[机器翻译-数据集] 批量获取所有WMT数据（初步解决方案）

前言

WMT是机器翻译和机器翻译研究的主要活动。 该会议每年与自然语言处理方面的大型会议联合举行。2006年，第一届机器翻译研讨会在计算语言学协会北美分会年会上举行。2016年，随着神经机器翻译的兴起，WMT成为了一个自己的会议。 机器翻译会议仍然主要被称为WMT[1]。
有些机器翻译工作会使用历年WMT公开的数据集作为他们的数据集[2]，如下图所示：

当笔者想要复现工作结果时，首先需要收集得到这样的数据集。而以WMT13[3]为例。如下图所示，笔者需要手动点击下载上面公开的每一个子数据集，然后汇总得到整个WMT13的训练、验证和测试集。而由于每一个子数据集的形式也不同，且数量较多…总的来说还是很麻烦的。

而笔者发现，huggingface[4]上面已经收集了部分年份的WMT数据，并提供了下载接口。以wmt14的所有hi-en数据为例，最终的下载结果如下图所示：

（笔者后知后觉意识到，只要想办法打开.arrow文件就可以得到对应数据了…艹）
本文旨在总结批量获取所有WMT数据的初步解决方案，通过修改huggingface datasets库的源码实现。

具体实现

第一步，pip install datasets安装datasets库。
第二步，通过git clone https://github.com/huggingface/datasets克隆datasets库，datasets/datasets路径下面包含了该库提供的所有数据集的相关代码：

第三步，创建主程序文件（run.py），代码如下，其中，py_file_path为上面说的datasets/datasets路径，save_dir为保存到本地的路径：

from datasets import load_dataset
import os
wmt_dict = {
    "wmt14": [(lang, "en") for lang in ["cs", "de", "fr", "hi", "ru"]],
    "wmt15": [(lang, "en") for lang in ["cs", "de", "fi", "fr", "ru"]],
    "wmt16": [(lang, "en") for lang in ["cs", "de", "fi", "ro", "ru", "tr"]],
    "wmt17": [(lang, "en") for lang in ["cs", "de", "fi", "lv", "ru", "tr", "zh"]],
    "wmt18": [(lang, "en") for lang in ["cs", "de", "et", "fi", "kk", "ru", "tr", "zh"]],
    "wmt19": [(lang, "en") for lang in ["cs", "de", "fi", "gu", "kk", "lt", "ru", "zh"]] + [("fr", "de")],
}

py_file_path = r"C:\Users\13359\PycharmProjects\for_fun\other\wmt_datasets\datasets\datasets"
save_dir = r"D:\dataset\mt"
for wmt in wmt_dict:
    for lang_tuple in wmt_dict[wmt]:
        lang_pair = "-".join(lang_tuple)
        print(f"wmt: {wmt} | lang_pair: {lang_pair}")
        load_dataset(os.path.join(py_file_path, wmt), name = lang_pair, cache_dir = save_dir)

第四步，在上述datasets/datasets路径下面随便选择一个wmt文件夹，比如wmt14，将里面的wmt_utils.py复制到run.py的同级目录下。（暂时不知道为何，尝试下来这样没有错），也就是文件目录结构如下：

第五步，如果此时运行run.py，则会像前言中的那样，得到所有wmt的所有语言对的数据，但数据格式是arrow的。笔者脑抽，没有直接去想如何把.arrow文件转成更好理解的格式，而是通过修改pip install下来的datasets源码，来直接修改保存数据的过程。具体来说，通过ctrl+B追溯run.py中load_dataset的执行顺序，最终找到了保存数据的源码位置：load_dataset(run.py)->builder_instance.download_and_prepare(load.py,1738行)->self._download_and_prepare(builder.py, 638行)->self._prepare_split(builder.py, 723行)，由于此时的self是一个wmtxx object，从具体的wmtxx.py（如wmt14.py，位于datasets/datasets/wmt14/wmt14.py）可知，wmtxx类的继承顺序是：wmtxx->Wmt(wmt_utils.py)->GeneratorBasedBuilder(builder.py)->DatasetBuilder(builder.py)，所以self._prepare_split最终方法实现是GeneratorBasedBuilder类的_prepare_split方法。
该方法中完成了.arrow数据的创建，具体代码如下所示：

with ArrowWriter(
    features=self.info.features,
    path=fpath,
    writer_batch_size=self._writer_batch_size,
    hash_salt=split_info.name,
    check_duplicates=check_duplicate_keys,
) as writer:
    try:
        for key, record in logging.tqdm(
            generator,
            unit=" examples",
            total=split_info.num_examples,
            leave=False,
            disable=not logging.is_progress_bar_enabled(),
            desc=f"Generating {split_info.name} split",
        ):
            example = self.info.features.encode_example(record)
            writer.write(example, key)
    finally:
        num_examples, num_bytes = writer.finalize()

split_generator.split_info.num_examples = num_examples
split_generator.split_info.num_bytes = num_bytes

