scripts/analyze_dataset.py

#!/usr/bin/env python3
"""
Analyze UVW 2026 dataset statistics.

UVW 2026: Underthesea Vietnamese Wikipedia Dataset
https://github.com/undertheseanlp/underthesea/issues/896
"""

import json
from collections import Counter
from pathlib import Path

from tqdm import tqdm


DATA_DIR = Path(__file__).parent.parent / "data" / "processed"
SPLITS_DIR = Path(__file__).parent.parent / "data" / "splits"


def load_jsonl(path: Path) -> list:
    """Load data from JSONL file."""
    data = []
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            data.append(json.loads(line))
    return data


def analyze_content_length(articles: list) -> dict:
    """Analyze content length distribution."""
    lengths = [a["num_chars"] for a in articles]
    lengths.sort()

    n = len(lengths)
    return {
        "min": lengths[0],
        "max": lengths[-1],
        "mean": sum(lengths) // n,
        "median": lengths[n // 2],
        "p10": lengths[int(n * 0.1)],
        "p25": lengths[int(n * 0.25)],
        "p75": lengths[int(n * 0.75)],
        "p90": lengths[int(n * 0.9)],
    }


def analyze_titles(articles: list) -> dict:
    """Analyze article titles."""
    titles = [a["title"] for a in articles]

    # First character distribution
    first_chars = Counter(t[0].upper() for t in titles if t)

    # Title length
    title_lengths = [len(t) for t in titles]

    return {
        "total": len(titles),
        "avg_title_length": sum(title_lengths) // len(title_lengths),
        "first_char_distribution": dict(first_chars.most_common(26)),
    }


def count_vietnamese_chars(text: str) -> int:
    """Count Vietnamese-specific characters."""
    vietnamese_chars = set("àáảãạăằắẳẵặâầấẩẫậèéẻẽẹêềếểễệìíỉĩịòóỏõọôồốổỗộơờớởỡợùúủũụưừứửữựỳýỷỹỵđ")
    vietnamese_chars.update(c.upper() for c in vietnamese_chars)
    return sum(1 for c in text if c in vietnamese_chars)


def analyze_language(articles: list) -> dict:
    """Analyze Vietnamese language characteristics."""
    total_chars = 0
    vietnamese_chars = 0

    for article in tqdm(articles[:1000], desc="Analyzing language"):  # Sample for speed
        content = article["content"]
        total_chars += len(content)
        vietnamese_chars += count_vietnamese_chars(content)

    return {
        "sample_size": min(1000, len(articles)),
        "vietnamese_char_ratio": vietnamese_chars / total_chars if total_chars else 0,
    }


def main():
    """Analyze dataset statistics."""
    jsonl_path = DATA_DIR / "uvw_2026.jsonl"

    if not jsonl_path.exists():
        print(f"Dataset not found: {jsonl_path}")
        print("Please run extract_articles.py first.")
        return

    print("Loading dataset...")
    articles = load_jsonl(jsonl_path)
    print(f"Total articles: {len(articles)}")

    print("\n" + "=" * 50)
    print("CONTENT LENGTH ANALYSIS")
    print("=" * 50)
    length_stats = analyze_content_length(articles)
    for key, value in length_stats.items():
        print(f"  {key}: {value:,} chars")

    print("\n" + "=" * 50)
    print("TITLE ANALYSIS")
    print("=" * 50)
    title_stats = analyze_titles(articles)
    print(f"  Total titles: {title_stats['total']:,}")
    print(f"  Avg title length: {title_stats['avg_title_length']} chars")
    print(f"  First character distribution (top 10):")
    for char, count in list(title_stats["first_char_distribution"].items())[:10]:
        print(f"    {char}: {count:,}")

    print("\n" + "=" * 50)
    print("LANGUAGE ANALYSIS")
    print("=" * 50)
    lang_stats = analyze_language(articles)
    print(f"  Sample size: {lang_stats['sample_size']}")
    print(f"  Vietnamese char ratio: {lang_stats['vietnamese_char_ratio']:.2%}")

    print("\n" + "=" * 50)
    print("OVERALL STATISTICS")
    print("=" * 50)
    total_chars = sum(a["num_chars"] for a in articles)
    total_sentences = sum(a["num_sentences"] for a in articles)
    print(f"  Total articles: {len(articles):,}")
    print(f"  Total characters: {total_chars:,}")
    print(f"  Total sentences: {total_sentences:,}")
    print(f"  Avg chars/article: {total_chars // len(articles):,}")
    print(f"  Avg sentences/article: {total_sentences // len(articles)}")

    # Size categories
    size_categories = {
        "small (<1K chars)": 0,
        "medium (1K-10K chars)": 0,
        "large (10K-100K chars)": 0,
        "very large (>100K chars)": 0,
    }
    for a in articles:
        chars = a["num_chars"]
        if chars < 1000:
            size_categories["small (<1K chars)"] += 1
        elif chars < 10000:
            size_categories["medium (1K-10K chars)"] += 1
        elif chars < 100000:
            size_categories["large (10K-100K chars)"] += 1
        else:
            size_categories["very large (>100K chars)"] += 1

    print("\n  Article size distribution:")
    for category, count in size_categories.items():
        pct = count / len(articles) * 100
        print(f"    {category}: {count:,} ({pct:.1f}%)")

    # Save analysis
    analysis = {
        "total_articles": len(articles),
        "total_characters": total_chars,
        "total_sentences": total_sentences,
        "content_length": length_stats,
        "title_stats": title_stats,
        "language_stats": lang_stats,
        "size_distribution": size_categories,
    }

    analysis_path = DATA_DIR / "analysis.json"
    with open(analysis_path, "w", encoding="utf-8") as f:
        json.dump(analysis, f, ensure_ascii=False, indent=2)
    print(f"\nAnalysis saved to: {analysis_path}")


if __name__ == "__main__":
    main()