ERRnews.py

import os

import datasets
import pandas as pd
import regex as re


class ERRNewsConfig(datasets.BuilderConfig):
    def __init__(self, data_url, features, recordings_url, **kwargs):
        super().__init__(version=datasets.Version("1.1.0"), **kwargs)
        self.data_url = data_url
        self.recordings_url = recordings_url
        self.features = features


class ERRNews(datasets.GeneratorBasedBuilder):
    data_url = "https://cs.taltech.ee/staff/heharm/ERRnews/data.zip"
    recordings_url = "https://cs.taltech.ee/staff/heharm/ERRnews/recordings.tar"
    features = ["name", "summary", "transcript", "url", "meta"]

    BUILDER_CONFIGS = [
        ERRNewsConfig(
            name="et",
            data_url=data_url,
            recordings_url=None,
            features=features
        ),
        ERRNewsConfig(
            name="audio",
            data_url=data_url,
            recordings_url=recordings_url,
            features=features + ["audio", "recording_id"]
        ),
        ERRNewsConfig(
            name="et_en",
            data_url=data_url,
            recordings_url=None,
            features=features + ["en_summary", "en_transcript"]
        ),
        ERRNewsConfig(
            name="full",
            data_url=data_url,
            recordings_url=recordings_url,
            features=features + ["audio", "recording_id", "en_summary", "en_transcript"]
        )
    ]

    DEFAULT_CONFIG_NAME = "et"

    def _info(self):
        description = (
            "ERRnews is an estonian language summaryzation dataset of ERR News broadcasts scraped from the ERR "
            "Archive (https://arhiiv.err.ee/err-audioarhiiv). The dataset consists of news story transcripts "
            "generated by an ASR pipeline paired with the human written summary from the archive. For leveraging "
            "larger english models the dataset includes machine translated (https://neurotolge.ee/) transcript and "
            "summary pairs."
        )

        citation = """\
        @article{henryabstractive,
          title={Abstractive Summarization of Broadcast News Stories for {Estonian}},
          author={Henry, H{\"a}rm and Tanel, Alum{\"a}e},
          journal={Baltic J. Modern Computing},
          volume={10},
          number={3},
          pages={511-524},
          year={2022}
        }
        """

        features = datasets.Features(
            {
                "name": datasets.Value("string"),
                "summary": datasets.Value("string"),
                "transcript": datasets.Value("string"),
                "url": datasets.Value("string"),
                "meta": datasets.Value("string"),
            })

        if self.config.name == "audio":
            features["audio"] = datasets.features.Audio(sampling_rate=16_000)
            features["recording_id"] = datasets.Value("int32")

        if self.config.name == "et_en":
            features["en_summary"] = datasets.Value("string")
            features["en_transcript"] = datasets.Value("string")

        if self.config.name == "full":
            features["en_summary"] = datasets.Value("string")
            features["en_transcript"] = datasets.Value("string")
            features["audio"] = datasets.features.Audio(sampling_rate=16_000)
            features["recording_id"] = datasets.Value("int32")

        return datasets.DatasetInfo(
            description=description,
            citation=citation,
            features=features,
            supervised_keys=None,
            version=self.config.version,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        train = "data/train.csv"
        test = "data/test.csv"
        val = "data/val.csv"

        data_archive = dl_manager.download_and_extract(self.config.data_url)

        if self.config.recordings_url:
            recordings = dl_manager.download(self.config.recordings_url)
            recordings_archive = dl_manager.extract(recordings) if not dl_manager.is_streaming else None
            audio_files = dl_manager.iter_archive(recordings)
        else:
            audio_files = None
            recordings_archive = None

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "file_path": train,
                    "audio_files": audio_files,
                    "recordings_archive": recordings_archive,
                    "data_archive": data_archive
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "file_path": val,
                    "audio_files": audio_files,
                    "recordings_archive": recordings_archive,
                    "data_archive": data_archive
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "file_path": test,
                    "audio_files": audio_files,
                    "recordings_archive": recordings_archive,
                    "data_archive": data_archive
                },
            ),
        ]

    def create_dict(self, data):
        res = dict()
        for key in self.config.features:
            res[key] = data[key]
        return res

    def _generate_examples(self, file_path, audio_files, recordings_archive, data_archive):
        data = pd.read_csv(os.path.join(data_archive, file_path))

        if audio_files:
            for path, f in audio_files:
                id = re.sub("^recordings\/", "", re.sub(".ogv$", "", path))
                row = data.loc[data['recording_id'] == int(id)]
                if len(row) > 0:
                    result = row.to_dict('records')[0]
                    # set the audio feature and the path to the extracted file
                    path = os.path.join(recordings_archive, path) if recordings_archive else path
                    result["audio"] = {"path": path, "bytes": f.read()}
                    yield row.index[0].item(), self.create_dict(result)
        else:
            for row in data.iterrows():
                result = row[1].to_dict()
                yield row[0], self.create_dict(result)