Datasets:
Commit
•
fa08c6b
1
Parent(s):
fc308e7
Update dataset card
Browse files
README.md
CHANGED
@@ -138,7 +138,7 @@ task_ids:
|
|
138 |
- language-modeling
|
139 |
- masked-language-modeling
|
140 |
paperswithcode_id: cc100
|
141 |
-
pretty_name:
|
142 |
dataset_info:
|
143 |
- config_name: am
|
144 |
features:
|
@@ -181,7 +181,7 @@ config_names:
|
|
181 |
- sr
|
182 |
---
|
183 |
|
184 |
-
# Dataset Card for
|
185 |
|
186 |
## Table of Contents
|
187 |
- [Dataset Description](#dataset-description)
|
@@ -210,8 +210,11 @@ config_names:
|
|
210 |
## Dataset Description
|
211 |
|
212 |
- **Homepage:** https://data.statmt.org/cc-100/
|
213 |
-
- **Repository:**
|
214 |
-
- **Paper:** https://
|
|
|
|
|
|
|
215 |
- **Leaderboard:** [More Information Needed]
|
216 |
- **Point of Contact:** [More Information Needed]
|
217 |
|
@@ -221,7 +224,7 @@ This corpus is an attempt to recreate the dataset used for training XLM-R. This
|
|
221 |
|
222 |
### Supported Tasks and Leaderboards
|
223 |
|
224 |
-
CC-100 is mainly
|
225 |
|
226 |
### Languages
|
227 |
|
@@ -319,6 +322,8 @@ Statistical Machine Translation at the University of Edinburgh makes no claims o
|
|
319 |
|
320 |
### Citation Information
|
321 |
|
|
|
|
|
322 |
```bibtex
|
323 |
@inproceedings{conneau-etal-2020-unsupervised,
|
324 |
title = "Unsupervised Cross-lingual Representation Learning at Scale",
|
@@ -332,12 +337,16 @@ Statistical Machine Translation at the University of Edinburgh makes no claims o
|
|
332 |
Ott, Myle and
|
333 |
Zettlemoyer, Luke and
|
334 |
Stoyanov, Veselin",
|
|
|
|
|
|
|
|
|
335 |
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
336 |
month = jul,
|
337 |
year = "2020",
|
338 |
address = "Online",
|
339 |
publisher = "Association for Computational Linguistics",
|
340 |
-
url = "https://
|
341 |
doi = "10.18653/v1/2020.acl-main.747",
|
342 |
pages = "8440--8451",
|
343 |
abstract = "This paper shows that pretraining multilingual language models at scale leads to significant performance gains for a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +14.6{\%} average accuracy on XNLI, +13{\%} average F1 score on MLQA, and +2.4{\%} F1 score on NER. XLM-R performs particularly well on low-resource languages, improving 15.7{\%} in XNLI accuracy for Swahili and 11.4{\%} for Urdu over previous XLM models. We also present a detailed empirical analysis of the key factors that are required to achieve these gains, including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling without sacrificing per-language performance; XLM-R is very competitive with strong monolingual models on the GLUE and XNLI benchmarks. We will make our code and models publicly available.",
|
@@ -354,12 +363,26 @@ Statistical Machine Translation at the University of Edinburgh makes no claims o
|
|
354 |
Guzm{\'a}n, Francisco and
|
355 |
Joulin, Armand and
|
356 |
Grave, Edouard",
|
357 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
358 |
month = may,
|
359 |
year = "2020",
|
360 |
address = "Marseille, France",
|
361 |
publisher = "European Language Resources Association",
|
362 |
-
url = "https://
|
363 |
pages = "4003--4012",
|
364 |
abstract = "Pre-training text representations have led to significant improvements in many areas of natural language processing. The quality of these models benefits greatly from the size of the pretraining corpora as long as its quality is preserved. In this paper, we describe an automatic pipeline to extract massive high-quality monolingual datasets from Common Crawl for a variety of languages. Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that deduplicates documents and identifies their language. We augment this pipeline with a filtering step to select documents that are close to high quality corpora like Wikipedia.",
|
365 |
language = "English",
|
|
|
138 |
- language-modeling
|
139 |
- masked-language-modeling
|
140 |
paperswithcode_id: cc100
|
141 |
+
pretty_name: CC-100
|
142 |
dataset_info:
|
143 |
- config_name: am
|
144 |
features:
|
|
|
181 |
- sr
|
182 |
---
|
183 |
|
184 |
+
# Dataset Card for CC-100
|
185 |
|
186 |
## Table of Contents
|
187 |
- [Dataset Description](#dataset-description)
|
|
|
210 |
## Dataset Description
|
211 |
|
212 |
- **Homepage:** https://data.statmt.org/cc-100/
|
213 |
+
- **Repository:** [More Information Needed]
|
214 |
+
- **Paper:** https://aclanthology.org/2020.acl-main.747/
|
215 |
+
- **Paper:** https://aclanthology.org/2020.lrec-1.494/
|
216 |
+
- **Paper:** https://arxiv.org/abs/1911.02116
|
217 |
+
- **Paper:** https://arxiv.org/abs/1911.00359
|
218 |
- **Leaderboard:** [More Information Needed]
|
219 |
- **Point of Contact:** [More Information Needed]
|
220 |
|
|
|
224 |
|
225 |
### Supported Tasks and Leaderboards
|
226 |
|
227 |
+
CC-100 is mainly intended to pretrain language models and word representations.
