Datasets:
ArXiv:
License:
Upload oscar_2201.py with huggingface_hub
Browse files- oscar_2201.py +340 -0
oscar_2201.py
ADDED
@@ -0,0 +1,340 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import gzip
|
3 |
+
import json
|
4 |
+
from pathlib import Path
|
5 |
+
from typing import Dict, List, Tuple
|
6 |
+
from urllib.parse import urljoin
|
7 |
+
|
8 |
+
import datasets
|
9 |
+
|
10 |
+
from seacrowd.utils import schemas
|
11 |
+
from seacrowd.utils.configs import SEACrowdConfig
|
12 |
+
from seacrowd.utils.constants import Tasks, Licenses
|
13 |
+
|
14 |
+
_CITATION = """\
|
15 |
+
@inproceedings{abadji2022cleaner,
|
16 |
+
author = {Julien Abadji and
|
17 |
+
Pedro Javier Ortiz Su{\'{a}}rez and
|
18 |
+
Laurent Romary and
|
19 |
+
Beno{\^{\i}}t Sagot},
|
20 |
+
title = {Towards a Cleaner Document-Oriented Multilingual Crawled Corpus},
|
21 |
+
booktitle = {Proceedings of the Thirteenth Language Resources and Evaluation Conference,
|
22 |
+
{LREC} 2022, Marseille, France, 20-25 June 2022},
|
23 |
+
pages = {4344--4355},
|
24 |
+
publisher = {European Language Resources Association},
|
25 |
+
year = {2022},
|
26 |
+
url = {https://aclanthology.org/2022.lrec-1.463},
|
27 |
+
}
|
28 |
+
|
29 |
+
@inproceedings{abadji2021ungoliant,
|
30 |
+
author = {Julien Abadji and
|
31 |
+
Pedro Javier Ortiz Su{\'a}rez and
|
32 |
+
Laurent Romary and
|
33 |
+
Beno{\^i}t Sagot},
|
34 |
+
title = {Ungoliant: An optimized pipeline for the generation of a very large-scale multilingual web corpus},
|
35 |
+
series = {Proceedings of the Workshop on Challenges in the Management of Large Corpora
|
36 |
+
(CMLC-9) 2021. Limerick, 12 July 2021 (Online-Event)},
|
37 |
+
editor = {Harald L{\"u}ngen and
|
38 |
+
Marc Kupietz and
|
39 |
+
Piotr Bański and
|
40 |
+
Adrien Barbaresi and
|
41 |
+
Simon Clematide and
|
42 |
+
Ines Pisetta},
|
43 |
+
publisher = {Leibniz-Institut f{\"u}r Deutsche Sprache},
|
44 |
+
address = {Mannheim},
|
45 |
+
doi = {10.14618/ids-pub-10468},
|
46 |
+
url = {https://nbn-resolving.org/urn:nbn:de:bsz:mh39-104688},
|
47 |
+
pages = {1 -- 9},
|
48 |
+
year = {2021},
|
49 |
+
abstract = {Since the introduction of large language models in Natural Language
|
50 |
+
Processing, large raw corpora have played a crucial role in Computational Linguistics.
