update

README.md

@@ -38,9 +38,8 @@ Tips:
 | tatoeba | [tatoeba](https://tatoeba.org/); [Tatoeba Paper](https://arxiv.org/abs/1812.10464v2) | TRAIN: 702895 | Tatoeba 是句子和翻译的集合。 | [tatoeba](https://huggingface.co/datasets/tatoeba) |
 | bucc2018 | [bucc2018](https://comparable.limsi.fr/bucc2018/bucc2018-task.html) | TRAIN: 2173318, TEST: 2125879 | 共享任务：识别可比语料库中的平行句子，语言：de, en, fr, ru, zh |  |
 | iwslt2017 | [2017.iwslt-1.1.pdf](https://aclanthology.org/2017.iwslt-1.1.pdf) | TRAIN: 2482649, VALID: 11480, TEST: 72470 | IWSLT 2017 多语言任务解决了文本翻译问题，涵盖英语、德语、荷兰语、意大利语和罗马尼亚语等所有方向。 | [iwslt2017](https://huggingface.co/datasets/iwslt2017) |
-| open_subtitles | [L16-1147.pdf](https://aclanthology.org/L16-1147.pdf) | 样本个数 | 我们推出了平行语料库 OpenSubtitles 集合的新主要版本。 该版本由大型电影和电视字幕数据库编译而成，共包含 1689 个双文本，涵盖 60 种语言的 26 亿个句子。 该版本还包含了字幕预处理和对齐方面的许多增强功能，例如自动更正 OCR 错误以及使用元数据来估计每个字幕的质量并对字幕对进行评分。 | [open_subtitles](https://huggingface.co/datasets/open_subtitles) |
-| bsd_ja_en | [2008.01940v1](https://arxiv.org/abs/2008.01940v1) | 样本个数 | 尽管由于并行语料库和基于语料库的训练技术的可用性不断增加，书面文本的机器翻译在过去几年中取得了长足的进步，但即使对于现代系统，口语文本和对话的自动翻译仍然具有挑战性。 在本文中，我们的目标是通过引入新构建的日语-英语商务会话平行语料库来提高会话文本的机器翻译质量。 | [bsd_ja_en](https://huggingface.co/datasets/bsd_ja_en) |
-| autshumato |  | 样本个数 | Autshumato 项目的目标之一是开发三种南非语言对的机器翻译系统。 | [autshumato](https://huggingface.co/datasets/autshumato) |
+| bsd_ja_en | [2008.01940v1](https://arxiv.org/abs/2008.01940v1) | TRAIN: 35755, VALID: 3636, TEST: 3702 | 尽管由于并行语料库和基于语料库的训练技术的可用性不断增加，书面文本的机器翻译在过去几年中取得了长足的进步，但即使对于现代系统，口语文本和对话的自动翻译仍然具有挑战性。 在本文中，我们的目标是通过引入新构建的日语-英语商务会话平行语料库来提高会话文本的机器翻译质量。 | [bsd_ja_en](https://huggingface.co/datasets/bsd_ja_en) |
+| autshumato |  | TRAIN: 652824 | Autshumato 项目的目标之一是开发三种南非语言对的机器翻译系统。 | [autshumato](https://huggingface.co/datasets/autshumato) |
 | chr_en | [2010.04791](https://arxiv.org/abs/2010.04791) | 样本个数 | ChrEn 是切罗基语-英语并行数据集，用于促进切罗基语和英语之间的机器翻译研究。 ChrEn 资源极少，总共包含 14k 个句子对，其分割方式有利于域内和域外评估。 ChrEn 还包含 5k 切罗基语单语数据以实现半监督学习。 | [chr_en](https://huggingface.co/datasets/chr_en) |
 | cmu_hinglish_dog | [CMU_DoG](https://github.com/festvox/datasets-CMU_DoG); [1809.07358](https://arxiv.org/abs/1809.07358) | 样本个数 | 这是印度英语（印地语-英语之间的代码混合）文本对话及其相应的英语版本的集合。 可用于两者之间的翻译。 该数据集由 CMU 的 Alan Black 教授团队提供。 | [cmu_hinglish_dog](https://huggingface.co/datasets/cmu_hinglish_dog) |
 | europa_eac_tm | [EAC-Translation Memory](https://joint-research-centre.ec.europa.eu/language-technology-resources/eac-translation-memory_en) | 样本个数 | 该数据集是从英语到多达 25 种语言的手动翻译的语料库，由欧盟教育和文化总局 (EAC) 于 2012 年发布。 | [europa_eac_tm](https://huggingface.co/datasets/europa_eac_tm) |
@@ -69,6 +68,7 @@ https://opus.nlpl.eu/
 | ecb | [ECB](https://opus.nlpl.eu/ECB/corpus/version/ECB); | 样本个数 |  | [ecb](https://huggingface.co/datasets/ecb) |
 | emea | [EMEA](https://opus.nlpl.eu/EMEA/corpus/version/EMEA); | 样本个数 |  | [emea](https://huggingface.co/datasets/emea) |
 | kde4 | [KDE4](https://opus.nlpl.eu/KDE4/corpus/version/KDE4); [apps.kde.org](https://apps.kde.org/zh-cn/); [opus.nlpl.eu](https://opus.nlpl.eu/) | 样本个数 |  | [kde4](https://huggingface.co/datasets/kde4) |
+| open_subtitles | [OpenSubtitles](https://opus.nlpl.eu/OpenSubtitles/corpus/version/OpenSubtitles); [L16-1147.pdf](https://aclanthology.org/L16-1147.pdf) | 样本个数 | 我们推出了平行语料库 OpenSubtitles 集合的新主要版本。 该版本由大型电影和电视字幕数据库编译而成，共包含 1689 个双文本，涵盖 60 种语言的 26 亿个句子。 该版本还包含了字幕预处理和对齐方面的许多增强功能，例如自动更正 OCR 错误以及使用元数据来估计每个字幕的质量并对字幕对进行评分。 | [open_subtitles](https://huggingface.co/datasets/open_subtitles) |
 | php | [PHP](https://opus.nlpl.eu/PHP/corpus/version/PHP) | 样本个数 | 最初从 http://se.php.net/download-docs.php 中提取的并行语料库。该语料库相当嘈杂。 | [php](https://huggingface.co/datasets/php) |
 
