transform_wikipedia.py

import nltk
from nltk.tokenize import sent_tokenize, BlanklineTokenizer
from datasets import load_dataset
import pyarrow as pa
import pyarrow.parquet as pq
import re
from pathlib import Path


nltk.download('punkt')
nltk.download('punkt_tab')


def strip_brackets(text: str) -> str:
    """
    Remove parentheses (), braces {}, and brackets [] along with their contents.
    """
    text = re.sub(r'\([^)]*\)', '', text)
    text = re.sub(r'\{[^}]*\}', '', text)
    text = re.sub(r'\[[^\]]*\]', '', text)
    text = re.sub(r'\s+', ' ', text).strip()
    return text


def is_valid_utterance(
    utterance: str,
    minimum_utterance_length: int = 10,
) -> bool:
    utterance = utterance.strip()

    # If _any_ newlines are present after stripping the end.
    if utterance.count('\n'):
        return False

    if len(utterance) < minimum_utterance_length:
        return False

    if not utterance.endswith(('.', '!', '?')):
        if not utterance.endswith(('"', "'")):
            return False
        if len(utterance) >= 2:
            if utterance[-2] not in ('.', '!', '?'):
                return False

    if any(char in utterance for char in '(){}[]'):
        return False

    return True


def estimate_speech_duration(
    text: str,
    wpm: int = 150,
) -> float:
    """
    Estimate speech duration in seconds.
    Average speaking rate is ~150 words per minute.
    """
    words = text.split()
    return (len(words) / wpm) * 60


def combine_sentences_in_paragraph(
    sentences: list[str],
    max_duration: float = 30.0,
) -> list[str]:
    """
    Combine consecutive sentences within a paragraph if the combined duration
    falls between min_duration and max_duration seconds.
    """
    if not sentences:
        return []

    result = []
    current_chunk = []
    current_duration = 0.0

    for sentence in sentences:
        valid_utterance = is_valid_utterance(sentence)
        sentence = sentence.strip()

        # If this is the first sentence in a chunk (and it's valid)
        if not current_chunk:
            if valid_utterance:
                current_chunk.append(sentence)
                current_duration = estimate_speech_duration(sentence)

            continue

        # If the utterance is not valid, finalize the current chunk
        if not valid_utterance:
            result.append(' '.join(current_chunk))
            current_chunk = []
            continue

        sentence_duration = estimate_speech_duration(sentence)
        potential_duration = current_duration + sentence_duration

        # If adding the sentence would not exceed max_duration, add it to the current chunk
        # Otherwise, finalize the current chunk and start a new one
        if potential_duration <= max_duration:
            current_chunk.append(sentence)
            current_duration = potential_duration
            continue

        # Otherwise, finalize the current chunk and start a new one
        result.append(' '.join(current_chunk))
        current_chunk = [sentence]
        current_duration = sentence_duration

    # Finalize a remaining chunk if it exists.
    if current_chunk:
        result.append(' '.join(current_chunk))

    return result


def transform_text(text: str) -> list[str]:
    """
    Transform Wikipedia text into valid utterances.
    Returns a list of valid utterance strings.
    """
    paragraph_tokenizer = BlanklineTokenizer()
    paragraphs = paragraph_tokenizer.tokenize(text)

    valid_utterances = []

    for segment in paragraphs:
        sentences = sent_tokenize(segment)
        combined_sentences = combine_sentences_in_paragraph(sentences)

        for sentence in combined_sentences:
            if is_valid_utterance(sentence):
                valid_utterances.append(sentence)

    return valid_utterances


def process_example(example: dict) -> list[dict]:
    """
    Process a single Wikipedia example, transforming the 'text' field
    into individual utterance rows.
    """
    utterances = transform_text(example['text'])
    return [
        {
            'title': example['title'],
            'text': utterance,
            'duration': estimate_speech_duration(utterance),
        }
        for utterance in utterances
    ]


def flush_batch(rows: list[dict], output_dir: Path, shard_index: int) -> None:
    """Write a batch of rows to a Parquet shard."""
    if not rows:
        return

    table = pa.Table.from_pydict({
        'title': [r['title'] for r in rows],
        'text': [r['text'] for r in rows],
        'duration': [r['duration'] for r in rows],
    })

    shard_path = output_dir / f"shard_{shard_index:05d}.parquet"
    pq.write_table(table, shard_path)
    print(f"  Wrote shard {shard_index} with {len(rows)} utterances")


def main():
    output_dir = Path('wikipedia_utterances')
    output_dir.mkdir(exist_ok=True)

    batch_size = 100_000  # Flush to disk every 100k utterances

    print("Loading Wikipedia dataset...")
    dataset = load_dataset(
        'wikimedia/wikipedia',
        '20231101.en',
        split='train',
        streaming=True,
    )

    print("Processing dataset...")
    rows = []
    shard_index = 0
    total_utterances = 0

    for i, example in enumerate(dataset):
        utterance_rows = process_example(example)
        rows.extend(utterance_rows)

        if len(rows) >= batch_size:
            flush_batch(rows, output_dir, shard_index)
            total_utterances += len(rows)
            shard_index += 1
            rows = []

        if (i + 1) % 10_000 == 0:
            print(f"Processed {i + 1} articles, {total_utterances + len(rows)} utterances so far")

    # Flush remaining rows
    if rows:
        flush_batch(rows, output_dir, shard_index)
        total_utterances += len(rows)

    print(f"\nDone! Saved {total_utterances} utterances across {shard_index + 1} shards to '{output_dir}/'")
    print(f"Load with: load_dataset('parquet', data_files='{output_dir}/*.parquet')")


if __name__ == "__main__":
    main()