|
51 |
+
However, most of these large raw corpora are either available only for English or not
|
52 |
+
available to the general public due to copyright issues. Nevertheless, there are some
|
53 |
+
examples of freely available multilingual corpora for training Deep Learning NLP
|
54 |
+
models, such as the OSCAR and Paracrawl corpora. However, they have quality issues,
|
55 |
+
especially for low-resource languages. Moreover, recreating or updating these corpora
|
56 |
+
is very complex. In this work, we try to reproduce and improve the goclassy pipeline
|
57 |
+
used to create the OSCAR corpus. We propose a new pipeline that is faster, modular,
|
58 |
+
parameterizable, and well documented. We use it to create a corpus similar to OSCAR
|
59 |
+
but larger and based on recent data. Also, unlike OSCAR, the metadata information is
|
60 |
+
at the document level. We release our pipeline under an open source license and
|
61 |
+
publish the corpus under a research-only license.},
|
62 |
+
language = {en}
|
63 |
+
}
|
64 |
+
|
65 |
+
@article{kreutzer2022quality,
|
66 |
+
title = {Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets},
|
67 |
+
author = {Kreutzer, Julia and
|
68 |
+
Caswell, Isaac and
|
69 |
+
Wang, Lisa and
|
70 |
+
Wahab, Ahsan and
|
71 |
+
van Esch, Daan and
|
72 |
+
Ulzii-Orshikh, Nasanbayar and
|
73 |
+
Tapo, Allahsera and
|
74 |
+
Subramani, Nishant and
|
75 |
+
Sokolov, Artem and
|
76 |
+
Sikasote, Claytone and
|
77 |
+
Setyawan, Monang and
|
78 |
+
Sarin, Supheakmungkol and
|
79 |
+
Samb, Sokhar and
|
80 |
+
Sagot, Beno{\^\i}t and
|
81 |
+
Rivera, Clara and
|
82 |
+
Rios, Annette and
|
83 |
+
Papadimitriou, Isabel and
|
84 |
+
Osei, Salomey and
|
85 |
+
Suarez, Pedro Ortiz and
|
86 |
+
Orife, Iroro and
|
87 |
+
Ogueji, Kelechi and
|
88 |
+
Rubungo, Andre Niyongabo and
|
89 |
+
Nguyen, Toan Q. and
|
90 |
+
M{\"u}ller, Mathias and
|
91 |
+
M{\"u}ller, Andr{\'e} and
|
92 |
+
Muhammad, Shamsuddeen Hassan and
|
93 |
+
Muhammad, Nanda and
|
94 |
+
Mnyakeni, Ayanda and
|
95 |
+
Mirzakhalov, Jamshidbek and
|
96 |
+
Matangira, Tapiwanashe and
|
97 |
+
Leong, Colin and
|
98 |
+
Lawson, Nze and
|
99 |
+
Kudugunta, Sneha and
|
100 |
+
Jernite, Yacine and
|
101 |
+
Jenny, Mathias and
|
102 |
+
Firat, Orhan and
|
103 |
+
Dossou, Bonaventure F. P. and
|
104 |
+
Dlamini, Sakhile and
|
105 |
+
de Silva, Nisansa and
|
106 |
+
{\c{C}}abuk Ball{\i}, Sakine and
|
107 |
+
Biderman, Stella and
|
108 |
+
Battisti, Alessia and
|
109 |
+
Baruwa, Ahmed and
|
110 |
+
Bapna, Ankur and
|
111 |
+
Baljekar, Pallavi and
|
112 |
+
Azime, Israel Abebe and
|
113 |
+
Awokoya, Ayodele and
|
114 |
+
Ataman, Duygu and
|
115 |
+
Ahia, Orevaoghene and
|
116 |
+
Ahia, Oghenefego and
|
117 |
+
Agrawal, Sweta and
|
118 |
+
Adeyemi, Mofetoluwa},
|
119 |
+
editor = {Roark, Brian and
|
120 |
+
Nenkova, Ani},
|
121 |
+
journal = {Transactions of the Association for Computational Linguistics},
|
122 |
+
volume = {10},
|
123 |
+
year = {2022},
|
124 |
+
address = {Cambridge, MA},
|
125 |
+
publisher = {MIT Press},
|
126 |
+
url = {https://aclanthology.org/2022.tacl-1.4},
|
127 |
+
doi = {10.1162/tacl_a_00447},
|
128 |
+
pages = {50--72},
|
129 |
+
abstract = {With the success of large-scale pre-training and multilingual modeling in
|
130 |
+
Natural Language Processing (NLP), recent years have seen a proliferation of large,
|
131 |
+
Web-mined text datasets covering hundreds of languages. We manually audit the quality
|
132 |
+
of 205 language-specific corpora released with five major public datasets (CCAligned,
|
133 |
+
ParaCrawl, WikiMatrix, OSCAR, mC4). Lower-resource corpora have systematic issues: At
|
134 |
+
least 15 corpora have no usable text, and a significant fraction contains less than
|
135 |
+
50{\%} sentences of acceptable quality. In addition, many are mislabeled or use
|
136 |
+
nonstandard/ambiguous language codes. We demonstrate that these issues are easy to
|
137 |
+
detect even for non-proficient speakers, and supplement the human audit with automatic
|
138 |
+
analyses. Finally, we recommend techniques to evaluate and improve multilingual
|
139 |
+
corpora and discuss potential risks that come with low-quality data releases.},
|
140 |
+
}
|
141 |
+
|
142 |
+
@inproceedings{ortizsuarez2020monolingual,