 

data/autshumato.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:016a32cb21c39511f8b541d298407dedafc13d20da8f65f63b28c5591fad1a96
+size 101604992

data/bsd_ja_en.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce59ca66051b94a699450ca16b233f27f021f9b083eddefe5d14b1c65167fa02
+size 5985101

dataset_details.md

@@ -69,6 +69,116 @@ zh-cn: 196260
 ![amazon_reviews_multi_text_length.jpg](docs/picture/amazon_reviews_multi_text_length.jpg)
 
 
+#### autshumato
+以下都是 train 训练集的信息
+
+```text
+语种数量：
+en: 292326
+ts: 207845
+tn: 125852
+zu: 26801
+```
+
+样本示例：
+
+| 数据 | 语种 | 样本 |
+| :---: | :---: | :---: |
+| autshumato | en | Good progress has been made with the staff composition project in each of the 15 faculties at the NWU . |
+| autshumato | en | The rector , Prof Thanyani Mariba , congratulated the newcomers on their choice to further their studies at the campus and emphasised the importance of choice and responsibility - both in terms of academic commitments and social endeavours . |
+| autshumato | en | Complaints against Correctional Services staff , court officials and members of the South African National Defence Force . |
+| autshumato | tn | Lo tla lemoga gore Thulaganyo ya Setheo ya 2012-2014 e e dirwang mo mafapheng otlhe ka tsamaiso ya ditumalano tsa go dira tiro ke ya gore YBB e fitlhelele maikemisetso a yone kgato ka kgato . |
+| autshumato | tn | Moreketoro , Mop Thanyani Mariba , o ne a akgolela batlabošeng tlhopho e ba e dirileng ya go tla go tswelela dithuto tsa bone mo khamphaseng eno mme o ne a gatelela botlhokwa jwa tlhopho le maikarabelo - malebana le go ineela ga bone mo dithutong le mo botshelong jwa bone jwa go tsalana le ba bangwe . |
+| autshumato | tn | Dingongorego kgatlhanong le badiredi ba Tirelo ya Ditshiamiso batlhankedi ba kgotlatshekelo le ditokololo tsa Mophato wa Phemelo wa Bosetšhaba wa Aforikaborwa . |
+| autshumato | zu | inkululeko yokwakha izinto ngokusebenzisa ubuciko; |
+| autshumato | zu | Lezozakhiwo ezithathwa njengabantu ngumthetho zingabanamalungelo akuMqulu wamaLungelo kuphela ngendlela edingwa uhlobo lwelungelo kanye nolwaleso sakhiwo esithathwa njengomuntu ngumthetho. |
+| autshumato | zu | Thina, Bantu baseNingizimu Afrika, Siyakukhumbula ukucekelwa phansi kwamalungelo okwenzeka eminyakeni eyadlula; |
+| autshumato | ts | Mahungu ya nkoka ya pfhumba ra dyondzo ra ndyangu wa hina hi lama landzelaka : |
+| autshumato | ts | xikan'we na nhlamuselo ya ntirho na xiyimo laha u nga ta tirha kona na leswaku nkarhi a wu nge hundzi malembe mambirhi |
+| autshumato | ts | loko xi laveka . |
+
+<details>
+<summary>文本长度</summary>
+<pre><code>0-10: 20573
+10-20: 47424
+20-30: 58434
+30-40: 61884
+40-50: 73557
+50-60: 70899
+60-70: 57249
+70-80: 42967
+80-90: 33702
+90-100: 26516
+100-110: 21149
+110-120: 18264
+120-130: 16390
+130-140: 14336
+140-150: 12944
+150-160: 11351
+160-170: 9839
+170-180: 8702
+180-190: 7294
+190-200: 6066
+200-210: 33284
+</code></pre>
+</details>
+
+文本长度统计图像：
+
+![autshumato_text_length.jpg](docs/picture/autshumato_text_length.jpg)
+
+
+#### bsd_ja_en
+以下都是 train 训练集的信息
+
+```text
+语种数量：
+ja: 18054
+en: 17701
+```
+
+样本示例：
+
+| 数据 | 语种 | 样本 |
+| :---: | :---: | :---: |
+| bsd_ja_en | en | Hi this is the systems development department of Company K. |
+| bsd_ja_en | en | My name is Takaichi from Company H. |
+| bsd_ja_en | en | Thank you as always. |
+| bsd_ja_en | ja | はい、K社システム開発部です。 |
+| bsd_ja_en | ja | H社の高市と申します。 |
+| bsd_ja_en | ja | いつもお世話になっております。 |
+
+<details>
+<summary>文本长度</summary>
+<pre><code>0-10: 1924
+10-20: 7921
+20-30: 7871
+30-40: 5637
