holylovenia
commited on
Commit
·
2f3297a
1
Parent(s):
13cc1a9
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
tags:
|
4 |
+
- self-supervised-pretraining
|
5 |
+
language:
|
6 |
+
- ind
|
7 |
+
- jav
|
8 |
+
- sun
|
9 |
+
---
|
10 |
+
|
11 |
+
# cc100
|
12 |
+
|
13 |
+
This corpus is an attempt to recreate the dataset used for training
|
14 |
+
|
15 |
+
XLM-R. This corpus comprises of monolingual data for 100+ languages and
|
16 |
+
|
17 |
+
also includes data for romanized languages (indicated by *_rom). This
|
18 |
+
|
19 |
+
was constructed using the urls and paragraph indices provided by the
|
20 |
+
|
21 |
+
CC-Net repository by processing January-December 2018 Commoncrawl
|
22 |
+
|
23 |
+
snapshots. Each file comprises of documents separated by
|
24 |
+
|
25 |
+
double-newlines and paragraphs within the same document separated by a
|
26 |
+
|
27 |
+
newline. The data is generated using the open source CC-Net repository.
|
28 |
+
|
29 |
+
No claims of intellectual property are made on the work of preparation
|
30 |
+
|
31 |
+
of the corpus.
|
32 |
+
|
33 |
+
## Dataset Usage
|
34 |
+
|
35 |
+
Run `pip install nusacrowd` before loading the dataset through HuggingFace's `load_dataset`.
|
36 |
+
|
37 |
+
## Citation
|
38 |
+
|
39 |
+
```
|
40 |
+
@inproceedings{conneau-etal-2020-unsupervised,
|
41 |
+
title = "Unsupervised Cross-lingual Representation Learning at Scale",
|
42 |
+
author = "Conneau, Alexis and
|
43 |
+
Khandelwal, Kartikay and
|
44 |
+
Goyal, Naman and
|
45 |
+
Chaudhary, Vishrav and
|
46 |
+
Wenzek, Guillaume and
|
47 |
+
Guzm{'a}n, Francisco and
|
48 |
+
Grave, Edouard and
|
49 |
+
Ott, Myle and
|
50 |
+
Zettlemoyer, Luke and
|
51 |
+
Stoyanov, Veselin",
|
52 |
+
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
53 |
+
month = jul,
|
54 |
+
year = "2020",
|
55 |
+
address = "Online",
|
56 |
+
publisher = "Association for Computational Linguistics",
|
57 |
+
url = "https://www.aclweb.org/anthology/2020.acl-main.747",
|
58 |
+
doi = "10.18653/v1/2020.acl-main.747",
|
59 |
+
pages = "8440--8451",
|
60 |
+
abstract = "This paper shows that pretraining multilingual language models
|
61 |
+
at scale leads to significant performance gains for a wide range of
|
62 |
+
cross-lingual transfer tasks. We train a Transformer-based masked language
|
63 |
+
model on one hundred languages, using more than two terabytes of filtered
|
64 |
+
CommonCrawl data. Our model, dubbed XLM-R, significantly outperforms
|
65 |
+
multilingual BERT (mBERT) on a variety of cross-lingual benchmarks,
|
66 |
+
including +14.6{%} average accuracy on XNLI, +13{%} average F1 score on
|
67 |
+
MLQA, and +2.4{%} F1 score on NER. XLM-R performs particularly well on
|
68 |
+
low-resource languages, improving 15.7{%} in XNLI accuracy for Swahili and
|
69 |
+
11.4{%} for Urdu over previous XLM models. We also present a detailed
|
70 |
+
empirical analysis of the key factors that are required to achieve these
|
71 |
+
gains, including the trade-offs between (1) positive transfer and capacity
|
72 |
+
dilution and (2) the performance of high and low resource languages at
|
73 |
+
scale. Finally, we show, for the first time, the possibility of
|
74 |
+
multilingual modeling without sacrificing per-language performance; XLM-R
|
75 |
+
is very competitive with strong monolingual models on the GLUE and XNLI
|
76 |
+
benchmarks. We will make our code and models publicly available.",
|
77 |
+
}
|
78 |
+
|
79 |
+
@inproceedings{wenzek-etal-2020-ccnet,
|
80 |
+
title = "{CCN}et: Extracting High Quality Monolingual Datasets from Web Crawl Data",
|
81 |
+
author = "Wenzek, Guillaume and
|
82 |
+
Lachaux, Marie-Anne and
|
83 |
+
Conneau, Alexis and
|
84 |
+
Chaudhary, Vishrav and
|
85 |
+
Guzm{'a}n, Francisco and
|
86 |
+
Joulin, Armand and
|
87 |
+
Grave, Edouard",
|
88 |
+
booktitle = "Proceedings of the 12th Language Resources and Evaluation Conference",
|
89 |
+
month = may,
|
90 |
+
year = "2020",
|
91 |
+
address = "Marseille, France",
|
92 |
+
publisher = "European Language Resources Association",
|
93 |
+
url = "https://www.aclweb.org/anthology/2020.lrec-1.494",
|
94 |
+
pages = "4003--4012",
|
95 |
+
abstract = "Pre-training text representations have led to significant
|
96 |
+
improvements in many areas of natural language processing. The quality of
|
97 |
+
these models benefits greatly from the size of the pretraining corpora as
|
98 |
+
long as its quality is preserved. In this paper, we describe an automatic
|
99 |
+
pipeline to extract massive high-quality monolingual datasets from Common
|
100 |
+
Crawl for a variety of languages. Our pipeline follows the data processing
|
101 |
+
introduced in fastText (Mikolov et al., 2017; Grave et al., 2018), that
|
102 |
+
deduplicates documents and identifies their language. We augment this
|
103 |
+
pipeline with a filtering step to select documents that are close to high
|
104 |
+
quality corpora like Wikipedia.",
|
105 |
+
language = "English",
|
106 |
+
ISBN = "979-10-95546-34-4",
|
107 |
+
}
|
108 |
+
```
|
109 |
+
|
110 |
+
## License
|
111 |
+
|
112 |
+
MIT
|
113 |
+
|
114 |
+
## Homepage
|
115 |
+
|
116 |
+
[https://data.statmt.org/cc-100/](https://data.statmt.org/cc-100/)
|
117 |
+
|
118 |
+
### NusaCatalogue
|
119 |
+
|
120 |
+
For easy indexing and metadata: [https://indonlp.github.io/nusa-catalogue](https://indonlp.github.io/nusa-catalogue)
|