|
143 |
+
title = {A Monolingual Approach to Contextualized Word Embeddings for Mid-Resource Languages},
|
144 |
+
author = {Ortiz Su{'a}rez, Pedro Javier and
|
145 |
+
Romary, Laurent and
|
146 |
+
Sagot, Benoit},
|
147 |
+
booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
|
148 |
+
month = {jul},
|
149 |
+
year = {2020},
|
150 |
+
address = {Online},
|
151 |
+
publisher = {Association for Computational Linguistics},
|
152 |
+
url = {https://www.aclweb.org/anthology/2020.acl-main.156},
|
153 |
+
pages = {1703--1714},
|
154 |
+
abstract = {We use the multilingual OSCAR corpus, extracted from Common Crawl via
|
155 |
+
language classification, filtering and cleaning, to train monolingual contextualized
|
156 |
+
word embeddings (ELMo) for five mid-resource languages. We then compare the
|
157 |
+
performance of OSCAR-based and Wikipedia-based ELMo embeddings for these languages on
|
158 |
+
the part-of-speech tagging and parsing tasks. We show that, despite the noise in the
|
159 |
+
Common-Crawl-based OSCAR data, embeddings trained on OSCAR perform much better than
|
160 |
+
monolingual embeddings trained on Wikipedia. They actually equal or improve the
|
161 |
+
current state of the art in tagging and parsing for all five languages. In particular,
|
162 |
+
they also improve over multilingual Wikipedia-based contextual embeddings
|
163 |
+
(multilingual BERT), which almost always constitutes the previous state of the art,
|
164 |
+
thereby showing that the benefit of a larger, more diverse corpus surpasses the
|
165 |
+
cross-lingual benefit of multilingual embedding architectures.},
|
166 |
+
}
|
167 |
+
|
168 |
+
@inproceedings{ortizsuarez2019asynchronous,
|
169 |
+
author = {Pedro Javier {Ortiz Su{'a}rez} and
|
170 |
+
Benoit Sagot and
|
171 |
+
Laurent Romary},
|
172 |
+
title = {Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures},
|
173 |
+
series = {Proceedings of the Workshop on Challenges in the Management of Large Corpora
|
174 |
+
(CMLC-7) 2019. Cardiff, 22nd July 2019},
|
175 |
+
editor = {Piotr Bański and
|
176 |
+
Adrien Barbaresi and
|
177 |
+
Hanno Biber and
|
178 |
+
Evelyn Breiteneder and
|
179 |
+
Simon Clematide and
|
180 |
+
Marc Kupietz and
|
181 |
+
Harald L{"u}ngen and
|
182 |
+
Caroline Iliadi},
|
183 |
+
publisher = {Leibniz-Institut f{"u}r Deutsche Sprache},
|
184 |
+
address = {Mannheim},
|
185 |
+
doi = {10.14618/ids-pub-9021},
|
186 |
+
url = {http://nbn-resolving.de/urn:nbn:de:bsz:mh39-90215},
|
187 |
+
pages = {9 -- 16},
|
188 |
+
year = {2019},
|
189 |
+
abstract = {Common Crawl is a considerably large, heterogeneous multilingual corpus
|
190 |
+
comprised of crawled documents from the internet, surpassing 20TB of data and
|
191 |
+
distributed as a set of more than 50 thousand plain text files where each contains
|
192 |
+
many documents written in a wide variety of languages. Even though each document has a
|
193 |
+
metadata block associated to it, this data lacks any information about the language in
|
194 |
+
which each document is written, making it extremely difficult to use Common Crawl for
|
195 |
+
monolingual applications. We propose a general, highly parallel, multithreaded
|
196 |
+
pipeline to clean and classify Common Crawl by language; we specifically design it so
|
197 |
+
that it runs efficiently on medium to low resource infrastructures where I/O speeds
|
198 |
+
are the main constraint. We develop the pipeline so that it can be easily reapplied to
|
199 |
+
any kind of heterogeneous corpus and so that it can be parameterised to a wide range
|
200 |
+
of infrastructures. We also distribute a 6.3TB version of Common Crawl, filtered,
|
201 |
+
classified by language, shuffled at line level in order to avoid copyright issues, and
|
202 |
+
ready to be used for NLP applications.},
|
203 |
+
language = {en}
|
204 |
+
}
|
205 |
+
"""
|
206 |
+
|
207 |
+
_DATASETNAME = "oscar_2201"
|
208 |
+
_DESCRIPTION = """\
|
209 |
+
OSCAR or Open Super-large Crawled Aggregated coRpus is a huge multilingual corpus
|
210 |
+
obtained by language classification and filtering of the Common Crawl corpus using
|
211 |
+
the ungoliant architecture. Data is distributed by language in both original and
|
212 |
+
deduplicated form.