+40-50: 3521
+50-60: 2557
+60-70: 1869
+70-80: 1399
+80-90: 944
+90-100: 721
+100-110: 496
+110-120: 324
+120-130: 224
+130-140: 123
+140-150: 85
+150-160: 51
+160-170: 33
+170-180: 19
+180-190: 18
+190-200: 9
+200-210: 9
+</code></pre>
+</details>
+
+文本长度统计图像：
+
+![bsd_ja_en_text_length.jpg](docs/picture/bsd_ja_en_text_length.jpg)
+
+
 #### bucc2018
 以下都是 train 训练集的信息
 

examples/make_subset_details.py

@@ -12,7 +12,7 @@ from project_settings import project_path
 
 def get_args():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--dataset_name", default="iwslt2017", type=str)
+    parser.add_argument("--dataset_name", default="bsd_ja_en", type=str)
     parser.add_argument(
         "--dataset_cache_dir",
         default=(project_path / "hub_datasets").as_posix(),

examples/preprocess/preprocess_autshumato.py

@@ -0,0 +1,93 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+import argparse
+from collections import defaultdict
+import json
+import os
+import sys
+
+pwd = os.path.abspath(os.path.dirname(__file__))
+sys.path.append(os.path.join(pwd, "../../"))
+
+from datasets import load_dataset, DownloadMode
+from tqdm import tqdm
+
+from language_identification import LANGUAGE_MAP
+from project_settings import project_path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dataset_path", default="autshumato", type=str)
+    parser.add_argument(
+        "--dataset_cache_dir",
+        default=(project_path / "hub_datasets").as_posix(),
+        type=str
+    )
+    parser.add_argument(
+        "--output_file",
+        default=(project_path / "data/autshumato.jsonl"),
+        type=str
+    )
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = get_args()
+
+    name_list = [
+        "autshumato-en-tn", "autshumato-en-zu", 
+        "autshumato-en-ts", "autshumato-en-ts-manual", 
+        "autshumato-tn", "autshumato-ts"
+    ]
+
+    text_set = set()
+    counter = defaultdict(int)
+    with open(args.output_file, "w", encoding="utf-8") as f:
+        for name in name_list:
+            dataset_dict = load_dataset(
+                path=args.dataset_path,
+                name=name,
+                cache_dir=args.dataset_cache_dir,
+                # download_mode=DownloadMode.FORCE_REDOWNLOAD
+            )
+            for k, v in dataset_dict.items():
+                split = k
+                if split not in ("train", "validation", "test"):
+                    print("skip split: {}".format(split))
+                    continue
+
+                for sample in tqdm(v):
+                    translation = sample.get("translation")
+                    if translation is None:
+                        break
+
+                    for language, text in translation.items():
+                        text = text.strip()
+
+                        if text in text_set:
+                            continue
+                        text_set.add(text)
+
+                        if language not in LANGUAGE_MAP.keys():
+                            raise AssertionError("language: {}, text: {}".format(language, text))
+
+                        row = {
+                            "text": text,
+                            "language": language,
+                            "data_source": "autshumato",
+                            "split": split
+                        }
+                        row = json.dumps(row, ensure_ascii=False)
+                        f.write("{}\n".format(row))
+                        counter[split] += 1
+
+    print("counter: {}".format(counter))
+
+    return
+
+
+if __name__ == "__main__":
+    main()

