wikidata_rubq.py

import datasets
import os
import json
import wikidata
import pickle
from wikidata.client import Client

client = Client()

_DESCRIPTION = """\
HuggingFace wrapper for https://github.com/vladislavneon/RuBQ dataset
"""

_HOMEPAGE = "https://zenodo.org/record/4345697#.Y01k81JBy3I"


_LICENSE = "Attribution-ShareAlike 4.0 International"

_LANGS = ["ru","en"]


_URLS = {
    "test": "https://raw.githubusercontent.com/vladislavneon/RuBQ/master/RuBQ_2.0/RuBQ_2.0_test.json",
    "dev": "https://raw.githubusercontent.com/vladislavneon/RuBQ/master/RuBQ_2.0/RuBQ_2.0_dev.json",
}


_DATA_DIRECTORY = "."
VERSION = datasets.Version("0.0.1")


class WikidataRuBQConfig(datasets.BuilderConfig):
    """BuilderConfig for WikidataRuBQ."""

    def __init__(self, **kwargs):
        """BuilderConfig for WikidataRuBQ.
        Args:
          **kwargs: keyword arguments forwarded to super.
        """
        super(WikidataRuBQConfig, self).__init__(**kwargs)


class WikidataRuBQ(datasets.GeneratorBasedBuilder):
    """HuggingFace wrapper https://github.com/vladislavneon/RuBQ/tree/master/RuBQ_2.0 dataset"""

    BUILDER_CONFIG_CLASS = WikidataRuBQConfig
    BUILDER_CONFIGS = []
    BUILDER_CONFIGS += [
        WikidataRuBQConfig(
            name=f"multiple_{ln}",
            version=VERSION,
            description="questions with russian multiple labels as answers",
        )
        for ln in _LANGS
    ]

    DEFAULT_CONFIG_NAME = "multiple_en"

    def _info(self):
        features = datasets.Features(
            {
                "object": datasets.Sequence(datasets.Value("string")),
                "question": datasets.Value("string")
            }
        )

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
        )

    def _split_generators(self, dl_manager):
        if self.config.name == "default":
            version, lang = "multiple", "en"
        else:
            version, lang = self.config.name.split("_")

        if lang not in _LANGS:
            raise ValueError(f"Language {lang} not supported")

        downloaded_files = dl_manager.download_and_extract(_URLS)

        data_dir = os.path.join(self.base_path, '')
        vocab_path = os.path.join(data_dir, "reverse_vocab_wikidata_en.json")

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "filepath": downloaded_files["dev"],
                    "lang": lang,
                    "vocab_path": vocab_path,
                    "split": 'train',
                    "data_dir": data_dir
                }),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "filepath": downloaded_files["dev"],
                    "lang": lang,
                    "vocab_path": vocab_path,
                    "split": 'validation',
                    "data_dir": data_dir
                }),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "filepath": downloaded_files["test"],
                    "lang": lang,
                    "vocab_path": vocab_path,
                    "split": 'test',
                    "data_dir": data_dir
                })
        ]

    def get_name(self, idd):
        '''
        This function returns a name of an entity and its description given WikiData id
        input:  (str) wikidata id, e.x. 'Q2'
        output: (str) concatenated 'name, description' of a given entity
        '''
        entity = client.get(idd, load=True)
        name = None
        try:
            name = entity.data["labels"]["en"]["value"]
        except:
            pass
        return name

    def _generate_examples(self, filepath, lang, vocab_path, split, data_dir):
        if split == 'test':
            direct_path = os.path.join(data_dir, f"test_direct_vocab_wikidata_en.pkl")
        else:
            direct_path = os.path.join(data_dir, f"train_direct_vocab_wikidata_en.pkl")

        with open(direct_path, 'rb') as handle:
            direct_vocab = pickle.load(handle)

        with open(filepath, encoding="utf-8") as f:
            item = json.load(f)
            uid_slide = 0
            for i in item:
                question = i['question_text'] if lang == 'ru' else i['question_eng']

                objects = list(set(
                    [answer['value'].split('entity/')[1] for answer in i['answers'] if '/Q' in answer['value']]
                ))

                if len(set(objects)) >= 1:
                    if split == 'train':
                        for obj in set(objects):
                            key = i['uid'] + uid_slide
                            resolved_obj = direct_vocab.get(obj, None)
                            if resolved_obj is not None:
                                resolved_obj = resolved_obj[0].upper() + resolved_obj[1:]

                                uid_slide += 1

                                yield (
                                    key,
                                    {
                                        "object": [resolved_obj],
                                        "question": question,
                                    }
                                )
                    else:
                        key = i['uid'] + uid_slide
                        yield (
                            key,
                            {
                                "object": objects,
                                "question": question,
                            }
                        )