其中，generator就是包含了所有数据的生成器。于是，笔者修改了上面的代码，完成了对数据保存的修改：

# ...其它代码
generator = self._generate_examples(**split_generator.gen_kwargs)

# 修改代码
user_name = "xushaoyang"
str_lst = fpath.split("\\")
index = str_lst.index(self.name)
lang_pair = str_lst[index + 1]
source, target = lang_pair.split("-")
path_lst = str_lst[:index + 2]
path_lst[index] += f"_{user_name}"
dir_path = os.path.join(*path_lst)
os.makedirs(dir_path, exist_ok=True)
source_file_name = f"{self.name}.{lang_pair}-{split_generator.name}.{source}"
source_path = os.path.join(dir_path, source_file_name)
target_file_name = f"{self.name}.{lang_pair}-{split_generator.name}.{target}"
target_path = os.path.join(dir_path, target_file_name)
# flag = "population of the US, or of the combined current population of India and China"
with open(source_path, mode="w", encoding="utf-8") as source_f:
    with open(target_path, mode="w", encoding="utf-8") as target_f:
        with ArrowWriter(
            features=self.info.features,
            path=fpath,
            writer_batch_size=self._writer_batch_size,
            hash_salt=split_info.name,
            check_duplicates=check_duplicate_keys,
        ) as writer:
            try:
                for key, record in logging.tqdm(
                    generator,
                    unit=" examples",
                    total=split_info.num_examples,
                    leave=False,
                    disable=not logging.is_progress_bar_enabled(),
                    desc=f"Generating {split_info.name} split",
                ):
                    assert list(record.keys()) == ["translation"]
                    lang_keys = list(record["translation"].keys())
                    if source not in lang_keys: # 问题1：zh的key在有些数据集里是ch
                        target_idx = lang_keys.index(target)
                        source_idx = (target_idx + 1) % 2
                        error_source_key = lang_keys[source_idx]
                        new_record = {"translation": {
                            source: record["translation"][error_source_key],
                            target: record["translation"][target]
                        }}
                        record = new_record
                        del new_record
                    # arrow write
                    example = self.info.features.encode_example(record)
                    writer.write(example, key)
                    # file write
                    source_sentence = record['translation'][source]
                    target_sentence = record['translation'][target]
                    source_sentence = source_sentence.replace("\r", "").replace("\n", "") # 问题2：多余的换行
                    target_sentence = target_sentence.replace("\r", "").replace("\n", "")
                    source_f.write(source_sentence + "\n")
                    target_f.write(target_sentence + "\n")
            finally:
                num_examples, num_bytes = writer.finalize()
split_generator.split_info.num_examples = num_examples
split_generator.split_info.num_bytes = num_bytes

完成修改后执行run.py，以wmt14的hi-en数据为例，得到的数据如下图所示，文件的命名仿照了OPUS100[5]：

不足之处

1. 总体来说，本篇博客价值不大。因为如果直接读取.arrow应该也行，比如参考：https://blog.csdn.net/wowotuo/article/details/110497489。如果这样的话，那完成前四步，直接运行run.py即可。
2. huggingface datasets库只提供了WMT14-19的下载，其它年份的WMT还是需要自行下载，当然也可以继续修改源码，把其它年份的加进去（TODO1）。
3. 下载的数据不是很全。wmt_utils.py中以SubDataset类的形式定义了所有WMT可能需要下载到的所有数据集，而以czeng_10为例，定义了manual_dl_files的数据集就是需要手动下载的。笔者目前的解决方案是，在下载过程中记录哪些数据集没有自动下载，所有下载完成之后再去补上（包括手动补上第2条中缺少的WMT数据），记录的代码见“补充”。
   