|
228 |
|
229 |
### Languages
|
230 |
|
|
|
322 |
|
323 |
### Citation Information
|
324 |
|
325 |
+
Please cite the following if you found the resources in this corpus useful:
|
326 |
+
|
327 |
```bibtex
|
328 |
@inproceedings{conneau-etal-2020-unsupervised,
|
329 |
title = "Unsupervised Cross-lingual Representation Learning at Scale",
|
|
|
337 |
Ott, Myle and
|
338 |
Zettlemoyer, Luke and
|
339 |
Stoyanov, Veselin",
|
340 |
+
editor = "Jurafsky, Dan and
|
341 |
+
Chai, Joyce and
|
342 |
+
Schluter, Natalie and
|
343 |
+
Tetreault, Joel",
|
344 |
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
345 |
month = jul,
|
346 |
year = "2020",
|
347 |
address = "Online",
|
348 |
publisher = "Association for Computational Linguistics",
|
349 |
+
url = "https://aclanthology.org/2020.acl-main.747",
|
350 |
doi = "10.18653/v1/2020.acl-main.747",
|
351 |
pages = "8440--8451",
|
352 |
abstract = "This paper shows that pretraining multilingual language models at scale leads to significant performance gains for a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +14.6{\%} average accuracy on XNLI, +13{\%} average F1 score on MLQA, and +2.4{\%} F1 score on NER. XLM-R performs particularly well on low-resource languages, improving 15.7{\%} in XNLI accuracy for Swahili and 11.4{\%} for Urdu over previous XLM models. We also present a detailed empirical analysis of the key factors that are required to achieve these gains, including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling without sacrificing per-language performance; XLM-R is very competitive with strong monolingual models on the GLUE and XNLI benchmarks. We will make our code and models publicly available.",
|
|
|
363 |
Guzm{\'a}n, Francisco and
|
364 |
Joulin, Armand and
|
365 |
Grave, Edouard",
|
366 |
+
editor = "Calzolari, Nicoletta and
|
367 |
+
B{\'e}chet, Fr{\'e}d{\'e}ric and
|
368 |
+
Blache, Philippe and
|
369 |
+
Choukri, Khalid and
|
370 |
+
Cieri, Christopher and
|
371 |
+
Declerck, Thierry and
|
372 |
+
Goggi, Sara and
|
373 |
+
Isahara, Hitoshi and
|
374 |
+
Maegaard, Bente and
|
375 |
+
Mariani, Joseph and
|
376 |
+
Mazo, H{\'e}l{\`e}ne and
|
377 |
+
Moreno, Asuncion and
|
378 |
+
Odijk, Jan and
|
379 |
+
Piperidis, Stelios",
|
380 |
+
booktitle = "Proceedings of the Twelfth Language Resources and Evaluation Conference",
|
381 |
month = may,
|
382 |
year = "2020",
|
383 |
address = "Marseille, France",
|
384 |
publisher = "European Language Resources Association",
|
385 |
+
url = "https://aclanthology.org/2020.lrec-1.494",
|
386 |
pages = "4003--4012",
|
387 |
abstract = "Pre-training text representations have led to significant improvements in many areas of natural language processing. The quality of these models benefits greatly from the size of the pretraining corpora as long as its quality is preserved. In this paper, we describe an automatic pipeline to extract massive high-quality monolingual datasets from Common Crawl for a variety of languages. Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that deduplicates documents and identifies their language. We augment this pipeline with a filtering step to select documents that are close to high quality corpora like Wikipedia.",
|
388 |
language = "English",
|
cc100.py
CHANGED
@@ -29,20 +29,23 @@ _CITATION = """\
|
|
29 |
Goyal, Naman and
|
30 |
Chaudhary, Vishrav and
|
31 |
Wenzek, Guillaume and
|
32 |
-
Guzm{'a}n, Francisco and
|
33 |
Grave, Edouard and
|
34 |
Ott, Myle and
|
35 |
Zettlemoyer, Luke and
|
36 |
Stoyanov, Veselin",
|
|
|
|
|
|
|
|
|
37 |
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
38 |
month = jul,
|
39 |
year = "2020",
|
40 |
address = "Online",
|
41 |
publisher = "Association for Computational Linguistics",
|
42 |
-
url = "https://
|
43 |
doi = "10.18653/v1/2020.acl-main.747",
|
44 |
pages = "8440--8451",
|
45 |
-
