nllb.py

# coding=utf-8
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# 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.
"""No Language Left Behind (NLLB)

The "No Language Left Behind" paper is a dataset with translation examples across 200 languages.
This paper reuses prior work, and for some language pairs is just reusing CC-Matrix published on statmt.org.
Depending on the language pair chosen, this script will fetch either the original
version from statmt.org, or the new one from AllenAI.
"""

import datasets
import typing as tp

NLLB_CITATION = (
    "@article{team2022NoLL,"
    "title={No Language Left Behind: Scaling Human-Centered Machine Translation},"
    r"author={Nllb team and Marta Ruiz Costa-juss{\`a} and James Cross and Onur Celebi and Maha Elbayad and Kenneth Heafield and Kevin Heffernan and Elahe Kalbassi and Janice Lam and Daniel Licht and Jean Maillard and Anna Sun and Skyler Wang and Guillaume Wenzek and Alison Youngblood and Bapi Akula and Lo{\"i}c Barrault and Gabriel Mejia Gonzalez and Prangthip Hansanti and John Hoffman and Semarley Jarrett and Kaushik Ram Sadagopan and Dirk Rowe and Shannon L. Spruit and C. Tran and Pierre Andrews and Necip Fazil Ayan and Shruti Bhosale and Sergey Edunov and Angela Fan and Cynthia Gao and Vedanuj Goswami and Francisco Guzm'an and Philipp Koehn and Alexandre Mourachko and Christophe Ropers and Safiyyah Saleem and Holger Schwenk and Jeff Wang},"
    "journal={ArXiv},"
    "year={2022},"
    "volume={abs/2207.04672}"
    "}"
)

CCMATRIX_CITATION = (
    "@inproceedings{schwenk2021ccmatrix,"
    "title={CCMatrix: Mining Billions of High-Quality Parallel Sentences on the Web},"
    "author={Schwenk, Holger and Wenzek, Guillaume and Edunov, Sergey and Grave, {'E}douard and Joulin, Armand and Fan, Angela},"
    "booktitle={Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},"
    "pages={6490--6500},"
    "year={2021}"
)


_DESCRIPTION = ""  # TODO

_HOMEPAGE = ""  # TODO

_LICENSE = "https://opendatacommons.org/licenses/by/1-0/"

from .nllb_lang_pairs import NLLB_PAIRS, CCMATRIX_PAIRS, CCMATRIX_MAPPING

_ALLENAI_URL = "https://storage.googleapis.com/allennlp-data-bucket/nllb/"
_STATMT_URL = "http://data.statmt.org/cc-matrix/"


class NLLBTaskConfig(datasets.BuilderConfig):
    """BuilderConfig for No Language Left Behind Dataset."""

    def __init__(self, src_lg, tgt_lg, url, **kwargs):
        super(NLLBTaskConfig, self).__init__(**kwargs)
        self.src_lg = src_lg
        self.tgt_lg = tgt_lg
        self.url = url
        self.source = "statmt" if url.startswith(_STATMT_URL) else "allenai"

def _builder_configs() -> tp.List[NLLBTaskConfig]:
    """
    Creates the dataset used for training NLLB model.

    Note we always return data from AllenAI if possible because CC-Matrix data
    is older, and most language pairs have been improved between the two versions.
    """
    configs = {}

    for (src_lg, tgt_lg) in NLLB_PAIRS:
        assert (src_lg, tgt_lg) not in configs
        configs[(src_lg, tgt_lg)] = NLLBTaskConfig(
            name=f"{src_lg}-{tgt_lg}",
            version=datasets.Version("1.0.0"),
            description=f"No Language Left Behind (NLLB): {src_lg} - {tgt_lg}",
            src_lg=src_lg,
            tgt_lg=tgt_lg,
            url=f"{_ALLENAI_URL}{src_lg}-{tgt_lg}.gz",
        )

    for (src_lg, tgt_lg) in CCMATRIX_PAIRS:
        # Prevent accidental override
        # Note: the lang pairs are not consistently alphabetically sorted,
        # because CCMatrix was using other language code which may swap the order.
        # Let's keep the original order because they correspond to the column order in the file.
        assert (src_lg, tgt_lg) not in configs
        assert (tgt_lg, src_lg) not in configs
        src_cc, tgt_cc = CCMATRIX_MAPPING[src_lg], CCMATRIX_MAPPING[tgt_lg]
        configs[(src_lg, tgt_lg)] = NLLBTaskConfig(
            name=f"{src_lg}-{tgt_lg}",
            version=datasets.Version("1.0.0"),
            description=f"No Language Left Behind (NLLB): {src_lg} - {tgt_lg}",
            src_lg=src_lg,
            tgt_lg=tgt_lg,
            # Use CCMatrix language code to fetch from statmt
            url=f"{_STATMT_URL}{src_cc}-{tgt_cc}.bitextf.tsv.gz",
        )

    return list(configs.values())


class NLLB(datasets.GeneratorBasedBuilder):
    """No Language Left Behind Dataset."""