4. 另外，笔者在完整跑完run.py之后发现，wmt19的测试集是缺失的。

补充

补充一：记录暂时只能手动下载的数据集

如"不足之处"的第3点所述，有一些文件没有提供自动下载的url，笔者的解决方案是在下载过程中记录哪些数据集没有自动下载，所有下载完成之后再去手动补上。具体来说，笔者在wmt_utils.py中的_split_generators函数中的 if dataset.get_manual_dl_files(source):语句下，加入了如下语句：

with open(f"{self.name}_error_log", mode="a", encoding="utf-8") as file:
    file.write(f"lang: {'-'.join(self.config.language_pair)} | data_name: {dataset.name} | url: {str(dataset.get_manual_dl_files(source))}" + "\n")

在运行run.py的过程中，出现数据集缺失的情况，这样的记录就会被保存在wmtxxx_error_log日志文件中，如下图所示：

补充二：添加wmt19的测试集

打开datasets/datasets中的wmt19文件夹，修改里面的wmt19.py和wmt_utils.py。

# wmt19.py：line79
datasets.Split.TEST: ["newstest2019", "newstest2019_csen", "newstest2019_frde"],
# wmt_utils.py：line617
SubDataset(
    name="newstest2019",
    target="en",
    sources={"de", "fi", "gu", "kk", "lt", "ru", "zh"},
    url="http://data.statmt.org/wmt19/translation-task/test.tgz",
    path=("sgm/newstest2019-{src}en-src.{src}.sgm", "sgm/newstest2019-{src}en-ref.en.sgm"),
),

SubDataset(
    name="newstest2019_csen",
    target="en",
    sources={"cs"},
    url="http://data.statmt.org/wmt19/translation-task/test.tgz",
    path=("sgm/newstest2019-en{src}-src.en.sgm", "sgm/newstest2019-en{src}-ref.{src}.sgm"),
),

SubDataset(
    name="newstest2019_frde",
    target="de",
    sources={"fr"},
    url="http://data.statmt.org/wmt19/translation-task/test.tgz",
    path=("sgm/newstest2019-frde-ref.de.sgm", "sgm/newstest2019-frde-src.fr.sgm"),
),

另外，直接这样运行run.py会报错，因为dataset_infos.json中的内容还没有修改。而程序会读取dataset_info.json中一些预设的信息，然后和下载下来的结果进行一些校验：verify_checksums、verify_splits，但是直接修改dataset_infos.json较为麻烦，所以笔者选择取消校验：

load_dataset(os.path.join(py_file_path, wmt), name = lang_pair, cache_dir = save_dir, save_infos = True)
# ignore_verifications = True，也可以

另外，还是要对dataset_infos.json做一个细微的修改，即在splits中增加”test“：

      "test": {
        "name": "test",
        "num_bytes": 3000, # 这个随便设置的，不影响下载
        "num_examples": 3000, # 同上
        "dataset_name": "wmt19"
      }

下载结果记录

平行语料数量统计

wmt14

cs-en:
	train:
		source:953621
		target:953621
	validation:
		source:3000
		target:3000
	test:
		source:3003
		target:3003
de-en:
	train:
		source:4508785
		target:4508785
	validation:
		source:3000
		target:3000
	test:
		source:3003
		target:3003
fr-en:
	train:
		source:40836715
		target:40836715
	validation:
		source:3000
		target:3000
	test:
		source:3003
		target:3003
hi-en:
	train:
		source:32863
		target:32863
	validation:
		source:520
		target:520
	test:
		source:2507
		target:2507
ru-en:
	train:
		source:1486965
		target:1486965
	validation:
		source:3000
		target:3000
	test:
		source:3003
		target:3003

wmt15

cs-en:
	train:
		source:959768
		target:959768
	validation:
		source:3003
		target:3003
	test:
		source:2656
		target:2656
de-en:
	train:
		source:4522998
		target:4522998
	validation:
		source:3003
		target:3003
	test:
		source:2169
		target:2169
fi-en:
	train:
		source:2073394
		target:2073394
	validation:
		source:1500
		target:1500
	test:
		source:1370
		target:1370
fr-en:
	train:
		source:40853137
		target:40853137
	validation:
		source:4503
		target:4503
	test:
		source:1500
		target:1500
ru-en:
	train:
		source:1495081
		target:1495081
	validation:
		source:3003
		target:3003
	test:
		source:2818
		target:2818

