lsoie.py

# -*- coding: utf-8 -*-
"""LSOIE: A Large-Scale Dataset for Supervised Open Information Extraction."""
import os
import datasets
from datasets.info import SupervisedKeysData
from zipfile import ZipFile
logger = datasets.logging.get_logger(__name__)


_CITATION = """\
@article{lsoie-2021,
  title={{LSOIE}: A Large-Scale Dataset for Supervised Open Information Extraction},
  author={{Solawetz}, Jacob and {Larson}, Stefan},
  journal={arXiv preprint arXiv:2101.11177},
  year={2019},
  url="https://arxiv.org/pdf/2101.11177.pdf"
}
"""

_DESCRIPTION = """
The Large Scale Open Information Extraction Dataset (LSOIE), is a dataset 20 
times larger than the next largest human-annotated Open Information Extraction
(OIE) dataset. LSOIE is a built upon the QA-SRL 2.0 dataset.
"""

_URL = "https://github.com/Jacobsolawetz/large-scale-oie/"
_URLS = {
    "zip": _URL+"raw/master/dataset_creation/lsoie_data/lsoie_data.zip"
}
_ARCHIVE_FILES = [
    "lsoie_science_train.conll",
    "lsoie_science_dev.conll",
    "lsoie_science_test.conll",
    "lsoie_wiki_train.conll",
    "lsoie_wiki_dev.conll",
    "lsoie_wiki_test.conll",
    ]


class LsoieConfig(datasets.BuilderConfig):
    """BuilderConfig for LSOIE."""

    def __init__(self,subset="wiki", **kwargs):
        """BuilderConfig for LSOIE.
        Args:
          subset: str - either "wiki" or "science"
          **kwargs: keyword arguments forwarded to super.
        """
        super(LsoieConfig, self).__init__(**kwargs)
        self.subset=subset


class Lsoie(datasets.GeneratorBasedBuilder):
    """LSOIE: A Large-Scale Dataset for Supervised Open Information Extraction"""

    BUILDER_CONFIGS = [
        LsoieConfig(
            name="wiki",
            description="LSOIE dataset from wikipedia and wikinews",
            subset="wiki",
        ),
        LsoieConfig(
            name="sci",
            description="LSOIE dataset build over scientific domain",
            subset="science",
        ),
    ]
    
    DEFAULT_CONFIG_NAME = "wiki"

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "word_ids": datasets.Sequence(datasets.Value("int16")),
                    "words": datasets.Sequence(datasets.Value("string")),
                    "pred": datasets.Value("string"),
                    "pred_ids": datasets.Sequence(datasets.Value("int16")),
                    "head_pred_id": datasets.Value("int16"),
                    "sent_id": datasets.Value("int16"),
                    "run_id": datasets.Value("int16"),
                    "label": datasets.Sequence(datasets.Value("string")),
                }
            ),
            supervised_keys=SupervisedKeysData(input="word_ids",output="label"),
            homepage=_URL,
            citation=_CITATION,
            #there is no default task for open information extraction yet
            #task_templates=[
            #    OpenInformationExtraction(
            #        question_column="question", context_column="context", answers_column="answers"
            #    )
            #],
        )

    def _split_generators(self, dl_manager):
        downloaded_archive = dl_manager.download(_URLS)['zip']          
        #name_pre=os.path.join("lsoie_data","lsoie_")+self.config.subset+"_"
        name_pre="lsoie_"+self.config.subset+"_"
        return [
            datasets.SplitGenerator(name=datasets.Split.TRAIN, 
                                    gen_kwargs={
                                        "archive_path": downloaded_archive,
                                        "file_name": name_pre+"train.conll",
                                        }),
            datasets.SplitGenerator(name=datasets.Split.VALIDATION, 
                                    gen_kwargs={
                                        "archive_path": downloaded_archive,
                                        "file_name": name_pre+"dev.conll",
                                        }),
            datasets.SplitGenerator(name=datasets.Split.TEST, 
                                    gen_kwargs={
                                        "archive_path": downloaded_archive,
                                        "file_name": name_pre+"test.conll",
                                        }),
        ]
                        
    def _generate_examples(self,archive_path,file_name):
        """This functions returns the samples in a raw format"""
        logger.info("generating examples from archive:{}".format(archive_path))
        columns={'word_ids':int,
                 'words':str, 
                 'pred':str,
                 'pred_ids':lambda x: [ num for num in x.strip('[]').split(',')],
                 'head_pred_id': int,
                 'sent_id':int,
                 'run_id': int,
                 'label':str}
        list_columns=["word_ids","words","label"]
        sep="\t"
        key=0
        sentence=dict()
        for column in list_columns:
            sentence[column]=[]
        with ZipFile(archive_path) as zipfile:
            with zipfile.open('lsoie_data/'+file_name,mode='r') as file:
                for line in file:
                    line=line.decode("utf-8").strip('\n').split(sep=sep)
                    if line[0]=='':
                        yield key, sentence
                        key+=1
                        for column in list_columns:
                            sentence[column]=[]
                        continue
                    for column, val in zip(columns.keys(),line):
                        val=columns[column](val)
                        if column in list_columns:
                            sentence[column].append(val)
                        else:
                            sentence[column]=val