doc2dial.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
"""Doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset v0.9.0"""

from __future__ import absolute_import, division, print_function

import json
import logging
import os

import datasets


_CITATION = """\
@inproceedings{feng-etal-2020-doc2dial,
    title = "doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset",
    author = "Feng, Song  and Wan, Hui  and Gunasekara, Chulaka  and Patel, Siva  and Joshi, Sachindra  and Lastras, Luis",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.652",
}
"""

_DESCRIPTION = """\
Doc2dial is dataset of goal-oriented dialogues that are grounded in the associated documents. \
It includes over 4500 annotated conversations with an average of 14 turns that are grounded \
in over 450 documents from four domains. Compared to the prior document-grounded dialogue datasets \
this dataset covers a variety of dialogue scenes in information-seeking conversations.
"""

_HOMEPAGE = "https://doc2dial.github.io/file/doc2dial/"

# TODO: Add the licence for the dataset here if you can find it
_LICENSE = ""

_URLs = "https://doc2dial.github.io/file/doc2dial.zip"


class Doc2dial(datasets.GeneratorBasedBuilder):
    "Doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset v0.9"

    VERSION = datasets.Version("1.1.0")

    # You will be able to load one or the other configurations in the following list with
    # data = datasets.load_dataset("my_dataset", "first_domain")
    # data = datasets.load_dataset("my_dataset", "second_domain")
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="dialogue_domain",
            version=VERSION,
            description="This part of the dataset covers the dialgoue domain that has questions, answers and the associated doc ids",
        ),
        datasets.BuilderConfig(
            name="document_domain",
            version=VERSION,
            description="This part of the dataset covers the document domain which details all the documents in the various domains",
        ),
        datasets.BuilderConfig(
            name="doc2dial_rc",
            version=VERSION,
            description="Load Doc2Dial dataset for machine reading comprehension tasks",
        ),
    ]

    DEFAULT_CONFIG_NAME = "dialogue_domain"

    def _info(self):

        if self.config.name == "dialogue_domain":
            features = datasets.Features(
                {
                    "dial_id": datasets.Value("string"),
                    "doc_id": datasets.Value("string"),
                    "domain": datasets.Value("string"),
                    "turns": [
                        {
                            "turn_id": datasets.Value("int32"),
                            "role": datasets.Value("string"),
                            "da": datasets.Value("string"),
                            "reference": [
                                {
                                    "keys": datasets.Value("string"),
                                    "values": datasets.Value("string"),
                                }
                            ],
                            "utterance": datasets.Value("string"),
                        }
                    ],
                }
            )
        elif self.config.name == "document_domain":
            features = datasets.Features(
                {
                    "domain": datasets.Value("string"),
                    "doc_id": datasets.Value("string"),
                    "title": datasets.Value("string"),
                    "doc_text": datasets.Value("string"),
                    "spans": [
                        {
                            "id_sp": datasets.Value("string"),
                            "tag": datasets.Value("string"),
                            "start_sp": datasets.Value("int32"),
                            "end_sp": datasets.Value("int32"),
                            "text_sp": datasets.Value("string"),
                            "title": datasets.Value("string"),
                            "parent_titles": datasets.Value("string"),
                            "id_sec": datasets.Value("string"),
                            "start_sec": datasets.Value("int32"),
                            "text_sec": datasets.Value("string"),
                            "end_sec": datasets.Value("int32"),
                        }
                    ],
                    "doc_html_ts": datasets.Value("string"),
                    "doc_html_raw": datasets.Value("string"),
                }
            )
        elif self.config.name == "doc2dial_rc":
            features = datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "question": datasets.Value("string"),
                    "answers": datasets.features.Sequence(
                        {
                            "text": datasets.Value("string"),
                            "answer_start": datasets.Value("int32"),
                            "answer_end": datasets.Value("int32"),
                            "sp_id": datasets.features.Sequence(datasets.Value("string")),
                        }
                    ),
                    "is_impossible": datasets.Value("bool"),
                    "dial_context": datasets.features.Sequence(
                        {
                            "turn_id": datasets.Value("int32"),
                            "role": datasets.Value("string"),
                            "da": datasets.Value("string"),
                            "utterance": datasets.Value("string"),
                            "references": datasets.features.Sequence(
                                {
                                    "text": datasets.Value("string"),
                                    "answer_start": datasets.Value("int32"),
                                    "answer_end": datasets.Value("int32"),
                                    "sp_id": datasets.features.Sequence(datasets.Value("string")),
                                }
                            ),
                        }
                    ),
                    "doc_context": datasets.Value("string"),
                    "title": datasets.Value("string"),
                    "domain": datasets.Value("string"),
                    "start_candidates": datasets.features.Sequence(datasets.Value("int32")),
                    "end_candidates": datasets.features.Sequence(datasets.Value("int32")),
                }
            )