wmt16

cs-en:
	train:
		source:997240
		target:997240
	validation:
		source:2656
		target:2656
	test:
		source:2999
		target:2999
de-en:
	train:
		source:4548885
		target:4548885
	validation:
		source:2169
		target:2169
	test:
		source:2999
		target:2999
fi-en:
	train:
		source:2073394
		target:2073394
	validation:
		source:1370
		target:1370
	test:
		source:6000
		target:6000
ro-en:
	train:
		source:610320
		target:610320
	validation:
		source:1999
		target:1999
	test:
		source:1999
		target:1999
ru-en:
	train:
		source:1516162
		target:1516162
	validation:
		source:2818
		target:2818
	test:
		source:2998
		target:2998
tr-en:
	train:
		source:205756
		target:205756
	validation:
		source:1001
		target:1001
	test:
		source:3000
		target:3000

wmt17

cs-en:
	train:
		source:1018291
		target:1018291
	validation:
		source:2999
		target:2999
	test:
		source:3005
		target:3005
de-en:
	train:
		source:5906184
		target:5906184
	validation:
		source:2999
		target:2999
	test:
		source:3004
		target:3004
fi-en:
	train:
		source:2656542
		target:2656542
	validation:
		source:6000
		target:6000
	test:
		source:6004
		target:6004
lv-en:
	train:
		source:3567528
		target:3567528
	validation:
		source:2003
		target:2003
	test:
		source:2001
		target:2001
ru-en:
	train:
		source:24782720
		target:24782720
	validation:
		source:2998
		target:2998
	test:
		source:3001
		target:3001
tr-en:
	train:
		source:205756
		target:205756
	validation:
		source:3000
		target:3000
	test:
		source:3007
		target:3007
zh-en:
	train:
		source:25134743
		target:25134743
	validation:
		source:2002
		target:2002
	test:
		source:2001
		target:2001

wmt18

cs-en:
	train:
		source:11046024
		target:11046024
	validation:
		source:3005
		target:3005
	test:
		source:2983
		target:2983
de-en:
	train:
		source:42271874
		target:42271874
	validation:
		source:3004
		target:3004
	test:
		source:2998
		target:2998
et-en:
	train:
		source:2175873
		target:2175873
	validation:
		source:2000
		target:2000
	test:
		source:2000
		target:2000
fi-en:
	train:
		source:3280600
		target:3280600
	validation:
		source:6004
		target:6004
	test:
		source:3000
		target:3000
kk-en:
	train:
		source:0
		target:0
	validation:
		source:0
		target:0
	test:
		source:0
		target:0
ru-en:
	train:
		source:36858512
		target:36858512
	validation:
		source:3001
		target:3001
	test:
		source:3000
		target:3000
tr-en:
	train:
		source:205756
		target:205756
	validation:
		source:3007
		target:3007
	test:
		source:3000
		target:3000
zh-en:
	train:
		source:25160346
		target:25160346
	validation:
		source:2001
		target:2001
	test:
		source:3981
		target:3981

wmt19

cs-en:
	train:
		source:7270695
		target:7270695
	validation:
		source:2983
		target:2983
	test:
		source:1997
		target:1997
de-en:
	train:
		source:38690334
		target:38690334
	validation:
		source:2998
		target:2998
	test:
		source:2000
		target:2000
fi-en:
	train:
		source:6587448
		target:6587448
	validation:
		source:3000
		target:3000
	test:
		source:1996
		target:1996
gu-en:
	train:
		source:11670
		target:11670
	validation:
		source:1998
		target:1998
	test:
		source:1016
		target:1016
kk-en:
	train:
		source:126583
		target:126583
	validation:
		source:2066
		target:2066
	test:
		source:1000
		target:1000
lt-en:
	train:
		source:2344893
		target:2344893
	validation:
		source:2000
		target:2000
	test:
		source:1000
		target:1000
ru-en:
	train:
		source:37492126
		target:37492126
	validation:
		source:3000
		target:3000
	test:
		source:2000
		target:2000
zh-en:
	train:
		source:25984574
		target:25984574
	validation:
		source:3981
		target:3981
	test:
		source:2000
		target:2000
fr-de:
	train:
		source:9824476
		target:9824476
	validation:
		source:1512
		target:1512
	test:
		source:1707
		target:1707

