wsdmt.py

import json
import bz2

import datasets
from datasets import DownloadManager, DatasetInfo


def _order_langs(lang1, lang2):
    return (lang1, lang2) if lang1 < lang2 else (lang2, lang1)


class WSDMTConfig(datasets.BuilderConfig):
    def __init__(self, *args, corpus, lang1, lang2, variety='base', challenge=False, **kwargs):
        lang1, lang2 = _order_langs(lang1, lang2)

        super().__init__(
            *args,
            name=f"{corpus}{'#challenge' if challenge else ''}@{lang1}-{lang2}@{variety}",
            **kwargs,
        )
        self.lang1 = lang1
        self.lang2 = lang2
        self.corpus = corpus
        self.variety = variety
        self.challenge = challenge

    def path_for(self, split, lang):
        split_path = ('challenge/' if self.challenge else '') + split
        return f"data/{self.corpus}/{self.variety}/{split_path}/{lang}.jsonl.bz2"


POS_TAGS = """ADJ
ADP
ADV
AUX
CCONJ
DET
INTJ
NOUN
NUM
PART
PRON
PROPN
PUNCT
SCONJ
SYM
VERB
X""".splitlines()


class WSDMTDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIG_CLASS = WSDMTConfig
    config: WSDMTConfig

    def _generate_examples(self, path_lang1, path_lang2):
        with bz2.open(path_lang1) as f1, bz2.open(path_lang2) as f2:
            for n, (line1, line2) in enumerate(zip(f1, f2)):
                sid1, data1 = self._read_json_line(line1)
                sid2, data2 = self._read_json_line(line2)

                assert sid1 == sid2, (
                    f"Different sentence id found for {self.config.lang1} and {self.config.lang2}: "
                    f"{sid1} != {sid2} at line {n}"
                )

                data_dict = {
                    'sid': sid1,
                    self.config.lang1: data1,
                    self.config.lang2: data2,
                }

                yield n, data_dict

    @classmethod
    def _read_json_line(cls, line):
        obj = json.loads(line)
        sid = obj.pop('sid')
        sentence = obj.pop('sentence')
        data = obj.pop('data')
        tokens, lemmas, pos_tags, senses, is_senses, is_polysemous = zip(*data)
        assert len(tokens) == len(lemmas) == len(pos_tags) == len(senses) == len(is_senses) == len(is_polysemous), (
            f"Inconsistent annotation lengths in sentence {sid}"
        )

        return sid, dict(
            sentence=sentence,
            tokens=tokens, lemmas=lemmas, pos_tags=pos_tags,
            sense=senses, identified_as_sense=is_senses, is_polysemous=is_polysemous,
        )

    def _info(self) -> DatasetInfo:
        language_features = dict(
            sentence=datasets.Value("string"),
            tokens=datasets.Sequence(datasets.Value("string")),
            sense=datasets.Sequence(datasets.Value("string")),
            identified_as_sense=datasets.Sequence(datasets.Value("bool")),
            is_polysemous=datasets.Sequence(datasets.Value("bool")),
            lemmas=datasets.Sequence(datasets.Value("string")),
            pos_tags=datasets.Sequence(datasets.ClassLabel(names=POS_TAGS)),
            # pos_tags=datasets.Sequence(datasets.Value("string")),
        )

        return datasets.DatasetInfo(
            description="empty description",
            features=datasets.Features(
                {
                    "sid": datasets.Value("string"),
                    self.config.lang1: language_features,
                    self.config.lang2: language_features
                },
            ),
            supervised_keys=None,
            homepage="no-homepage",
            citation="no-citation",
        )

    def _split_generators(self, dl_manager: DownloadManager):

        if self.config.challenge:
            split_names = ['wsd_bias', 'adversarial']
        else:
            splits_file = dl_manager.download(f'data/{self.config.corpus}/splits.txt')

            with open(splits_file) as f:
                split_names = [line.rstrip() for line in f]

        urls = {
            split: {
                self.config.lang1: self.config.path_for(split, self.config.lang1),
                self.config.lang2: self.config.path_for(split, self.config.lang2),
            }
            for split in split_names
            if not (split == 'wsd_bias' and 'adv.' in self.config.lang1)
        }
        downloaded = dl_manager.download(urls)

        return [
            datasets.SplitGenerator(name=split,
                                    gen_kwargs=dict(
                                        path_lang1=paths[self.config.lang1],
                                        path_lang2=paths[self.config.lang2],
                                    ))
            for split, paths in downloaded.items()
        ]


if __name__ == '__main__':
    from datasets import load_dataset
    load_dataset('Valahaar/wsdmt', corpus='wmt', variety='all', lang1='en', lang2='de', script_version='main')