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            supervised_keys=None,
            homepage=_HOMEPAGE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):

        my_urls = _URLs
        data_dir = dl_manager.download_and_extract(my_urls)

        if self.config.name == "dialogue_domain":
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        "filepath": os.path.join(data_dir, "doc2dial/v0.9/data/woOOD/doc2dial_dial_train.json"),
                    },
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.VALIDATION,
                    gen_kwargs={
                        "filepath": os.path.join(data_dir, "doc2dial/v0.9/data/woOOD/doc2dial_dial_dev.json"),
                    },
                ),
            ]
        elif self.config.name == "document_domain":
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        "filepath": os.path.join(data_dir, "doc2dial/v0.9/data/doc2dial_doc.json"),
                    },
                )
            ]
        elif self.config.name == "doc2dial_rc":
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.VALIDATION,
                    gen_kwargs={
                        "filepath": os.path.join(
                            data_dir, "doc2dial", "v0.9", "data", "woOOD", "doc2dial_dial_dev.json"
                        ),
                    },
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        "filepath": os.path.join(
                            data_dir,
                            "doc2dial",
                            "v0.9",
                            "data",
                            "woOOD",
                            "doc2dial_dial_train.json",
                        ),
                    },
                ),
            ]

    def _load_doc_data_rc(self, filepath):
        doc_filepath = os.path.join(os.path.dirname(filepath), "..", "doc2dial_doc.json")
        with open(doc_filepath, encoding="utf-8") as f:
            data = json.load(f)["doc_data"]
        return data

    def _get_start_end_candidates_rc(self, spans):
        """Get the start and end positions of all the spans"""
        start_candidates, end_candidates = [], []
        for _, sp in spans.items():
            start_candidates.append(sp["start_sp"])
            end_candidates.append(sp["end_sp"])
        return start_candidates, end_candidates

    def _create_answers_merging_text_ref_rc(self, refs, spans, doc_text):
        """Combine the consecutive spans. Create answers with the start and end position of merged spans and corresponding text content in the document."""
        output = []
        if not refs:
            return output
        all_consecutive_spans = []
        consecutive_spans = []
        for id_, _ in sorted(refs.items(), key=lambda x: int(x[0])):
            if not consecutive_spans or int(id_) == int(consecutive_spans[-1]) + 1:
                consecutive_spans.append(id_)
            else:
                all_consecutive_spans.append(consecutive_spans)
                consecutive_spans = [id_]
        all_consecutive_spans.append(consecutive_spans)
        if len(all_consecutive_spans) > 1:
            all_consecutive_spans.reverse()
        for con_spans in all_consecutive_spans:
            answer = {
                "answer_start": spans[con_spans[0]]["start_sp"],
                "answer_end": spans[con_spans[-1]]["end_sp"],
                "text": doc_text[spans[con_spans[0]]["start_sp"] : spans[con_spans[-1]]["end_sp"]],
                "sp_id": con_spans,
            }
            output.append(answer)
        return output

    def _generate_examples(self, filepath):
        """This function returns the examples in the raw (text) form."""

        if self.config.name == "dialogue_domain":
            logging.info("generating examples from = %s", filepath)
            with open(filepath, encoding="utf-8") as f:
                data = json.load(f)
                for domain in data["dial_data"]:
                    for doc_id in data["dial_data"][domain]:
                        for dialogue in data["dial_data"][domain][doc_id]:

                            x = {
                                "dial_id": dialogue["dial_id"],
                                "domain": domain,
                                "doc_id": doc_id,
                                "turns": [
                                    {
                                        "turn_id": i["turn_id"],
                                        "role": i["role"],
                                        "da": i["da"],
                                        "reference": [
                                            {
                                                "keys": ref,
                                                "values": str(i["reference"][ref]),
                                            }
                                            for ref in i["reference"]
                                        ],
                                        "utterance": i["utterance"],
                                    }
                                    for i in dialogue["turns"]
                                ],
                            }

