kowiki.py

from datasets import load_dataset
import datasets
import xml.etree.ElementTree as ET
import re
import requests
from collections import defaultdict
import os

XML_FILES = [
    ['kowiki-20230420-pages-articles.xml'],
    ['kowiki-20230501-pages-articles.xml'],
    ['kowiki-20230520-pages-articles.xml'],
    ['kowiki-20230601-pages-articles.xml'],
    ['kowiki-20230620-pages-articles.xml'],
    ['kowiki-20230701-pages-articles.xml'],
    ['kowiki-20230720-pages-articles.xml'],
    ['kowiki-20230801-pages-articles.xml'],
    ['kowiki-20230820-pages-articles.xml'],
    ['kowiki-20230901-pages-articles.xml'],
    ['kowiki-20231101-pages-articles-multistream.xml']
]

PROCESSED_FILES = [
    [f"20230420/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230501/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230520/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230601/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230620/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230701/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230720/train-{i:05d}-of-00010.parquet" for i in range(10)],
    [f"20230801/train-{i:05d}-of-00010.parquet" for i in range(10)],
]



def take_behind_linked_text(text):
    _PATTERN = "\[\[[^\[\]\|]+\|([^\[\]]+)\]\]"
    result = ""
    while True:
        result = re.sub(_PATTERN, "\\1", text)
        if result == text:
            break
        text = result
    return text


def remove_linked_text(text):
    result = ""
    while True:
        result = re.sub("\[\[([^\[\]\|]+)\]\]", "\\1", text)
        if result == text:
            break
        text = result
    return text


def remove_attribute_in_table(text):
    text = re.sub("<bgcolor=#[^>]+>", "", text)
    text = re.sub("<-[0-9]>", "", text)
    text = re.sub("\|\|<table[^\n]+\n", "", text)
    text = re.sub("<tablewidth\=[^>]+>", "", text)
    text = re.sub("<width\=[^>]+>", "", text)
    text = re.sub("(?<=코멘트\-)\|\|(?=\n)", "", text)

    return text


def replace_link(text):
    text = re.sub("\[youtube\([^\]]+\)\]", "[YOUTUBE LINK]", text)
    return text


def remove_double_bracket_text(text):
    text = re.sub("\{\{([^|\{\}]+)\}\}", "\\1", text)
    text = re.sub("\|style[^\}]+(\}\})", "}}", text)

    text = re.sub("\{\{ll?ang\|[a-z]+\|([^\{\}]+)\}\}", "\\1", text)
    text = re.sub("\{\{[^\=]+\|([^\=\{\}]+)\}\}", "\\1", text)
    matched = re.search("{{([^\|\{\}]+)\n\|([^\}]+)\}\}", text)
    if matched is not None:
        s, e = matched.span()
        values = matched.group(2).split("\n|")
        text = text[:s] + f"[{matched.group(1)}]" + ", ".join(
            [v for v in values if v.split("=")[-1].strip() not in ("", "값찾기")]) + "\n" + text[e:]
    text = re.sub("\{\{보호 문서\|크기\=작게\}\}", "", text)
    text = re.sub("{{말풍선\|1\=NUMBER+[^}]+\}\}", "<NUMBER>", text)
    return text


def process_text(text: str):
    if not isinstance(text, str):
        return ""
    text = text.strip()
    text = re.sub("<[^>]+>", "", text)
    text = re.sub("\[\[(파일:[^\]]+)", "[\\1", text)

    text = remove_linked_text(text)
    text = take_behind_linked_text(text)
    text = text.replace("[파일:", "[[파일:")
    text = re.sub("'''", "", text)
    text = replace_link(text)
    text = remove_attribute_in_table(text)
    text = remove_double_bracket_text(text)

    text = re.sub("[0-9]+px\|sub\|link[^\n]+", "", text)
    text = text.replace("\n]]\n", "\n")
    text = re.sub("섬네일\|[^\n]+\|([^\n]+)", "\\1", text)
    text = re.sub("\n{2,}", "\n\n", text)
    text = re.sub("\t{2,}", "\t\t", text)
    text = re.sub("[“”]", '"', text)
    text = re.sub("[‘’]", "'", text)
    text = text.replace("《", "<<").replace("》", ">>")
    return text.strip()


class WikiConfig(datasets.BuilderConfig):
    """BuilderConfig for Wikibuilder."""

    def __init__(self, name, data_url, data_dir, citation, **kwargs):
        """BuilderConfig for Wikibuilder.

        Args:
          name: `string`, name of the dataset
          data_url: `string`, url to the dataset
          data_dir: `string`, directory to which the data is downloaded and extracted
          citation: `string`, citation for the dataset
          **kwargs: keyword arguments forwarded to super.
        """
        super(WikiConfig, self).__init__(
            name=name,
            version=datasets.Version("1.0.0", ""),
            description=("Wikipedia dataset"),
            **kwargs
        )
        self.data_url = data_url
        self.data_dir = data_dir
        self.citation = citation

class KoWiki(datasets.GeneratorBasedBuilder):
    """KoWiki: Korean Wikipedia"""

    BUILDER_CONFIGS = [
        WikiConfig(
            name=name[0].split("-")[1],
            data_url=[f"https://huggingface.co/datasets/psyche/kowiki/resolve/main/data/{n}" for n in name],
            data_dir="data",
            citation="",
        )
        for name in XML_FILES
    ]+[
        WikiConfig(
            name = os.path.dirname(name[0]) +".process",
            data_url=[f"https://huggingface.co/datasets/psyche/kowiki/resolve/main/data/{n}" for n in name],
            data_dir=name,
            citation="",
        )
        for name in PROCESSED_FILES
    ]


    def _info(self):
        return datasets.DatasetInfo(
            description=("Korean Wikipedia", ),
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "title": datasets.Value("string"),
                    "text": datasets.Value("string"),
                }
            ),
            supervised_keys=None,
            homepage="https://dumps.wikimedia.org/kowiki/",
            citation=self.config.citation,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        data_files = dl_manager.download_and_extract(self.config.data_url)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"data_files": data_files},
            )
        ]

    def _generate_examples(self, data_files):
        n = 0
        if self.config.name.endswith(".process"):
            _dataset = load_dataset("parquet", data_files={"train": data_files}, token=self.use_auth_token, split= "train", streaming=True)
            for d in _dataset:
                yield n, d
                n += 1
        else:
            output = {}
            for data_file in data_files:
                if data_file.startswith(("https://", "http://")):
                    data_file = requests.get(data_file, stream=True).raw

                for event, node in ET.iterparse(data_file):
                    tag_name = node.tag
                    if tag_name.endswith("title"):
                        output["title"] = node.text

                    elif tag_name.endswith("text"):
                        output["text"] = node.text

                    elif tag_name.endswith("page"):
                        output["id"] = f"{n:08d}"
                        yield n, output
                        output = {}
                        n += 1