wmt_vat.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.
"""Variance-Aware Machine Translation Test Sets"""

import os
import json
import textwrap
from typing import List

import datasets
from datasets.utils.download_manager import DownloadManager


_CITATION = """\
@inproceedings{
    zhan2021varianceaware,
    title={Variance-Aware Machine Translation Test Sets},
    author={Runzhe Zhan and Xuebo Liu and Derek F. Wong and Lidia S. Chao},
    booktitle={Thirty-fifth Conference on Neural Information Processing Systems, Datasets and Benchmarks Track},
    year={2021},
    url={https://openreview.net/forum?id=hhKA5k0oVy5}
}
"""

_DESCRIPTION = """\
The Variance-Aware Machine Translation corpus contains 70 small and discriminative test sets for machine translation (MT) 
evaluation called variance-aware test sets (VAT), covering 35 translation directions from WMT16 to WMT20 competitions. 
VAT is automatically created by a novel variance-aware filtering method that filters the indiscriminative test instances 
of the current MT benchmark without any human labor. Experimental results show that VAT outperforms the original WMT benchmark 
in terms of the correlation with human judgment across mainstream language pairs and test sets. Further analysis on the properties 
of VAT reveals the challenging linguistic features (e.g., translation of low-frequency words and proper nouns) for the competitive 
MT systems, providing guidance for constructing future MT test sets. 
"""

_HOMEPAGE = "https://github.com/NLP2CT/Variance-Aware-MT-Test-Sets"

_LICENSE = "https://raw.githubusercontent.com/NLP2CT/Variance-Aware-MT-Test-Sets/main/LICENSE"

_BASE_URL = "https://github.com/NLP2CT/Variance-Aware-MT-Test-Sets/raw/main/VAT_data"
_META_URL = "https://raw.githubusercontent.com/NLP2CT/Variance-Aware-MT-Test-Sets/raw/main/VAT_meta"

_CONFIGS = {
    "wmt16": ["tr_en", "ru_en", "ro_en", "de_en", "en_ru", "fi_en", "cs_en"],
    "wmt17": ["en_lv", "zh_en", "en_tr", "lv_en", "en_de", "ru_en", "en_fi", "tr_en", "en_zh", "en_ru", "fi_en", "en_cs", "de_en", "cs_en"],
    "wmt18": ["en_cs", "cs_en", "en_fi", "en_tr", "en_et", "ru_en", "et_en", "tr_en", "fi_en", "zh_en", "en_zh", "en_ru", "de_en", "en_de"],
    "wmt19": ["zh_en", "en_cs", "de_en", "en_gu", "fr_de", "en_zh", "fi_en", "en_fi", "kk_en", "de_cs", "lt_en", "en_lt", "ru_en", "en_kk", "en_ru", "gu_en", "de_fr", "en_de"],
    "wmt20": ["km_en", "cs_en", "en_de", "ja_en", "ps_en", "en_zh", "en_ta", "de_en", "zh_en", "en_ja", "en_cs", "en_pl", "en_ru", "pl_en", "iu_en", "ru_en", "ta_en"],
}

_PATHS = {
    f"{year}_{pair}": {
        "src" : os.path.join(_BASE_URL, year, f"vat_newstest20{year[3:]}-{pair.replace('_', '')}-src.{pair.split('_')[0]}"),
        "ref" : os.path.join(_BASE_URL, year, f"vat_newstest20{year[3:]}-{pair.replace('_', '')}-ref.{pair.split('_')[1]}")
    } for year, pairs in _CONFIGS.items() for pair in pairs
}

_METADATA_PATHS = {k:os.path.join(_META_URL, k, "bert-r_filter-std60.json") for k in _CONFIGS.keys()}


class WmtVatConfig(datasets.BuilderConfig):
    def __init__(
        self,
        campaign: str,
        source: str, 
        reference: str,
        **kwargs
    ):
        """BuilderConfig for Variance-Aware MT Test Sets.

        Args:
            campaign: `str`, WMT campaign from which the test set was extracted
            source: `str`, source for translation.
            reference: `str`, reference translation.
            **kwargs: keyword arguments forwarded to super.
        """
        super().__init__(**kwargs)
        self.campaign = campaign
        self.source = source
        self.reference = reference


class WmtVat(datasets.GeneratorBasedBuilder):
    """Variance-Aware Machine Translation Test Sets"""
    VERSION = datasets.Version("1.0.0")

    BUILDER_CONFIGS = [
        WmtVatConfig(
            name=cfg,
            campaign=cfg.split("_")[0],
            source=cfg.split("_")[1],
            reference=cfg.split("_")[2],
        ) for cfg in _PATHS.keys()
    ]

    def _info(self):
        features = datasets.Features(
            {
                "orig_id": datasets.Value("int32"),
                "source": datasets.Value("string"),
                "reference": datasets.Value("string")
            }
        )
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager: DownloadManager):
        """Returns SplitGenerators."""
        src_file = dl_manager.download_and_extract(_PATHS[self.config.name]["src"])
        ref_file = dl_manager.download_and_extract(_PATHS[self.config.name]["ref"])
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "src_path": src_file,
                    "ref_path": ref_file,
                    "pair": self.config.name[5:].replace("_", "-"),
                    "meta_path": _METADATA_PATHS[self.config.name[:5]] # Only wmtXX
                },
            )
        ]

    def _generate_examples(
        self, src_path: str, ref_path: str, pair: str, meta_path: str
    ):
        """ Yields examples as (key, example) tuples. """
        with open(meta_path, encoding="utf-8") as meta:
            ids = json.load(meta)[pair]
        with open(src_path, encoding="utf-8") as src:
            with open(src_path, encoding="utf-8") as ref:
                for id_, (src_ex, ref_ex, orig_idx) in enumerate(zip(src, ref, ids)):
                    yield id_, {
                        "orig_id": orig_idx,
                        "source": src_ex,
                        "reference": ref_ex,
                    }