dureader.py

import json
import os

import datasets
from datasets import DatasetInfo, DownloadManager


class DuReaderConfig(datasets.BuilderConfig):
    """Config for DuReader dataset"""

    def __init__(self, name, data_url, **kwargs):
        super().__init__(name=name, version=datasets.Version("1.0.0", ""))
        self.data_url = data_url


class DuReader(datasets.GeneratorBasedBuilder):
    """ """

    BUILDER_CONFIGS = [
        DuReaderConfig(
            name="robust",
            data_url="https://dataset-bj.cdn.bcebos.com/qianyan/dureader_robust-data.tar.gz",
        ),
        DuReaderConfig(
            name="checklist",
            data_url="https://dataset-bj.cdn.bcebos.com/qianyan/dureader_checklist-data.tar.gz",
        ),
        # DuReaderConfig(
        #     name="yesno",
        #     data_url="https://dataset-bj.cdn.bcebos.com/qianyan/dureader_yesno-data.tar.gz",
        # ),
    ]

    def _info(self) -> DatasetInfo:
        if self.config.name == "robust":
            features = {
                "id": datasets.Value("string"),
                "context": datasets.Value("string"),
                "question": datasets.Value("string"),
                "answers": datasets.Sequence(
                    {
                        "text": datasets.Value("string"),
                        "answer_start": datasets.Value("int32"),
                    }
                ),
            }
            return datasets.DatasetInfo(
                description="",
                citation="",
                homepage="",
                features=datasets.Features(features),
            )
        if self.config.name == "checklist":
            features = {
                "id": datasets.Value("string"),
                "title": datasets.Value("string"),
                "context": datasets.Value("string"),
                "question": datasets.Value("string"),
                "is_impossible": datasets.Value("string"),
                "answers": datasets.Sequence(
                    {
                        "text": datasets.Value("string"),
                        "answer_start": datasets.Value("int32"),
                    }
                ),
                "type": datasets.Value("string"),
            }
            return datasets.DatasetInfo(
                description="",
                citation="",
                homepage="",
                features=datasets.Features(features),
            )

        return None

    def _split_generators(self, dl_manager: DownloadManager):
        """Split generators"""

        def _build(train_files, valid_files, test_files):
            train_split = datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "data_file": train_files,
                    "split": "train",
                },
            )
            valid_split = datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "data_file": valid_files,
                    "split": "dev",
                },
            )
            test_split = datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "data_file": test_files,
                    "split": "test",
                },
            )
            return [train_split, valid_split, test_split]

        if self.config.name == "robust":
            dl_dir = dl_manager.download_and_extract(self.config.data_url)
            splits = _build(
                train_files=os.path.join(dl_dir, "dureader_robust-data", "train.json"),
                valid_files=os.path.join(dl_dir, "dureader_robust-data", "dev.json"),
                test_files=os.path.join(dl_dir, "dureader_robust-data", "test.json"),
            )
            return splits

        if self.config.name == "checklist":
            dl_dir = dl_manager.download_and_extract(self.config.data_url)
            splits = _build(
                train_files=os.path.join(dl_dir, "dureader_checklist-data", "train.json"),
                valid_files=os.path.join(dl_dir, "dureader_checklist-data", "dev.json"),
                test_files=os.path.join(dl_dir, "dureader_checklist-data", "test.json"),
            )
            return splits

        return []

    def _generate_examples(self, data_file, split):
        if self.config.name == "robust":
            if split == "train" or split == "dev":
                return self._generate_robust_examples(data_file)
            return self._generate_robust_test_examples(data_file)

        if self.config.name == "checklist":
            if split == "train" or split == "dev":
                return self._generate_checklist_examples(data_file)
            return self._generate_checklist_test_examples(data_file)

    def _generate_robust_examples(self, data_file):
        with open(data_file, mode="rt", encoding="utf-8") as fin:
            data = json.load(fin)["data"]
        for d in data:
            for p in d["paragraphs"]:
                context = p["context"]
                for qa in p["qas"]:
                    starts = [x["answer_start"] for x in qa["answers"]]
                    answers = [x["text"] for x in qa["answers"]]
                    example = {
                        "id": qa["id"],
                        "context": context,
                        "question": qa["question"],
                        "answers": {
                            "text": answers,
                            "answer_start": starts,
                        },
                    }
                    yield example["id"], example

    def _generate_robust_test_examples(self, data_file):
        with open(data_file, mode="rt", encoding="utf-8") as fin:
            data = json.load(fin)["data"]
        for d in data:
            for p in d["paragraphs"]:
                context = p["context"]
                for qa in p["qas"]:
                    qid = qa["id"]
                    example = {
                        "id": qid,
                        "context": context,
                        "question": qa["question"],
                        "answers": {
                            "text": [],
                            "answer_start": [],
                        },
                    }
                    yield example["id"], example

    def _generate_checklist_examples(self, data_file):
        with open(data_file, mode="rt", encoding="utf-8") as fin:
            data = json.load(fin)["data"]
        exist_ids = set()
        for d in data:
            for p in d["paragraphs"]:
                title = p["title"].strip()
                context = p["context"].strip()
                for qa in p["qas"]:
                    qid = qa["id"]
                    # skip dumplicate keys
                    if qid in exist_ids:
                        continue
                    exist_ids.add(qid)
                    starts = [x["answer_start"] for x in qa["answers"]]
                    answers = [x["text"].strip() for x in qa["answers"]]
                    example = {
                        "id": qid,
                        "title": title,
                        "context": context,
                        "question": qa["question"].strip(),
                        "is_impossible": qa["is_impossible"],
                        "answers": {
                            "text": answers,
                            "answer_start": starts,
                        },
                        "type": qa["type"].strip(),
                    }
                    yield example["id"], example

    def _generate_checklist_test_examples(self, data_file):
        with open(data_file, mode="rt", encoding="utf-8") as fin:
            data = json.load(fin)["data"]
        exist_ids = set()
        for d in data:
            for p in d["paragraphs"]:
                title = p["title"]
                context = p["context"]
                for qa in p["qas"]:
                    qid = qa["id"]
                    if qid in exist_ids:
                        continue
                    exist_ids.add(qid)
                    example = {
                        "id": qid,
                        "title": title,
                        "context": context,
                        "question": qa["question"],
                        "is_impossible": None,
                        "answers": {
                            "text": [],
                            "answer_start": [],
                        },
                        "type": None,
                    }
                    yield example["id"], example