bucc-bitext-mining.py

# coding=utf-8

"""BUCC"""

import json
import datasets
import os
import gzip

logger = datasets.logging.get_logger(__name__)

_DESCRIPTION = """\
        Tatoeba multilingual test set
"""

_LANGUAGES = [
  'de-en',
  'fr-en',
  'ru-en',
  'zh-en'
]

class BUCC(datasets.GeneratorBasedBuilder):
    """BUCC Cross-lingual"""

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name=name,
            version=datasets.Version("1.0.0"),
            description=f"BUCC cross-lingual dataset for {name} language pair.",
        )
        for name in _LANGUAGES
    ]

    DEFAULT_CONFIG_NAME = 'fr-en'

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "sentence1": datasets.Sequence(datasets.Value("string")),
                    "sentence2": datasets.Sequence(datasets.Value("string")),
                    "gold": datasets.Sequence([datasets.Value("int32")])
                },
            ),
            supervised_keys=None,
            homepage="https://comparable.limsi.fr/bucc2018/bucc2018-task.html",
        )

    def _split_generators(self, dl_manager):
        url = f'{self.config.name}/test.json.gz'
        archive_path = dl_manager.download(url)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "text_path": archive_path,
                },
            ),
        ]

    def _generate_examples(self, text_path):
        """Yields examples."""
        with gzip.open(text_path,'rt') as f:
            for i, line in enumerate(f):
                yield i, json.loads(line)