                            yield dialogue["dial_id"], x

        elif self.config.name == "document_domain":

            logging.info("generating examples from = %s", filepath)
            with open(filepath, encoding="utf-8") as f:
                data = json.load(f)
                for domain in data["doc_data"]:
                    for doc_id in data["doc_data"][domain]:
                        for dialogue in data["doc_data"][domain][doc_id]:

                            yield doc_id, {
                                "domain": domain,
                                "doc_id": doc_id,
                                "title": data["doc_data"][domain][doc_id]["title"],
                                "doc_text": data["doc_data"][domain][doc_id]["doc_text"],
                                "spans": [
                                    {
                                        "id_sp": data["doc_data"][domain][doc_id]["spans"][i]["id_sp"],
                                        "tag": data["doc_data"][domain][doc_id]["spans"][i]["tag"],
                                        "start_sp": data["doc_data"][domain][doc_id]["spans"][i]["start_sp"],
                                        "end_sp": data["doc_data"][domain][doc_id]["spans"][i]["end_sp"],
                                        "text_sp": data["doc_data"][domain][doc_id]["spans"][i]["text_sp"],
                                        "title": data["doc_data"][domain][doc_id]["spans"][i]["title"],
                                        "parent_titles": str(
                                            data["doc_data"][domain][doc_id]["spans"][i]["parent_titles"]
                                        ),
                                        "id_sec": data["doc_data"][domain][doc_id]["spans"][i]["id_sec"],
                                        "start_sec": data["doc_data"][domain][doc_id]["spans"][i]["start_sec"],
                                        "text_sec": data["doc_data"][domain][doc_id]["spans"][i]["text_sec"],
                                        "end_sec": data["doc_data"][domain][doc_id]["spans"][i]["end_sec"],
                                    }
                                    for i in data["doc_data"][domain][doc_id]["spans"]
                                ],
                                "doc_html_ts": data["doc_data"][domain][doc_id]["doc_html_ts"],
                                "doc_html_raw": data["doc_data"][domain][doc_id]["doc_html_raw"],
                            }
        elif self.config.name == "doc2dial_rc":
            """Load dialog data in the reading comprehension task setup, where context is the grounding document,
            input query is dialog history in reversed order, and output to predict is the next agent turn."""

            logging.info("generating examples from = %s", filepath)

            doc_data = self._load_doc_data_rc(filepath)
            with open(filepath, encoding="utf-8") as f:
                dial_data = json.load(f)["dial_data"]
                for domain, d_doc_dials in dial_data.items():
                    for doc_id, dials in d_doc_dials.items():
                        doc = doc_data[domain][doc_id]
                        (
                            start_pos_char_candidates,
                            end_pos_char_candidates,
                        ) = self._get_start_end_candidates_rc(doc["spans"])
                        for dial in dials:
                            all_prev_utterances = []
                            all_prev_turns = []
                            for idx, turn in enumerate(dial["turns"]):
                                all_prev_utterances.append(turn["utterance"])
                                if "references" not in turn:
                                    turn["references"] = self._create_answers_merging_text_ref_rc(
                                        turn["reference"], doc["spans"], doc["doc_text"]
                                    )
                                turn.pop("reference", None)
                                all_prev_turns.append(turn)
                                if turn["role"] == "agent":
                                    continue
                                if idx + 1 < len(dial["turns"]):
                                    if dial["turns"][idx + 1]["role"] == "agent":
                                        turn_to_predict = dial["turns"][idx + 1]
                                    else:
                                        continue
                                question = " ".join(list(reversed(all_prev_utterances)))
                                id_ = dial["dial_id"] + "_" + str(turn["turn_id"])
                                qa = {
                                    "id": id_,
                                    "question": question,
                                    "answers": [],
                                    "dial_context": all_prev_turns,
                                    "doc_context": doc["doc_text"],
                                    "title": doc_id,
                                    "domain": domain,
                                    "start_candidates": start_pos_char_candidates,
                                    "end_candidates": end_pos_char_candidates,
                                }
                                if "references" not in turn_to_predict:
                                    turn_to_predict["references"] = self._create_answers_merging_text_ref_rc(
                                        turn_to_predict["reference"], doc["spans"], doc["doc_text"]
                                    )
                                if not turn_to_predict["references"]:
                                    qa["is_impossible"] = True
                                else:
                                    qa["is_impossible"] = False
                                    qa["answers"] = turn_to_predict["references"]
                                    assert (
                                        len((qa["answers"])) >= 1
                                    ), "Ensure the answers are not empty if the question is answerable"
                                yield id_, qa