需要手动下载的语料统计

• wmt14:
  • lang: cs-en | data_name: czeng_10 | url: [‘data-plaintext-format.0.tar’, ‘data-plaintext-format.1.tar’, ‘data-plaintext-format.2.tar’, ‘data-plaintext-format.3.tar’, ‘data-plaintext-format.4.tar’, ‘data-plaintext-format.5.tar’, ‘data-plaintext-format.6.tar’, ‘data-plaintext-format.7.tar’, ‘data-plaintext-format.8.tar’, ‘data-plaintext-format.9.tar’]
  • lang: ru-en | data_name: yandexcorpus | url: [‘1mcorpus.zip’]
  • lang: hi-en | data_name: hindencorp_01 | url: [‘hindencorp0.1.gz’]
• wmt15：
  • lang: cs-en | data_name: czeng_10 | url: [‘data-plaintext-format.0.tar’, ‘data-plaintext-format.1.tar’, ‘data-plaintext-format.2.tar’, ‘data-plaintext-format.3.tar’, ‘data-plaintext-format.4.tar’, ‘data-plaintext-format.5.tar’, ‘data-plaintext-format.6.tar’, ‘data-plaintext-format.7.tar’, ‘data-plaintext-format.8.tar’, ‘data-plaintext-format.9.tar’]
  • lang: ru-en | data_name: yandexcorpus | url: [‘1mcorpus.zip’]
• wmt16：
  • lang: cs-en | data_name: czeng_16pre | url: [‘czeng16pre.deduped-ignoring-sections.txt.gz’]
  • lang: ru-en | data_name: yandexcorpus | url: [‘1mcorpus.zip’]
• wmt17：
  • lang: cs-en | data_name: czeng_16 | url: [‘data-plaintext-format.0.tar’, ‘data-plaintext-format.1.tar’, ‘data-plaintext-format.2.tar’, ‘data-plaintext-format.3.tar’, ‘data-plaintext-format.4.tar’, ‘data-plaintext-format.5.tar’, ‘data-plaintext-format.6.tar’, ‘data-plaintext-format.7.tar’, ‘data-plaintext-format.8.tar’, ‘data-plaintext-format.9.tar’]
  • lang: ru-en | data_name: yandexcorpus | url: [‘1mcorpus.zip’]
• wmt18：
  • lang: cs-en | data_name: czeng_17 | url: [‘data-plaintext-format.0.tar’, ‘data-plaintext-format.1.tar’, ‘data-plaintext-format.2.tar’, ‘data-plaintext-format.3.tar’, ‘data-plaintext-format.4.tar’, ‘data-plaintext-format.5.tar’, ‘data-plaintext-format.6.tar’, ‘data-plaintext-format.7.tar’, ‘data-plaintext-format.8.tar’, ‘data-plaintext-format.9.tar’]
  • lang: ru-en | data_name: yandexcorpus | url: [‘1mcorpus.zip’]
• wmt19：
  • lang: cs-en | data_name: czeng_17 | url: [‘data-plaintext-format.0.tar’, ‘data-plaintext-format.1.tar’, ‘data-plaintext-format.2.tar’, ‘data-plaintext-format.3.tar’, ‘data-plaintext-format.4.tar’, ‘data-plaintext-format.5.tar’, ‘data-plaintext-format.6.tar’, ‘data-plaintext-format.7.tar’, ‘data-plaintext-format.8.tar’, ‘data-plaintext-format.9.tar’]
  • lang: ru-en | data_name: yandexcorpus | url: [‘1mcorpus.zip’]

可以看到存在大量的重复。以上语料的下载链接统计如下：

• czeng_10：https://ufal.mff.cuni.cz/legacy/czeng/czeng10/
• yandexcorpus：https://translate.yandex.ru/corpus?lang=en
• hindencorp_01：http://ufallab.ms.mff.cuni.cz/~bojar/hindencorp/
• czeng_16pre：https://ufal.mff.cuni.cz/czeng/czeng16pre
• czeng_16：https://ufal.mff.cuni.cz/czeng/czeng16
• czeng_17：https://ufal.mff.cuni.cz/czeng/czeng17

基本都需要提交申请

参考

[1]https://machinetranslate.org/wmt
[2]https://arxiv.org/pdf/2105.09259v1.pdf
[3]https://www.statmt.org/wmt14/translation-task.html
[4]https://github.com/huggingface/datasets
[5]https://github.com/EdinburghNLP/opus-100-corpus