|
213 |
+
"""
|
214 |
+
|
215 |
+
_HOMEPAGE = "https://huggingface.co/datasets/oscar-corpus/OSCAR-2201"
|
216 |
+
_LICENSE = Licenses.CC0_1_0.value
|
217 |
+
_BASE_URL = "https://huggingface.co/datasets/oscar-corpus/OSCAR-2201/resolve/main/compressed/{lang}_meta/"
|
218 |
+
|
219 |
+
_LOCAL = False
|
220 |
+
_LANGUAGES = ["war", "ceb", "min", "vie", "ilo", "tgl", "lao", "khm", "mya", "jav", "ind", "tha", "sun", "zlm"]
|
221 |
+
|
222 |
+
_SUPPORTED_TASKS = [Tasks.SELF_SUPERVISED_PRETRAINING]
|
223 |
+
_SOURCE_VERSION = "2022.1.0"
|
224 |
+
_SEACROWD_VERSION = "2024.06.20"
|
225 |
+
|
226 |
+
|
227 |
+
class Oscar2201Dataset(datasets.GeneratorBasedBuilder):
|
228 |
+
"""OSCAR subset for SEA languages, version 2201."""
|
229 |
+
|
230 |
+
SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
|
231 |
+
SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
|
232 |
+
|
233 |
+
SEACROWD_SCHEMA_NAME = "ssp"
|
234 |
+
SUBSETS = ["war", "ceb", "min", "vi", "ta", "ilo", "tl", "lo", "km", "my", "jv", "id", "th", "su", "ms"]
|
235 |
+
|
236 |
+
BUILDER_CONFIGS = [
|
237 |
+
SEACrowdConfig(
|
238 |
+
name=f"{_DATASETNAME}_{subset}_source",
|
239 |
+
version=datasets.Version(_SOURCE_VERSION),
|
240 |
+
description=f"{_DATASETNAME} {subset} source schema",
|
241 |
+
schema="source",
|
242 |
+
subset_id=subset,
|
243 |
+
) for subset in SUBSETS
|
244 |
+
] + [
|
245 |
+
SEACrowdConfig(
|
246 |
+
name=f"{_DATASETNAME}_{subset}_seacrowd_ssp",
|
247 |
+
version=datasets.Version(_SEACROWD_VERSION),
|
248 |
+
description=f"{_DATASETNAME} {subset} SEACrowd schema",
|
249 |
+
schema="seacrowd_ssp",
|
250 |
+
subset_id=subset,
|
251 |
+
)
|
252 |
+
for subset in SUBSETS
|
253 |
+
]
|
254 |
+
|
255 |
+
DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_jv_source"
|
256 |
+
|
257 |
+
def _info(self) -> datasets.DatasetInfo:
|
258 |
+
if self.config.schema == "source":
|
259 |
+
features = datasets.Features(
|
260 |
+
{
|
261 |
+
"id": datasets.Value("int64"),
|
262 |
+
"text": datasets.Value("string"),
|
263 |
+
"meta": {
|
264 |
+
"warc_headers": {
|
265 |
+
"warc-record-id": datasets.Value("string"),
|
266 |
+
"warc-date": datasets.Value("string"),
|
267 |
+
"content-type": datasets.Value("string"),
|
268 |
+
"content-length": datasets.Value("int32"),
|
269 |
+
"warc-type": datasets.Value("string"),
|
270 |
+
"warc-identified-content-language": datasets.Value("string"),
|
271 |
+
"warc-refers-to": datasets.Value("string"),
|
272 |
+
"warc-target-uri": datasets.Value("string"),
|
273 |
+
"warc-block-digest": datasets.Value("string"),
|
274 |
+
},
|
275 |
+
"identification": {
|
276 |
+
"label": datasets.Value("string"),
|
277 |
+
