datasets/task_specific_example/README.md

Özel Görevler İçin Datasets

Bu klasör, specific NLP task'leri için dataset örnekleri içerir.

Task'ler

❓ Question Answering

Extractive QA (SQuAD-style)

{
    'context': 'Paris is the capital of France...',
    'question': 'What is the capital of France?',
    'answers': {
        'text': ['Paris'],
        'answer_start': [0]
    }
}

Multiple Choice QA

{
    'question': 'What is 2+2?',
    'choices': ['3', '4', '5', '6'],
    'answer': 1  # Index of correct answer
}

Best Practices:

• Validate answer spans
• Handle impossible questions
• Question type classification
• Context length management

📝 Summarization

News Summarization

{
    'article': 'Long news article...',
    'summary': 'Brief summary...',
    'compression_ratio': 0.24
}

Metrics:

• ROUGE scores
• Compression ratio (20-30% optimal)
• Abstractive vs Extractive

Best Practices:

• Multiple reference summaries
• Length constraints
• Quality validation

🏷️ Named Entity Recognition

BIO Tagging

{
    'tokens': ['John', 'Smith', 'works', 'at', 'Google'],
    'ner_tags': ['B-PER', 'I-PER', 'O', 'O', 'B-ORG']
}

Tag Schema:

• B-PER, I-PER (Person)
• B-ORG, I-ORG (Organization)
• B-LOC, I-LOC (Location)
• O (Outside)

Best Practices:

• Consistent tagging scheme
• Entity type taxonomy
• Nested entities handling
• Entity linking (optional)

😊 Sentiment Analysis

Binary/Multi-class

{
    'text': 'This product is amazing!',
    'label': 2,  # 0: neg, 1: neutral, 2: pos
    'confidence': 0.95
}

Aspect-Based

{
    'text': 'Great product but slow delivery',
    'aspect_sentiments': {
        'product': 'positive',
        'delivery': 'negative'
    }
}

Best Practices:

• Multi-level granularity
• Confidence scores
• Domain-specific lexicons
• Emotion detection

📊 Text Classification

Topic Classification

{
    'text': 'Article text...',
    'label': 'technology',
    'label_id': 0
}

Best Practices:

• Balanced classes
• Hierarchical categories
• Multi-label support
• Class imbalance handling

🎯 Multi-Task Learning

Unified Format

{
    'text': 'Sample text...',
    'sentiment': 'positive',
    'topic': 'technology',
    'quality_score': 0.85
}

Best Practices:

• Consistent preprocessing
• Task-specific heads
• Shared representations
• Task weighting

Dataset Statistics

Task	Örnekler	Format
QA	300	Extractive + MC
Summarization	100	News articles
NER	100	BIO tagged
Sentiment	350	Multi-class + Aspect
Classification	200	Topic
Multi-Task	100	Unified

Quality Metrics

QA

• Exact Match (EM)
• F1 Score
• Answer span accuracy

Summarization

• ROUGE-1, ROUGE-2, ROUGE-L
• Compression ratio
• Factual consistency

NER

• Precision, Recall, F1 per entity type
• Exact match
• Partial match

Sentiment

• Accuracy
• Macro/Micro F1
• Confusion matrix

Classification

• Accuracy
• Per-class F1
• Macro/Weighted F1

Best Practices (Genel)

✅ Clear annotation guidelines
✅ Inter-annotator agreement
✅ Quality control checks
✅ Regular dataset updates
✅ Version control
✅ Documentation
✅ Ethical considerations
✅ Bias analysis