar)

Multiple random grammatical errors (1-3 per sentence) including:

• Case suffix errors: Wrong or missing case markers (ގެ، އަށް، ގައި، އިން، އާ)
• Plural marker errors: Incorrect plural forms (ތައް، ން، ކަން)
• Demonstrative errors: Wrong demonstrative usage (މި، ތި، އެ)
• Particle errors: Incorrect sentence-ending particles (އެވެ، ވާނެ، ބައްޔެވެ)
• Negation errors: Misplaced or missing negation (ނު، ނޫން، ނެތް)
• Tense errors: Wrong tense markers (ނީ، ނި، ފި، ނަން)

Grammar patterns covered: 17 categories including case suffixes, plurals, demonstratives, definiteness, adverbials, tense/aspect, conditionals, auxiliaries, particles, negation, questions, postpositions, emphasis, conjunctions, verb endings, and honorifics.

3. Punctuation Errors (punctuation)

Issues with punctuation marks:

• Missing sentence-ending punctuation (., ؟, !, ؛)
• Incorrect spacing around punctuation
• Missing Arabic comma (،)
• Mixed Western/Arabic punctuation issues

Both naturally occurring errors from source data and synthetically generated errors.

Data Format

Each record contains three fields:

{
  "clean_text": "އައިއޯއެސް 26 އަށް އަޕްގްރޭޑްކުރުމަށް އެޕަލްއިން ބާރުއަޅަނީ.",
  "noisy_text": "އައިއޯއެސް 26 އަށް އަޕްގްރޭޑްކުރުުމަށް އެޕަލްއިން ބާރުުއަޅަނީ.",
  "error_type": "spelling"
}

Source Data

Generated from a new corpus of Dhivehi news articles and web content. All sentences have:

• Minimum 4 words
• Valid Thaana script characters
• Preserved punctuation in source sentences

Use Cases

This dataset is suitable for:

• Training Dhivehi spell checkers
• Building grammar correction tools
• Developing punctuation restoration models
• Creating general text correction systems for Dhivehi
• Fine-tuning multilingual models for Dhivehi

Loading the Dataset

from datasets import load_from_disk

# Load the dataset
dataset = load_dataset("alakxender/dhivehi-noisy-sentences")

# Access splits
train_data = dataset['train']
val_data = dataset['validation']
test_data = dataset['test']

# Iterate through examples
for example in train_data:
    clean = example['clean_text']
    noisy = example['noisy_text']
    error_type = example['error_type']  # 'spelling', 'grammar', or 'punctuation'