import re from typing import List import unicodedata from pathlib import Path from bs4 import BeautifulSoup import datasets as ds _DESCRIPTION = "Parallel passages from novels." _CITATION = """ 内山将夫,高橋真弓.(2003) 日英対訳文対応付けデータ. Masao Utiyama and Mayumi Takahashi. (2003) English-Japanese Translation Alignment Data. """.strip() _HOMEPAGE = "https://www2.nict.go.jp/astrec-att/member/mutiyama/align/" _LICENSE = None _DOWNLOAD_URL = ( "https://www2.nict.go.jp/astrec-att/member/mutiyama/align/download/align-070215.zip" ) def preprocess(text: str): text = re.sub(r"<注[0-9]+>", "", text.strip()) text = re.sub(r"《.*?》", "", text) text = re.sub(r"[#.*?]", "", text) text = re.sub(r"([\u3040-\u309F]+)", "", text) text = re.sub(r" − (.+) − ", "――\\1――", text) text = re.sub(r"_(.+)_", "\\1", text) text = re.sub(r" ``$", "''", text.strip()) text = re.sub(r"^――", "", text.strip()) text = re.sub(r"^..第", "第", text.strip()) return text.strip() def parse_html_table(path: Path): try: with path.open(encoding="shift_jis") as f: content = f.read() except UnicodeDecodeError: try: with path.open(encoding="utf-8") as f: content = f.read() except UnicodeDecodeError: try: with path.open(encoding="cp932") as f: content = f.read() except UnicodeDecodeError: return [], [] soup = BeautifulSoup(content, "lxml") tables = soup.find_all("table") texts_en, texts_ja = [], [] cur_text_en, cur_text_ja = "", "" cur_left_parens, cur_right_parens = 0, 0 cur_left_quote, cur_right_quote = 0, 0 cur_left_parens_ja, cur_right_parens_ja = 0, 0 cur_left_parens_ja2, cur_right_parens_ja2 = 0, 0 for table in tables: for tr in table.find_all("tr"): text_en, _, text_ja = (preprocess(td.text) for td in tr.find_all("td")) text_ja = unicodedata.normalize("NFKC", text_ja) cur_left_parens += text_en.count("(") cur_right_parens += text_en.count(")") cur_left_quote += len(list(re.findall(r"``", text_en))) cur_right_quote += len(list(re.findall(r"''", text_en))) # cur_right_quote += max( # len(list(re.findall(r"''", text_en))) # - len(list(re.findall(r"'''", text_en))), # 0, # ) cur_left_parens_ja += text_ja.count("「") cur_right_parens_ja += text_ja.count("」") cur_left_parens_ja2 += text_ja.count("『") cur_right_parens_ja2 += text_ja.count("』") if ( text_ja.strip().endswith("。") and text_en.strip().endswith(".") and cur_left_parens == cur_right_parens and cur_left_quote == cur_right_quote and cur_left_parens_ja == cur_right_parens_ja and cur_left_parens_ja2 == cur_right_parens_ja2 ): texts_en.append((cur_text_en + " " + text_en).strip()) texts_ja.append((cur_text_ja + text_ja).strip()) cur_text_en, cur_text_ja = "", "" cur_left_parens, cur_right_parens = 0, 0 cur_left_quote, cur_right_quote = 0, 0 cur_left_parens_ja, cur_right_parens_ja = 0, 0 cur_left_parens_ja2, cur_right_parens_ja2 = 0, 0 else: cur_text_en += " " + text_en cur_text_ja += text_ja texts_en.append(cur_text_en.strip()) texts_ja.append(cur_text_ja.strip()) return texts_en, texts_ja class EnJaAlignDataset(ds.GeneratorBasedBuilder): VERSION = ds.Version("1.0.0") DEFAULT_CONFIG_NAME = "default" BUILDER_CONFIGS = [ ds.BuilderConfig( name="default", version=VERSION, description="", ), ] def _info(self) -> ds.DatasetInfo: if self.config.name == "default": features = ds.Features( { "id": ds.Value("string"), "en": ds.Value("string"), "ja": ds.Value("string"), "source": ds.Value("string"), } ) return ds.DatasetInfo( description=_DESCRIPTION, citation=_CITATION, homepage=_HOMEPAGE, license=_LICENSE, features=features, ) def _split_generators(self, dl_manager: ds.DownloadManager): data_path = dl_manager.download_and_extract(_DOWNLOAD_URL) paths = list(Path(data_path, "align/htmPages").glob("*.htm")) return [ ds.SplitGenerator( name=ds.Split.TRAIN, gen_kwargs={"paths": paths}, ) ] def _preprocess_ja(self, text: str) -> str: text = re.sub(r"\d+\.(\d|\.)*", "", text.strip()).strip() text = re.sub(r"^――", "", text).strip() return text def _preprocess_en(self, text: str) -> str: text = re.sub(r"\d+\.(\d|\.)*", "", text.strip()).strip() text = re.sub(r"```(.*?)'", "``\1", text).strip() text = re.sub(r"``(.*?)''", r'"\1"', text).strip() return text def _generate_examples(self, paths: List[Path]): for path in paths: idx = 0 texts_en, texts_ja = parse_html_table(path) for text_en, text_ja in zip(texts_en, texts_ja, strict=True): row = { "id": f"{path.stem}/{idx}", "en": self._preprocess_en(text_en), "ja": self._preprocess_ja(text_ja), "source": path.name, } if ( isinstance(row["en"], str) and isinstance(row["ja"], str) and len(row["en"]) > 0 and len(row["ja"]) > 0 ): yield f"{path.name}/{idx}", row idx += 1