mmarco-train.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
"""mMARCO dataset."""

from collections import defaultdict
from gc import collect
import datasets
from tqdm import tqdm
import random


_CITATION = """
@misc{bonifacio2021mmarco,
      title={mMARCO: A Multilingual Version of the MS MARCO Passage Ranking Dataset},
      author={Luiz Henrique Bonifacio and Israel Campiotti and Vitor Jeronymo and Hugo Queiroz Abonizio and Roberto Lotufo and Rodrigo Nogueira},
      year={2021},
      eprint={2108.13897},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
"""

_URL = "https://github.com/unicamp-dl/mMARCO"

_DESCRIPTION = """
mMARCO translated datasets
"""


_BASE_URLS = {
    "collections": "https://huggingface.co/datasets/unicamp-dl/mmarco/resolve/main/data/google/collections/",
    "queries-train": "https://huggingface.co/datasets/unicamp-dl/mmarco/resolve/main/data/google/queries/train/",
    "queries-dev": "https://huggingface.co/datasets/unicamp-dl/mmarco/resolve/main/data/google/queries/dev/",
    "runs": "https://huggingface.co/datasets/unicamp-dl/mmarco/resolve/main/data/google/runs/",
    "train": "https://huggingface.co/datasets/unicamp-dl/mmarco/resolve/main/data/triples.train.ids.small.tsv",
}

LANGUAGES = [
    "arabic",
    "chinese",
    "dutch",
    "english",
    "french",
    "german",
    "hindi",
    "indonesian",
    "italian",
    "japanese",
    "portuguese",
    "russian",
    "spanish",
    "vietnamese",
]


class MMarco(datasets.GeneratorBasedBuilder):

    BUILDER_CONFIGS = (
        [
            datasets.BuilderConfig(
                name=language,
                description=f"{language.capitalize()} triples",
                version=datasets.Version("2.0.0"),
            )
            for language in LANGUAGES
        ]
        + [
            datasets.BuilderConfig(
                name=f"collection-{language}",
                description=f"{language.capitalize()} collection version v2",
                version=datasets.Version("2.0.0"),
            )
            for language in LANGUAGES
        ]
        + [
            datasets.BuilderConfig(
                name=f"queries-{language}",
                description=f"{language.capitalize()} queries version v2",
                version=datasets.Version("2.0.0"),
            )
            for language in LANGUAGES
        ]
        + [
            datasets.BuilderConfig(
                name=f"runs-{language}",
                description=f"{language.capitalize()} runs version v2",
                version=datasets.Version("2.0.0"),
            )
            for language in LANGUAGES
        ]
        + [
            datasets.BuilderConfig(
                name=f"all",
                description=f"All training data version v2",
                version=datasets.Version("2.0.0"),
            )
        ]
    )
    
    size_per_lang = {lang: 398792 for lang in LANGUAGES}
    # $ cat triples.train.ids.small.tsv  | cut -f 1  | sort | uniq | wc -l 
    # 398792

    DEFAULT_CONFIG_NAME = "english"

    def _info(self):
        name = self.config.name
        assert name in LANGUAGES + ["all"], f"Does not support languge {name}. Must be one of {LANGUAGES}."

        features = {
            "query_id": datasets.Value("string"),
            "query": datasets.Value("string"),
            "positive_passages": [
                {'docid': datasets.Value('string'), 'text': datasets.Value('string')}
            ],
            "negative_passages": [
                {'docid': datasets.Value('string'), 'text': datasets.Value('string')}
            ],
        }

        return datasets.DatasetInfo(
            description=f"{_DESCRIPTION}\n{self.config.description}",
            features=datasets.Features(features),
            supervised_keys=None,
            homepage=_URL,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        languages = [self.config.name] if self.config.name in LANGUAGES else LANGUAGES
        urls = {
            "collection": {lang: _BASE_URLS["collections"] + lang + "_collection.tsv" for lang in languages},
            "queries": {lang: _BASE_URLS["queries-train"] + lang + "_queries.train.tsv" for lang in languages}, 
            "train": _BASE_URLS["train"],
        }
        dl_path = dl_manager.download_and_extract(urls)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "files": dl_path["train"],
                    "args": {
                        "collection": dl_path["collection"],
                        "queries": dl_path["queries"],
                    },
                },
            )
        ]


    def _generate_examples(self, files, args=None):
        """Yields examples."""

        languages = [self.config.name] if self.config.name in LANGUAGES else LANGUAGES

        # loading
        runs = dict() # each query: [set(pos_passages), set(neg_passages)]
        with open(files, encoding="utf-8") as f:
            for (idx, line) in enumerate(f):
                query_id, pos_id, neg_id = line.rstrip().split("\t")
                if query_id not in runs:
                    runs[query_id] = [set(pos_id), set(neg_id)]
                else:
                    runs[query_id][0].add(pos_id)
                    runs[query_id][1].add(neg_id)

        # it would generate language by language so that it would be easier to constrain that each batch only contain one language;
        for lang in tqdm(languages, desc=f"Preparing training example for {len(languages)} languages."):
            n_missed_q = 0
            n_missed_d = 0

            collection_path, queries_path = args["collection"][lang], args["queries"][lang]

            collection = {}
            with open(collection_path, encoding="utf-8") as f:
                collection = dict(line.rstrip().split("\t") for line in f)

            queries = {}
            with open(queries_path, encoding="utf-8") as f:
                for line in f:
                    queries = dict(line.rstrip().split("\t") for line in f)
 
            assert len(runs) == self.size_per_lang[lang]
            for query_id, (pos_ids, neg_ids) in runs.items():
                if query_id not in queries:
                    n_missed_q += 1
                    continue
 
                pos_ids, neg_ids = list(pos_ids), list(neg_ids)
                pos_ids = [d for d in pos_ids if d in collection]
                neg_ids = [d for d in neg_ids if d in collection]
                if len(neg_ids) == 0 or len(pos_ids) == 0:
                    n_missed_d += 1
                    continue

                NNEG = min(10, len(neg_ids))
                neg_ids = random.choices(neg_ids, k=NNEG)

                features = {
                    "query_id": query_id,
                    "query": queries[query_id],
                    "positive_passages": [{
                        "docid": pos_id, 
                        "text": collection[pos_id],
                    } for pos_id in pos_ids],
                    "negative_passages": [{
                        "docid": neg_id, 
                        "text": collection[neg_id],
                    } for neg_id in neg_ids],
                }
                yield f"{lang}-{query_id}-{idx}", features
            print(f'Number of missed Q: {n_missed_q}. Number of missed D: {n_missed_d}')