    BUILDER_CONFIGS = _builder_configs()
    BUILDER_CONFIG_CLASS = NLLBTaskConfig

    def _info(self):
        # define feature types
        citation = NLLB_CITATION
        features = datasets.Features(
            {
                "translation": datasets.Translation(
                    languages=(self.config.src_lg, self.config.tgt_lg)
                ),
                "laser_score": datasets.Value("float32"),
                "source_sentence_lid": datasets.Value("float32"),
                "target_sentence_lid": datasets.Value("float32"),
                "source_sentence_source": datasets.Value("string"),
                "source_sentence_perplexity": datasets.Value("float32"),
                "source_sentence_url": datasets.Value("string"),
                "target_sentence_source": datasets.Value("string"),
                "target_sentence_perplexity": datasets.Value("float32"),
                "target_sentence_url": datasets.Value("string"),
            }
        )
        if self.config.source == "statmt":
            citation = CCMATRIX_CITATION
            # MT stats didn't published all the metadata
            features = datasets.Features(
                {
                    "translation": datasets.Translation(
                        languages=(self.config.src_lg, self.config.tgt_lg)
                    ),
                    "laser_score": datasets.Value("float32"),
                }
            )

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

    def _split_generators(self, dl_manager):
        """Returns one training generator. NLLB200 is meant for training.

        If you're interested in evaluation look at https://huggingface.co/datasets/facebook/flores
        """
        local_file = dl_manager.download_and_extract(self.config.url)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "filepath": local_file,
                    "source_lg": self.config.src_lg,
                    "target_lg": self.config.tgt_lg,
                },
            )
        ]

    def _generate_examples(self, filepath, source_lg, target_lg, source_max_perplexity=None, target_max_perplexity=None):
        if self.config.source == "statmt":
            # MT stats didn't published all the metadata
            return self._generate_minimal_examples(filepath, source_lg, target_lg)

        return self._generate_full_examples(filepath, source_lg, target_lg, source_max_perplexity=source_max_perplexity, target_max_perplexity=target_max_perplexity)

    def _generate_full_examples(self, filepath, source_lg, target_lg, source_max_perplexity=None, target_max_perplexity=None):
        with open(filepath, encoding="utf-8") as f:
            # reader = csv.reader(f, delimiter="\t")
            for id_, example in enumerate(f):
                try:
                    datarow = example.rstrip("\n").split("\t")
                    row = {}
                    # create translation json
                    row["translation"] = {
                        source_lg: datarow[0],
                        target_lg: datarow[1],
                    }

                    row["source_sentence_perplexity"] = None # TODO: compute the perplexity for the source sentence here
                    row["target_sentence_perplexity"] = None # TODO: compute the perplexity for the target sentence here
                    row["laser_score"] = float(datarow[2])
                    row["source_sentence_lid"] = float(datarow[3])
                    row["target_sentence_lid"] = float(datarow[4])
                    row["source_sentence_source"] = datarow[5]
                    row["source_sentence_url"] = datarow[6]
                    row["target_sentence_source"] = datarow[7]
                    row["target_sentence_url"] = datarow[8]
                    # replace empty values
                    row = {k: None if not v else v for k, v in row.items()}
                except:
                    print(datarow)
                    raise
                if source_max_perplexity is not None and row["source_sentence_perplexity"] > source_max_perplexity:
                    continue
                if target_max_perplexity is not None and row["target_sentence_perplexity"] > target_max_perplexity:
                    continue
                yield id_, row

    def _generate_minimal_examples(self, filepath, source_lg, target_lg):
        with open(filepath, encoding="utf-8") as f:
            for i, example in enumerate(f):
                try:
                    (score, src, tgt) = example.rstrip("\n").split("\t")
                    row = {
                        "translation": {
                            source_lg: src,
                            target_lg: tgt,
                        },
                        "laser_score": score,
                    }
                except:
                    print(example)
                    raise
                yield i, row


# to test the script, go to the root folder of the repo (nllb) and run:
# datasets-cli test nllb --save_infos --all_configs