kilt_tasks.py

# coding=utf-8
# Copyright 2020 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Lint as: python3
"""KILT tasks training and evaluation data"""

from __future__ import absolute_import, division, print_function

import json
import logging

import datasets


_CITATION = """\
@inproceedings{fb_kilt,
    author    = {Fabio Petroni and
                 Aleksandra Piktus and
                 Angela Fan and
                 Patrick Lewis and
                 Majid Yazdani and
                 Nicola De Cao and
                 James Thorne and
                 Yacine Jernite and
                 Vassilis Plachouras and
                 Tim Rockt\"aschel and
                 Sebastian Riedel},
    title     = {{KILT:} a {B}enchmark for {K}nowledge {I}ntensive {L}anguage {T}asks},
    journal   = {CoRR},
    archivePrefix = {arXiv},
    year      = {2020},
"""

_DESCRIPTION = """\
KILT tasks training and evaluation data.
- [FEVER](https://fever.ai) | Fact Checking | fever
- [AIDA CoNLL-YAGO](https://www.mpi-inf.mpg.de/departments/databases-and-information-systems/research/ambiverse-nlu/aida/downloads) | Entity Linking | aidayago2
- [WNED-WIKI](https://github.com/U-Alberta/wned) | Entity Linking | wned
- [WNED-CWEB](https://github.com/U-Alberta/wned) | Entity Linking | cweb
- [T-REx](https://hadyelsahar.github.io/t-rex) | Slot Filling | trex
- [Zero-Shot RE](http://nlp.cs.washington.edu/zeroshot) | Slot Filling | structured_zeroshot
- [Natural Questions](https://ai.google.com/research/NaturalQuestions) | Open Domain QA  | nq
- [HotpotQA](https://hotpotqa.github.io) | Open Domain QA | hotpotqa
- [TriviaQA](http://nlp.cs.washington.edu/triviaqa) | Open Domain QA | triviaqa
- [ELI5](https://facebookresearch.github.io/ELI5/explore.html) | Open Domain QA | eli5
- [Wizard of Wikipedia](https://parl.ai/projects/wizard_of_wikipedia) | Dialogue | wow

To finish linking TriviaQA questions to the IDs provided, follow the instructions [here](http://github.com/huggingface/datasets/datasets/kilt_tasks/README.md).
"""


_DATA_URLS = {
    "fever": {
        "train": "http://dl.fbaipublicfiles.com/KILT/fever-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/fever-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/fever-test_without_answers-kilt.jsonl",
    },
    "aidayago2": {
        "train": "http://dl.fbaipublicfiles.com/KILT/aidayago2-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/aidayago2-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/aidayago2-test_without_answers-kilt.jsonl",
    },
    "wned": {
        "validation": "http://dl.fbaipublicfiles.com/KILT/wned-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/wned-test_without_answers-kilt.jsonl",
    },
    "cweb": {
        "validation": "http://dl.fbaipublicfiles.com/KILT/cweb-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/cweb-test_without_answers-kilt.jsonl",
    },
    "trex": {
        "train": "http://dl.fbaipublicfiles.com/KILT/trex-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/trex-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/trex-test_without_answers-kilt.jsonl",
    },
    "structured_zeroshot": {
        "train": "http://dl.fbaipublicfiles.com/KILT/structured_zeroshot-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/structured_zeroshot-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/structured_zeroshot-test_without_answers-kilt.jsonl",
    },
    "nq": {
        "train": "http://dl.fbaipublicfiles.com/KILT/nq-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/nq-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/nq-test_without_answers-kilt.jsonl",
    },
    "hotpotqa": {
        "train": "http://dl.fbaipublicfiles.com/KILT/hotpotqa-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/hotpotqa-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/hotpotqa-test_without_answers-kilt.jsonl",
    },
    "triviaqa": {
        "train": "http://dl.fbaipublicfiles.com/KILT/triviaqa-train_id-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/triviaqa-dev_id-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/triviaqa-test_id_without_answers-kilt.jsonl",
    },
    "eli5": {
        "train": "http://dl.fbaipublicfiles.com/KILT/eli5-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/eli5-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/eli5-test_without_answers-kilt.jsonl",
    },
    "wow": {
        "train": "http://dl.fbaipublicfiles.com/KILT/wow-train-kilt.jsonl",
        "validation": "http://dl.fbaipublicfiles.com/KILT/wow-dev-kilt.jsonl",
        "test": "http://dl.fbaipublicfiles.com/KILT/wow-test_without_answers-kilt.jsonl",
    },
}


class KILTTasksConfig(datasets.BuilderConfig):
    """BuilderConfig for KILTTasks."""

    def __init__(self, **kwargs):
        """BuilderConfig for KILTTasks.

            Args:
        .
              **kwargs: keyword arguments forwarded to super.
        """
        super(KILTTasksConfig, self).__init__(
            version=datasets.Version("1.0.0", "KILT tasks training and evaluation data"), **kwargs
        )


class KILTTasks(datasets.GeneratorBasedBuilder):
    """WikipediaKILT: Wikipedia pre-processed for KILT. Version 1.0."""

