Dataset Card for PolyNews

Dataset Summary

PolyNews is a multilingual dataset containing news titles in 77 languages and 19 scripts.

Uses

This dataset can be used for domain adaptation of language models, language modeling or text generation.

Languages

There are 77 languages available:

Code	Language	Script	#Articles (K)
amh_Ethi	Amharic	Ethiopic	0.551
arb_Arab	Modern Standard Arabic	Arabic	10.882
ayr_Latn	Central Aymara	Latin	12.878
bam_Latn	Bambara	Latin	2.916
bbj_Latn	Ghomálá’	Latin	1.737
ben_Beng	Bengali	Bengali	2.268
bos_Latn	Bosnian	Latin	0.298
bul_Cyrl	Bulgarian	Cyrillic	1.791
cat_Latn	Catalan	Latin	30.410
ces_Latn	Czech	Latin	58.382
ckb_Arab	Central Kurdish	Arabic	0.014
dan_Latn	Danish	Latin	9.456
deu_Latn	German	Latin	145.484
ell_Grek	Greek	Greek	50.176
eng_Latn	English	Latin	981.430
est_Latn	Estonian	Latin	3.942
ewe_Latn	Éwé	Latin	2.003
fil_Latn	Filipino	Latin	3.3132
fin_Latn	Finnish	Latin	19.602
fon_Latn	Fon	Latin	2.610
fra_Latn	French	Latin	481.117
guj_Gujr	Gujarati	Gujarati	0.690
guw_Latn	Gun	Latin	1.068
hau_Latn	Hausa	Latin	7.898
heb_Hebr	Hebrew	Hebrew	0.355
hin_Deva	Hindi	Devanagari	0.707
hun_Latn	Hungarian	Latin	22.219
ibo_Latn	Igbo	Latin	7.709
ind_Latn	Indonesian	Latin	17.749
ita_Latn	Italian	Latin	163.396
jpn_Jpan	Japanese	Japanese	20.778
kaz_Cyrl	Kazakh	Cyrillic	0.763
khm_Khmr	Khmer	Khmer	0.227
kor_Hang	Korean	Hangul	3.527
lav_Latn	Latvian	Latin	3.971
lin_Latn	Lingala	Latin	0.602
lit_Latn	Lithuanian	Latin	3.948
lug_Latn	Ganda	Latin	4.769
luo_Latn	Luo	Latin	4.250
mkd_Cyrl	Macedonian	Cyrillic	10.537
mos_Latn	Mossi	Latin	2.458
mya_Mymr	Burmese	Myanmar	0.583
nld_Latn	Dutch	Latin	53.184
nor_Latn	Norwegian	Latin	0.529
npi_Deva	Nepali	Devanagari	0.220
orm_Latn	Oromo	Latin	1.124
ory_Orya	Odia	Oriya	0.038
pan_Guru	Eastern Panjabi	Gurmukhi	0.336
pcm_Latn	Nigerian Pidgin	Latin	5.742
pes_Arab	Western Persian	Arabic	1.431
plt_Latn	Malagasy	Latin	393.767
pol_Latn	Polish	Latin	80.960
por_Latn	Portuguese	Latin	156.039
ron_Latn	Romanian	Latin	10.472
run_Latn	Rundi	Latin	1.113
rus_Cyrl	Russian	Cyrillic	143.283
sna_Latn	Shona	Latin	1.128
som_Latn	Somali	Latin	1.019
spa_Latn	Spanish	Latin	681.121
sqi_Latn	Albanian	Latin	7.274
srp_Cyrl	Serbian	Cyrillic	1.056
srp_Latn	Serbian	Latin	58.012
swe_Latn	Swedish	Latin	12.323
swh_Latn	Swahili	Latin	47.337
tam_Taml	Tamil	Tamil	0.358
tet_Latn	Tetun	Latin	0.626
tha_Thai	Thai	Thai	0.091
tir_Ethi	Tigrinya	Ethiopic	0.079
tsn_Latn	Tswana	Latin	2.075
tur_Latn	Turkish	Latin	19.793
twi_Latn	Twi	Latin	3.012
ukr_Cyrl	Ukrainian	Cyrillic	0.292
urd_Arab	Urdu	Arabic	0.804
wol_Latn	Wolof	Latin	3.344
xho_Latn	Xhosa	Latin	0.709
yor_Latn	Yorùbá	Latin	8.011
zho_Hans	Chinese	Han (Simplified)	59.771
zho_Hant	Chinese	Han (Traditional)	54.561
zul_Latn	Zulu	Latin	3.376

Dataset Structure

Data Instances

>>> from datasets import load_dataset
>>> data = load_dataset('aiana94/polynews', 'ron_Latn')

# Please, specify the language code,

# A data point example is below:

{
"text": "Un public numeros. Este uimitor succesul după doar trei ediții . ",
"provenance": "globalvoices"
}

Data Fields

• text (string): news text
• provenance (string) : source dataset for the news example

Data Splits

For all languages, there is only the train split.

Dataset Creation

Curation Rationale

Multiple multilingual, human-translated, datasets containing news texts have been released in recent years. However, these datasets are stored in different formats and various websites, and many contain numerous near duplicates. With PolyNews, we aim to provide an easily-accessible, unified and deduplicated dataset that combines these disparate data sources. It can be used for domain adaptation of language models, language modeling or text generation in both high-resource and low-resource languages.

Source Data

The source data consists of five multilingual news datasets.

• Wikinews (latest dump available in May 2024)
• GlobalVoices (v2018q4)
• WMT-News (v2019)
• MasakhaNews (train split)
• MAFAND (train split)

Data Collection and Processing

We processed the data using a working script which covers the entire processing pipeline. It can be found here.

The data processing pipeline consists of:

1. Downloading the WMT-News and GlobalVoices News from OPUS.
2. Downloading the latest dump from WikiNews.
3. Loading the MasakhaNews and MAFAND datasets from Hugging Face Hub (only the train splits).
4. Concatenating, per language, all news texts from the source datasets.
5. Data cleaning (e.g., removal of exact duplicates, short texts, texts in other scripts)
6. MinHash near-deduplication per language.

Annotations

We augment the original samples with the provenance annotation which specifies the original data source from which a particular examples stems.

Personal and Sensitive Information

The data is sourced from newspaper sources and contains mentions of public figures and individuals.

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

Users should keep in mind that the dataset contains short news texts (e.g., mostly titles), which might limit the applicability of the developed systems to other domains.

Additional Information

Licensing Information

The dataset is released under the CC BY-NC Attribution-NonCommercial 4.0 International license.

Citation Infomation

BibTeX:

@misc{iana2024news,
      title={News Without Borders: Domain Adaptation of Multilingual Sentence Embeddings for Cross-lingual News Recommendation}, 
      author={Andreea Iana and Fabian David Schmidt and Goran Glavaš and Heiko Paulheim},
      year={2024},
      eprint={2406.12634},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2406.12634}
}