Data Quality Standards for Somali Dialect Classifier

Quality Dimensions

1. Completeness

• All required fields present (text, label, source, timestamp)
• No null or empty text fields
• Labels properly assigned (Northern/Southern/Central)

2. Accuracy

• Text is in Somali (not English, Arabic, or other languages)
• Labels match actual dialect (validated by native speakers)
• Geographic metadata aligns with dialect labels

3. Consistency

• Uniform text encoding (UTF-8)
• Consistent label format (standardized names)
• Timestamp format standardized (ISO 8601)

4. Uniqueness

• No exact duplicates
• Near-duplicate detection (>95% similarity flagged)
• Source URL deduplication

5. Validity

• Text length within acceptable range (10-5000 characters)
• No corrupted/garbled text
• No HTML tags or formatting artifacts

Quality Metrics

Critical Metrics

Language Purity:

• Target: >98% Somali text
• Method: Language detection (langdetect, fastText)
• Action: Remove non-Somali text

Duplicate Rate:

• Target: <2% duplicates
• Method: Exact match + fuzzy matching (Levenshtein distance)
• Action: Keep first occurrence, remove duplicates

Label Confidence:

• Target: >90% inter-annotator agreement
• Method: Multiple annotators for sample
• Action: Re-label low-confidence examples

Text Quality Score:

• Target: Average score >7/10
• Components: Length, vocabulary richness, grammar
• Action: Filter texts with score <5

Validation Pipeline

Stage 1: Basic Validation

def basic_validation(record):
    checks = {
        'has_text': bool(record.get('text', '').strip()),
        'has_label': record.get('label') in ['Northern', 'Southern', 'Central'],
        'valid_length': 10 <= len(record.get('text', '')) <= 5000,
        'valid_encoding': is_valid_utf8(record['text'])
    }
    return all(checks.values()), checks

Stage 2: Language Detection

from langdetect import detect

def validate_language(text):
    try:
        lang = detect(text)
        return lang == 'so'  # Somali ISO code
    except:
        return False

Stage 3: Duplicate Detection

from difflib import SequenceMatcher

def is_near_duplicate(text1, text2, threshold=0.95):
    similarity = SequenceMatcher(None, text1, text2).ratio()
    return similarity >= threshold

Stage 4: Quality Scoring

def compute_quality_score(text):
    score = 0
    # Length appropriateness (1-3 points)
    if 50 <= len(text) <= 1000:
        score += 3
    elif 20 <= len(text) < 50 or 1000 < len(text) <= 3000:
        score += 2
    else:
        score += 1

    # Vocabulary richness (1-3 points)
    unique_words = len(set(text.split()))
    total_words = len(text.split())
    if total_words > 0:
        vocab_ratio = unique_words / total_words
        if vocab_ratio > 0.7:
            score += 3
        elif vocab_ratio > 0.5:
            score += 2
        else:
            score += 1

    # No HTML/formatting artifacts (1-2 points)
    if not ('<' in text or '>' in text or '{' in text):
        score += 2

    # Proper sentences (1-2 points)
    if text.count('.') >= 1:  # At least one sentence
        score += 2

    return min(score, 10)  # Cap at 10

Acceptance Criteria

Minimum Quality Thresholds

For Training Set:

• Language purity: >98% Somali
• Duplicate rate: <1%
• Quality score: Average >7.5
• Label confidence: >95%

For Validation/Test Sets:

• Language purity: >99% Somali
• Duplicate rate: 0% (strict)
• Quality score: Average >8.0
• Label confidence: >98% (manually validated)

Quality Guardrails

Automatic Filters

1. Remove if:

   • Non-Somali language detected
   • Exact duplicate found
   • Text length <10 or >5000 characters
   • Quality score <5
   • Contains >20% numbers/special characters

2. Flag for review if:

   • Near-duplicate (>95% similarity)
   • Quality score 5-7
   • Label confidence <90%
   • Unusual character patterns

3. Accept if:

   • All validation checks pass
   • Quality score ≥7
   • No duplicates
   • Language = Somali

Quality Reporting

Metrics to Track

Dataset-Level:

• Total records
• Records passing validation (%)
• Average quality score
• Duplicate count
• Language distribution (% Somali)

Per-Source:

• Source name
• Records contributed
• Average quality score
• Duplicate rate
• Rejection rate

Per-Dialect:

• Dialect label
• Record count
• Average quality score
• Inter-annotator agreement

Example Report:

Dataset Quality Report - 2025-11-06

Total Records: 10,000
Passing Validation: 9,200 (92%)
Average Quality Score: 7.8/10
Duplicates Removed: 600 (6%)
Language Purity: 98.5% Somali

Per-Source Quality:
- Wikipedia: 8.5/10 (3,000 records)
- BBC Somali: 8.2/10 (2,500 records)
- Social Media: 6.9/10 (4,500 records, 30% rejected)

Per-Dialect Distribution:
- Northern: 5,500 (59.8%)
- Southern: 2,200 (23.9%)
- Central: 1,500 (16.3%)

When This Skill Activates

This skill auto-invokes when you mention:

• Data quality, data validation, quality checks
• Duplicates, deduplication, duplicate detection
• Quality metrics, quality score, quality standards
• Data cleaning, data filtering, guardrails
• Language detection, language purity
• Acceptance criteria, validation rules

Version: 1.0.0 Last Updated: 2025-11-06 Project: Somali Dialect Classifier