src/create_nl_dataset.py

import json
import math
import zipfile

import bs4
import datasets
import dateutil.parser
import pandas as pd
from tqdm import tqdm


def yield_file_contents(zip_path, train_df, val_df):
    with (zipfile.ZipFile(zip_path, 'r') as zip_file):
        for file_info in zip_file.infolist():
            with zip_file.open(file_info, 'r') as file:
                content = file.read()
            soup = bs4.BeautifulSoup(content, 'xml')

            id_blk = soup.find('idno', type="titelcode")
            text_id = id_blk.text.strip() if id_blk is not None else file_info.filename.replace('.xml', '')
            ti_id = '_'.join(text_id.split('_')[:-1])

            train_row = train_df[train_df['ti_id'] == ti_id]
            val_row = val_df[val_df['ti_id'] == ti_id]
            is_train = len(train_row) > 0
            is_val = len(val_row) > 0
            if is_train:
                meta = train_row.iloc[0].to_dict()
                split = 'train'
            elif is_val:
                meta = val_row.iloc[0].to_dict()
                split = 'validation'
            else:
                print(f'Did not find meta for {text_id}!')

            for key, value in list(meta.items()):
                if isinstance(value, float) and math.isnan(value):
                    meta[key] = ''

            edition_blk = soup.find('edition')
            edition = edition_blk.text.strip() if edition_blk is not None else None

            lang_blk = soup.find('language')
            language = lang_blk.get('id').strip() if lang_blk is not None else None
            
            date_blk = soup.find('revisionDesc')
            if date_blk is not None:
                date_blk = date_blk.find('date')
            if date_blk is not None:
                try:
                    date = dateutil.parser.parse(
                        date_blk.text.strip(),
                        yearfirst=True,
                        dayfirst=True
                    ).isoformat() if date_blk is not None else None
                except Exception:
                    date = None
            else:
                date = None

            meta['revision_date'] = date
            meta['edition'] = edition
            meta['language'] = language
            
            for chap_idx, chapter in enumerate(soup.find_all('div', type='chapter')):
                meta['chapter'] = chap_idx + 1
                for sec_idx, section in enumerate(chapter.find_all('div', type='section')):
                    meta['section'] = sec_idx + 1
                    text = section.text.strip()
                    yield {'meta': meta, 'text': text, 'id': f"{text_id}_{chap_idx}_{sec_idx}"}, split


if __name__ == '__main__':
    train_fraction = 0.90
    metadata_path = '../origin/titels_pd.csv'
    meta_df = pd.read_csv(metadata_path, header=1, sep='|')

    meta_df = meta_df.sample(frac=1, random_state=0)

    num_train = round(train_fraction*len(meta_df))
    train_df = meta_df.iloc[:num_train]
    val_df = meta_df.iloc[num_train:]

    with open('tmp/train.jsonl', 'w') as train_file:
        with open('tmp/val.jsonl', 'w') as val_file:
            for item, split in tqdm(yield_file_contents('../origin/xml_pd.zip', train_df, val_df)):
                if split == 'train':
                    train_file.write('{}\n'.format(json.dumps(item)))
                if split == 'validation':
                    val_file.write('{}\n'.format(json.dumps(item)))

    datasets.Dataset.from_json('tmp/train.jsonl', split='train').to_parquet('../data/train.parquet')
    datasets.Dataset.from_json('tmp/val.jsonl', split='validation').to_parquet('../data/validation.parquet')