    BUILDER_CONFIGS = [
        KILTTasksConfig(
            name="all_tasks",
            description="All KILT tasks traiing and evaluation data",
        ),
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "input": datasets.Value("string"),
                    "meta": datasets.Features(
                        {
                            "left_context": datasets.Value("string"),
                            "mention": datasets.Value("string"),
                            "right_context": datasets.Value("string"),
                            "partial_evidence": datasets.features.Sequence(
                                {
                                    "start_paragraph_id": datasets.Value("int32"),
                                    "end_paragraph_id": datasets.Value("int32"),
                                    "title": datasets.Value("string"),
                                    "section": datasets.Value("string"),
                                    "wikipedia_id": datasets.Value("string"),
                                    "meta": datasets.features.Sequence(
                                        {
                                            "evidence_span": datasets.Value("string"),
                                        }
                                    ),
                                }
                            ),
                            "obj_surface": datasets.features.Sequence({"text": datasets.Value("string")}),
                            "sub_surface": datasets.features.Sequence({"text": datasets.Value("string")}),
                            "subj_aliases": datasets.features.Sequence({"text": datasets.Value("string")}),
                            "template_questions": datasets.features.Sequence({"text": datasets.Value("string")}),
                        }
                    ),
                    "output": datasets.features.Sequence(
                        {
                            "answer": datasets.Value("string"),
                            "meta": datasets.Features({"score": datasets.Value("int32")}),
                            "provenance": datasets.features.Sequence(
                                {
                                    "bleu_score": datasets.Value("float32"),
                                    "start_character": datasets.Value("int32"),
                                    "start_paragraph_id": datasets.Value("int32"),
                                    "end_character": datasets.Value("int32"),
                                    "end_paragraph_id": datasets.Value("int32"),
                                    "meta": datasets.Features(
                                        {
                                            "fever_page_id": datasets.Value("string"),
                                            "fever_sentence_id": datasets.Value("int32"),
                                            "annotation_id": datasets.Value("string"),  # int runs into overflow issues
                                            "yes_no_answer": datasets.Value("string"),
                                            "evidence_span": datasets.features.Sequence(
                                                {"text": datasets.Value("string")}
                                            ),
                                        }
                                    ),
                                    "section": datasets.Value("string"),
                                    "title": datasets.Value("string"),
                                    "wikipedia_id": datasets.Value("string"),
                                }
                            ),
                        }
                    ),
                }
            ),
            # No default supervised_keys (as we have to pass both premise
            # and hypothesis as input).
            supervised_keys=None,
            homepage="https://github.com/facebookresearch/KILT",
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        file_paths = {}
        for task_name, task_urls in _DATA_URLS.items():
            file_paths[task_name] = dl_manager.download_and_extract(task_urls)

        return [
            datasets.SplitGenerator(name=split + "_" + task, gen_kwargs={"filepath": downloaded_path})
            for task, split_paths in file_paths.items()
            for split, downloaded_path in split_paths.items()
        ]

    def _generate_examples(self, filepath):
        """Generate Wikipedia articles for KILT.

        Args:
          filepath: a string

        Yields:
          dictionaries representing article data and metadata
        """
        logging.info("generating examples from = %s", filepath)
        with open(filepath, encoding="utf-8") as f:
            for idx, line in enumerate(f):
                article = json.loads(line.strip())
                article["input"] = article.get("input", "")
                # meta
                article["meta"] = article.get("meta", {})
                for k in ["left_context", "mention", "right_context"]:
                    article["meta"][k] = article["meta"].get(k, "")
                for k in ["obj_surface", "sub_surface", "subj_aliases", "template_questions"]:
                    article["meta"][k] = {"text": article["meta"].get(k, [])}
                article["meta"]["partial_evidence"] = article["meta"].get("partial_evidence", [])
                if "partial_evidence" in article["meta"]:
                    dct_list = {}
                    for k in ["start_paragraph_id", "end_paragraph_id"]:
                        dct_list[k] = [dct.get(k, -1) for dct in article["meta"]["partial_evidence"]]
                    for k in ["title", "section", "wikipedia_id"]:
                        dct_list[k] = [dct.get(k, "") for dct in article["meta"]["partial_evidence"]]
                    if any(["meta" in dct for dct in article["meta"]["partial_evidence"]]):
                        dct_list["meta"] = [dct.get("meta", {}) for dct in article["meta"]["partial_evidence"]]
                        for meta in dct_list["meta"]:
                            meta["evidence_span"] = meta.get("evidence_span", [])
                    else:
                        dct_list["meta"] = []
                    article["meta"]["partial_evidence"] = dct_list
                # output
                article["output"] = article.get("output", [])
                dct_list = {}
                dct_list["answer"] = [dct.get("answer", "") for dct in article["output"]]
                if any(["meta" in dct for dct in article["output"]]):
                    dct_list["meta"] = [dct.get("meta", {"score": 0}) for dct in article["output"]]
                else:
                    dct_list["meta"] = []
                dct_list["provenance"] = []
                for dct in article["output"]:
                    if "provenance" in dct:
                        prov_list = dct["provenance"]
                        prov_dct_list = {}
                        prov_dct_list["bleu_score"] = [prov.get("bleu_score", 0.0) for prov in prov_list]
                        if any(["meta" in prov for prov in prov_list]):
                            prov_dct_list["meta"] = [prov.get("meta", {}) for prov in prov_list]
                            for meta_dct in prov_dct_list["meta"]:
                                meta_dct["fever_page_id"] = meta_dct.get("fever_page_id", "")
                                meta_dct["fever_sentence_id"] = meta_dct.get("fever_sentence_id", -1)
                                meta_dct["yes_no_answer"] = meta_dct.get("yes_no_answer", "")
                                meta_dct["annotation_id"] = str(meta_dct.get("annotation_id", -1))
                                meta_dct["evidence_span"] = {"text": meta_dct.get("evidence_span", [])}
                        else:
                            prov_dct_list["meta"] = []
                        for k in ["start_character", "start_paragraph_id", "end_character", "end_paragraph_id"]:
                            prov_dct_list[k] = [prov.get(k, -1) for prov in prov_list]
                        for k in ["section", "title", "wikipedia_id"]:
                            prov_dct_list[k] = [prov.get(k, "") for prov in prov_list]
                        dct_list["provenance"] += [prov_dct_list]
                article["output"] = dct_list
                yield idx, article