examples/preprocess/preprocess_bsd_ja_en.py

@@ -0,0 +1,86 @@
+#!/usr/bin/python3
+# -*- coding: utf-8 -*-
+import argparse
+from collections import defaultdict
+import json
+import os
+import sys
+
+pwd = os.path.abspath(os.path.dirname(__file__))
+sys.path.append(os.path.join(pwd, "../../"))
+
+from datasets import load_dataset, DownloadMode
+from tqdm import tqdm
+
+from language_identification import LANGUAGE_MAP
+from project_settings import project_path
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dataset_path", default="bsd_ja_en", type=str)
+    parser.add_argument(
+        "--dataset_cache_dir",
+        default=(project_path / "hub_datasets").as_posix(),
+        type=str
+    )
+    parser.add_argument(
+        "--output_file",
+        default=(project_path / "data/bsd_ja_en.jsonl"),
+        type=str
+    )
+
+    args = parser.parse_args()
+    return args
+
+
+def main():
+    args = get_args()
+
+    dataset_dict = load_dataset(
+        path=args.dataset_path,
+        cache_dir=args.dataset_cache_dir,
+        # download_mode=DownloadMode.FORCE_REDOWNLOAD
+    )
+    print(dataset_dict)
+
+    text_set = set()
+    counter = defaultdict(int)
+    with open(args.output_file, "w", encoding="utf-8") as f:
+        for k, v in dataset_dict.items():
+            split = k
+            if split not in ("train", "validation", "test"):
+                print("skip split: {}".format(split))
+                continue
+
+            for sample in tqdm(v):
+
+                en_sentence = sample["en_sentence"]
+                ja_sentence = sample["ja_sentence"]
+                for language, text in [("en", en_sentence), ("ja", ja_sentence)]:
+                    text = text.strip()
+    
+                    if text in text_set:
+                        continue
+                    text_set.add(text)
+
+                    if language not in LANGUAGE_MAP.keys():
+                        raise AssertionError(language)
+
+                    row = {
+                        "text": text,
+                        "language": language,
+                        "data_source": "bsd_ja_en",
+                        "split": split
+                    }
+                    row = json.dumps(row, ensure_ascii=False)
+                    f.write("{}\n".format(row))
+                    counter[split] += 1
+
+    print("counter: {}".format(counter))
+
+    return
+
+
+if __name__ == '__main__':
+    main()

language_identification.py

@@ -10,6 +10,8 @@ import datasets
 
 _URLS = {
     "amazon_reviews_multi": "data/amazon_reviews_multi.jsonl",
+    "autshumato": "data/autshumato.jsonl",
+    "bsd_ja_en": "data/bsd_ja_en.jsonl",
     "bucc2018": "data/bucc2018.jsonl",
     "iwslt2017": "data/iwslt2017.jsonl",
     "mike0307": "data/mike0307.jsonl",
@@ -64,12 +66,15 @@ LANGUAGE_MAP = {
     "sw": "swahili",
     "sv": "swedish",
     "th": "thai",
+    "tn": "sepedi",
     "tr": "turkish",
+    "ts": "dzonga",
     "ur": "urdu",
     "vi": "vietnamese",
     "zh": "chinese",
     "zh-cn": "simplified chinese",
     "zh-tw": "traditional chinese",
+    "zu": "zulu, south africa",
 }
 
 
@@ -78,6 +83,8 @@ class LanguageIdentification(datasets.GeneratorBasedBuilder):
 
     BUILDER_CONFIGS = [
         datasets.BuilderConfig(name="amazon_reviews_multi", version=VERSION, description="amazon_reviews_multi"),
+        datasets.BuilderConfig(name="autshumato", version=VERSION, description="autshumato"),
+        datasets.BuilderConfig(name="bsd_ja_en", version=VERSION, description="bsd_ja_en"),
         datasets.BuilderConfig(name="bucc2018", version=VERSION, description="bucc2018"),
         datasets.BuilderConfig(name="iwslt2017", version=VERSION, description="iwslt2017"),
         datasets.BuilderConfig(name="mike0307", version=VERSION, description="mike0307"),