abstract = "This paper shows that pretraining multilingual language models at scale leads to significant performance gains for a wide range of cross-lingual transfer tasks. We train a Transformer-based masked language model on one hundred languages, using more than two terabytes of filtered CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms multilingual BERT (mBERT) on a variety of cross-lingual benchmarks, including +14.6{%} average accuracy on XNLI, +13{%} average F1 score on MLQA, and +2.4{%} F1 score on NER. XLM-R performs particularly well on low-resource languages, improving 15.7{%} in XNLI accuracy for Swahili and 11.4{%} for Urdu over previous XLM models. We also present a detailed empirical analysis of the key factors that are required to achieve these gains, including the trade-offs between (1) positive transfer and capacity dilution and (2) the performance of high and low resource languages at scale. Finally, we show, for the first time, the possibility of multilingual modeling without sacrificing per-language performance; XLM-R is very competitive with strong monolingual models on the GLUE and XNLI benchmarks. We will make our code and models publicly available.",
|
46 |
}
|
47 |
@inproceedings{wenzek-etal-2020-ccnet,
|
48 |
title = "{CCN}et: Extracting High Quality Monolingual Datasets from Web Crawl Data",
|
@@ -50,17 +53,30 @@ _CITATION = """\
|
|
50 |
Lachaux, Marie-Anne and
|
51 |
Conneau, Alexis and
|
52 |
Chaudhary, Vishrav and
|
53 |
-
Guzm{'a}n, Francisco and
|
54 |
Joulin, Armand and
|
55 |
Grave, Edouard",
|
56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
month = may,
|
58 |
year = "2020",
|
59 |
address = "Marseille, France",
|
60 |
publisher = "European Language Resources Association",
|
61 |
-
url = "https://
|
62 |
pages = "4003--4012",
|
63 |
-
abstract = "Pre-training text representations have led to significant improvements in many areas of natural language processing. The quality of these models benefits greatly from the size of the pretraining corpora as long as its quality is preserved. In this paper, we describe an automatic pipeline to extract massive high-quality monolingual datasets from Common Crawl for a variety of languages. Our pipeline follows the data processing introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that deduplicates documents and identifies their language. We augment this pipeline with a filtering step to select documents that are close to high quality corpora like Wikipedia.",
|
64 |
language = "English",
|
65 |
ISBN = "979-10-95546-34-4",
|
66 |
}
|
|
|
29 |
Goyal, Naman and
|
30 |
Chaudhary, Vishrav and
|
31 |
Wenzek, Guillaume and
|
32 |
+
Guzm{\\'a}n, Francisco and
|
33 |
Grave, Edouard and
|
34 |
Ott, Myle and
|
35 |
Zettlemoyer, Luke and
|
36 |
Stoyanov, Veselin",
|
37 |
+
editor = "Jurafsky, Dan and
|
38 |
+
Chai, Joyce and
|
39 |
+
Schluter, Natalie and
|
40 |
+
Tetreault, Joel",
|
41 |
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
42 |
month = jul,
|
43 |
year = "2020",
|
44 |
address = "Online",
|
45 |
publisher = "Association for Computational Linguistics",
|
46 |
+
url = "https://aclanthology.org/2020.acl-main.747",
|
47 |
doi = "10.18653/v1/2020.acl-main.747",
|
48 |
pages = "8440--8451",
|
|
|
49 |
}
|
50 |
@inproceedings{wenzek-etal-2020-ccnet,
|
51 |
title = "{CCN}et: Extracting High Quality Monolingual Datasets from Web Crawl Data",
|
|
|
53 |
Lachaux, Marie-Anne and
|
54 |
Conneau, Alexis and
|
55 |
Chaudhary, Vishrav and
|
56 |
+
Guzm{\\'a}n, Francisco and
|
57 |
Joulin, Armand and
|
58 |
Grave, Edouard",
|
59 |
+
editor = "Calzolari, Nicoletta and
|
60 |
+
B{\\'e}chet, Fr{\\'e}d{\\'e}ric and
|
61 |
+
Blache, Philippe and
|
62 |
+
Choukri, Khalid and
|
63 |
+
Cieri, Christopher and
|
64 |
+
Declerck, Thierry and
|
65 |
+
Goggi, Sara and
|
66 |
+
Isahara, Hitoshi and
|
67 |
+
Maegaard, Bente and
|
68 |
+
Mariani, Joseph and
|
69 |
+
Mazo, H{\\'e}l{\\`e}ne and
|
70 |
+
Moreno, Asuncion and
|
71 |
+
Odijk, Jan and
|
72 |
+
Piperidis, Stelios",
|
73 |
+
booktitle = "Proceedings of the Twelfth Language Resources and Evaluation Conference",
|
74 |
month = may,
|
75 |
year = "2020",
|
76 |
address = "Marseille, France",
|
77 |
publisher = "European Language Resources Association",
|
78 |
+
url = "https://aclanthology.org/2020.lrec-1.494",
|
79 |
pages = "4003--4012",
|
|
|
80 |
language = "English",
|
81 |
ISBN = "979-10-95546-34-4",
|
82 |
}
|