Dataset Card for "open-lid-dataset"

Dataset Summary

The OpenLID dataset covers 201 languages and is designed for training language identification models. The majority of the source datasets were derived from news sites, Wikipedia, or religious text, though some come from other domains (e.g. transcribed conversations, literature, or social media). A sample of each language in each source was manually audited to check it was in the attested language (see the paper) for full details.

Supported tasks

This dataset is intended for training high-coverage language identification models (e.g. OpenLID). It is compatible with the FLORES-200 evaluation benchmark.

Languages

There are 201 languages included in the dataset with varying amounts of data: the largest class (English) contains 7.5 million lines of data, and the smallest (South Azerbaijani) contains 532 lines of data. The mean number of lines per language is 602,812. A full breakdown of lines of data per language is available on the repo.

Dataset Structure

Data Instances

Each entry in the dataset consists of a line of data, a language label included script information, and a tag indicating the source.

{
  "text": "¿Serás exaltada hasta el cielo?",
  "language": "spa_Latn",
  "dataset_source": "lti" 
}

Data Splits

Only a train split is provided. The dataset is designed to be compatible with the FLORES-200 evaluation benchmark.

Dataset Creation

Curation Rationale

Recent work has found that existing language identification algorithms perform poorly in practice compared to test performance. The problem is particularly acute for low-resource languages: Kreutzer et al. (2022) found a positive Spearman rank correlation between quality of data and size of language for all of the \ac{lid}-filtered multilingual datasets they studied. In addition, for a significant fraction of the language corpora they studied, less than half of the sentences were in the correct language. They point out that such low-quality data not only leads to poor performance in downstream tasks, but that it also contributes to `representation washing', where the community is given a false view of the actual progress of low-resource natural language processing.

There are several open language identification models offering quick classification and high language coverage (e.g. CLD3, No Language Left Behind). However, to the best of our knowledge, none of the commonly-used scalable language identificaiton systems make their training data public.

This dataset aims to address that gap by curating and combining sources of open training data for language identification and by auditing a sample of all languages in each source to check reliability.

Source Data

The majority of the source datasets were derived from news sites, Wikipedia, or religious text, though some come from other domains (e.g. transcribed conversations, literature, or social media). We provide a full list at the end of this model card along with the licensing information for each source.

Initial Data Collection and Normalisation

Our initial aim was to cover the same languages present in the FLORES-200 Evaluation Benchmark so that we could use this dataset for evaluation. However, during the curation process, we decided to exclude three languages. Firstly, though Akan and Twi are both included as separate languages in FLORES-200, Akan is actually a macrolanguage covering a language continuum which includes Twi. Given the other languages in FLORES-200 are individual languages, we decided to exclude Akan. Secondly, FLORES-200 includes Modern Standard Arabic (MSA) written in Latin script. It is true that Arabic dialects are often written in Latin characters in informal situations (e.g. social media). However, MSA is a form of standardised Arabic which is not usually used in informal situations. Since we could not any find naturally-occurring training data, we excluded MSA from the dataset. Finally, we excluded Minangkabau in Arabic script because it is now rarely written this way, making it difficult to find useful training data.

The first step in our manual audit was to check and standardise language labels, as these are often inconsistent or idiosyncratic. We chose to copy the language codes in FLORES-200 and reassign macrolanguage or ambiguous language codes in the data sources we found to the dominant individual language. Whilst this resulted in more useful data for some languages, for other languages we had to be more conservative. For example, we originally reassigned text labelled as the macrolanguage Malay (msa_Latn) to Standard Malay, but this led to a large drop in performance as the former covers a very diverse set of languages.

Two of the authors then carried out a manual audit of a random sample of all data sources and languages: one a native Bulgarian speaker (able to read Cyrillic and Latin scripts and Chinese characters), and the other a native English speaker (able to read Latin, Arabic and Hebrew scripts). For languages we knew, we checked the language was what we expected. For unfamiliar languages in a script we could read, we compared the sample to the Universal Declaration of Human Rights or failing that, to a sample of text on Wikipedia. We compared features of the text which are common in previous language identification algorithms and could be identified easily by humans: similar diacritics, word lengths, common words, loan words matching the right cultural background, similar suffixes and prefixes, and vowel/consonant patterns. For scripts we could not read, we checked that all lines of the sample matched the script in the Universal Declaration of Human Rights.

We kept preprocessing minimal so that the process was as language agnostic as possible. We used the scripts provided with Moses to remove non-printing characters and detokenise the data where necessary. We then filtered the data so that each line contained at least one character in the expected script (as defined by Perl) to allow for borrowings. Finally, we sampled proportionally to $ p_l^{0.3} $, where $ p_l $ is the fraction of lines in the dataset which are in language $ l $. This aims to ameliorate class skew issues.

Considerations for Using the Data

Social Impact of Dataset

This dataset covers a number of low-resourced languages. This makes it a potentially useful resource, but due to the limited amount of data and domains, care must be taken not to overclaim performance or coverage.

Discussion of Biases

Our work aims to broaden natural language processing coverage by allowing practitioners to identify relevant data in more languages. However, we note that language identification is inherently a normative activity that risks excluding minority dialects, scripts, or entire microlanguages from a macrolanguage. Choosing which languages to cover may reinforce power imbalances, as only some groups gain access to language processing technologies.