"prob": datasets.Value("float"),
|
278 |
+
},
|
279 |
+
"annotations": datasets.Sequence(datasets.Value("string")),
|
280 |
+
"line_identifications": [
|
281 |
+
{
|
282 |
+
"label": datasets.Value("string"),
|
283 |
+
"prob": datasets.Value("float"),
|
284 |
+
}
|
285 |
+
],
|
286 |
+
},
|
287 |
+
}
|
288 |
+
)
|
289 |
+
elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}":
|
290 |
+
features = schemas.ssp_features
|
291 |
+
|
292 |
+
return datasets.DatasetInfo(
|
293 |
+
description=_DESCRIPTION,
|
294 |
+
features=features,
|
295 |
+
homepage=_HOMEPAGE,
|
296 |
+
license=_LICENSE,
|
297 |
+
citation=_CITATION,
|
298 |
+
)
|
299 |
+
|
300 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
|
301 |
+
"""Returns SplitGenerators."""
|
302 |
+
base_path = _BASE_URL.format(lang=self.config.name.split("_")[2])
|
303 |
+
|
304 |
+
checksum_url = urljoin(base_path, "checksum.sha256")
|
305 |
+
checksum_path = Path(dl_manager.download(checksum_url))
|
306 |
+
with open(checksum_path, encoding="utf-8") as f:
|
307 |
+
filenames = [line.split()[1] for line in f if line]
|
308 |
+
filenames = sorted(filenames, key=lambda x: int(re.search(r"\d+", x).group()) if re.search(r"\d+", x) else x)
|
309 |
+
data_urls = [urljoin(base_path, filename) for filename in filenames]
|
310 |
+
|
311 |
+
data_paths = list(map(Path, dl_manager.download([url for url in data_urls if url.endswith(".jsonl.gz")])))
|
312 |
+
|
313 |
+
return [
|
314 |
+
datasets.SplitGenerator(
|
315 |
+
name=datasets.Split.TRAIN,
|
316 |
+
gen_kwargs={
|
317 |
+
"filepaths": data_paths,
|
318 |
+
"split": "train",
|
319 |
+
},
|
320 |
+
)
|
321 |
+
]
|
322 |
+
|
323 |
+
def _generate_examples(self, filepaths: [Path], split: str) -> Tuple[int, Dict]:
|
324 |
+
"""Yields examples as (key, example) tuples."""
|
325 |
+
key = 0
|
326 |
+
for filepath in filepaths:
|
327 |
+
with gzip.open(open(filepath, "rb"), "rt", encoding="utf-8") as f:
|
328 |
+
for line in f:
|
329 |
+
doc = json.loads(line)
|
330 |
+
if self.config.schema == "source":
|
331 |
+
meta = dict()
|
332 |
+
meta["warc_headers"] = doc["warc_headers"]
|
333 |
+
meta["warc_headers"]["warc-identified-content-language"] = doc["warc_headers"].get("warc-identified-content-language")
|
334 |
+
meta["identification"] = doc["metadata"]["identification"]
|
335 |
+
meta["annotations"] = doc["metadata"]["annotation"]
|
336 |
+
meta["line_identifications"] = doc["metadata"]["sentence_identifications"]
|
337 |
+
yield key, {"id": key, "text": doc["content"], "meta": meta}
|
338 |
+
elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}":
|
339 |
+
yield key, {"id": str(key), "text": doc["content"]}
|
340 |
+
key += 1
|