In addition, errors in language identification can have a significant impact on downstream performance, particularly (as is often the case) when a system is used as a `black box'. The performance of our classifier is not equal across languages which could lead to worse downstream performance for particular groups. We mitigate this by providing metrics by class.

Additional information

The dataset was curated from the sources listed below by Laurie Burchell and Nikolay Bogoychev.

Licensing Information

License considerations for each source are given below. Open use for non-commercial purposes is covered by all licences.

If you view any part of this dataset as a violation of intellectual property rights, please let us know and we will remove it.

Source	Description	License
Arabic Dialects Dataset	Dataset of Arabic dialects for Gulf, Egyptian, Levantine, and Tunisian Arabic dialects plus MSA	No explicit license; website describes data as "some free and useful Arabic corpora that I have created for researchers working on Arabic Natural Language Processing, Corpus and Computational Linguistics."
BLTR	Monolingual Bhojpuri corpus	CC BY-NC-SA 4.0
Global Voices	A parallel corpus of news stories from the web site Global Voices	The website for Global Voices is licensed as Creative Commons Attribution 3.0. There is no explicit additional license accompanying the dataset.
Guaraní Parallel Set	Parallel Guaraní-Spanish news corpus sourced from Paraguyan websites	No explicit license
HKCanCor	Transcribed conversations in Hong Kong Cantonese	CC BY 4.0
IADD	Arabic dialect identification dataset covering 5 regions (Maghrebi, Levantine, Egypt, Iraq, and Gulf) and 9 countries (Algeria, Morocco, Tunisia, Palestine, Jordan, Syria, Lebanon, Egypt and Iraq). It is created from five corpora: DART, SHAMI, TSAC, PADIC, and AOC.	Multiple licenses: Apache License 2.0 (SHAMI); GNU Lesser General Public License v3.0 (TSAC); GNU General Public License v3 (PADIC). DART and AOC had no explicit license.
Leipzig Corpora Collection	A collection of corpora in different languages with an identical format.	The Terms of Usage states "Permission for use is granted free of charge solely for non-commercial personal and scientific purposes licensed under the Creative Commons License CC BY-NC."
LTI	Training data for language identification	From the README: "With the exception of the contents of the Europarl/, ProjectGutenberg/, and PublicDomain/ directories, all code and text in this corpus are copyrighted. However, they may be redistributed under the terms of various Creative Commons licenses and the GNU GPL. Copying the unmodified archive noncommercially is permitted by all of the licenses. For commercial redistribution or redistribution of modified versions, please consult the individual licenses."
MADAR Shared Task 2019, subtask 1	Dialectal Arabic in the travel domain	The MADAR Corpus has a custom license, the text of which can be found in this repo.
EM corpus	Parallel Manipuri-English sentences crawled from The Sangai Express	CC BY-NC 4.0
MIZAN	Parallel Persian-English corpus from literature domain	CC BY 4.0
MT560 v1	A machine translation dataset for over 500 languages to English. We have filtered out data from OPUS-100, Europarl, Open Subtitles, Paracrawl, Wikimedia, Wikimatrix, Wikititles, and Common Crawl due to issues with the fidelity of the language labels.	Apache License 2.0
NLLB Seed	Around 6000 sentences in 39 languages sampled from Wikipedia, intended to cover languages lacking training data.	CC BY-SA 4.0
SETIMES	A parallel corpus of news articles in the Balkan languages	CC-BY-SA 3.0
Tatoeba	Collaborative sentence translations	CC BY 2.0 FR
Tehran English-Persian parallel corpus (TEP)	Parallel Persian-English sentences sourced from subtitles	GNU General Public License
Turkic Interlingua (TIL) Corpus	A large-scale parallel corpus combining most of the public datasets for 22 Turkic languages	CC BY-NC-SA 4.0
WiLI-2018	Wikipedia language identification benchmark containing 235K paragraphs of 235 languages	Open Data Commons Open Database License (ODbL) v1.0
XL-Sum	Summarisation dataset covering 44 languages, sourced from BBC News	CC BY-NC-SA 4.0

Citation Information

If you use this dataset, please cite all the authors in the citation file who compiled the source datasets, plus the OpenLID paper:

@inproceedings{burchell-etal-2023-open,
    title = "An Open Dataset and Model for Language Identification",
    author = "Burchell, Laurie  and
      Birch, Alexandra  and
      Bogoychev, Nikolay  and
      Heafield, Kenneth",
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-short.75",
    doi = "10.18653/v1/2023.acl-short.75",
    pages = "865--879",
    abstract = "Language identification (LID) is a fundamental step in many natural language processing pipelines. However, current LID systems are far from perfect, particularly on lower-resource languages. We present a LID model which achieves a macro-average F1 score of 0.93 and a false positive rate of 0.033{\%} across 201 languages, outperforming previous work. We achieve this by training on a curated dataset of monolingual data, which we audit manually to ensure reliability. We make both the model and the dataset available to the research community. Finally, we carry out detailed analysis into our model{'}s performance, both in comparison to existing open models and by language class.",
}

Contributions

Thanks to @hac541309 and @davanstrien for adding this dataset.