AfricaNLP-Public-Datasets

A repository for publicly/freely available Natural Language Processing (NLP) datasets for African languages.

Datasets per task (Randomly ordered)

Machine Translation

• JW300: A parallel text dataset of 417 languages, including 101 African languages.

• TANZIL: A translated Quran to 45 languages, including African languages such as Amharic, Hausa, Somali, and Swahili.

• MENYO-20k: A Yorùbá-English multi-domain parallel text dataset.

• FFR: A Fon-French parallel text dataset.

• Hausa Corpus: A Hausa-English parallel text dataset.

• CCAligned: A parallel text dataset for English and 137 languages, including 30 African Languages.

• ParaCrawl: A parallel text dataset for 41 languages, including Somali and Swahili.

• WikiMatrix: A parallel text dataset for 85 languages, including Swahili, Malagasy, and Egyptian Arabic.

• Ethiopian MT datasets: A parallel text dataset for English paired with 7 Ethiopian languages.

• English-Luganda: An English-Luganda parallel text dataset.

• French-Fon and French-Ewe: A parallel text dataset for French paired with Fon and Ewe.

• Amharic-English: An Amharic-English parallel text dataset.

• Tigrinya-English: A Tigrinya-English parallel text dataset (Free registration required).

• Lingala-French: A Lingala-English parallel text dataset (Free registration required).

• Congolese Swahili-French (Min,Small,Medium): Congolese Swahili-French parallel text datasets (Free registration required).

• Swahili-French: A synthetic Swahili-French parallel text dataset (Free registration required).

• English-Hausa (Min, Small): English-Hausa parallel text datasets (Free registration required).

• English-Swahili: An English-Swahili parallel text dataset (Free registration required).

• English-Kanuri: An English-Kanuri parallel text dataset (Free registration required).

• English-Akuapem Twi: An English-Akwapem Twi parallel text dataset.

• FLORES-101: A parallel text dataset for 101 languages, including 18 African languages.

Text Classification

• KINNEWS and KIRNEWS: News Classification datasets for Kinyarwanda (KINNEWS) and Kirundi (KIRNEWS).

• Setswana and Sepedi: News classification datasets for Setswana and Sepedi.

• Swahili News: A news classification dataset for Swahili.

• Amharic News Text classification: News text classification dataset for Amharic.

• VOA Hausa and BBC Yoruba news classification: News title classification dataset for Hausa and Yoruba.

Sentiment Analysis

• TUNIZI: A Tunizian Arabizi sentiment analysis dataset.

Text Summarization

• Amharic Summarization: A dataset for Amharic abstractive text summarization.

Named Entity Recognition

• MasakhaNER: A dataset for Named Entity Recognition of 10 African languages.

• WikiANN: A dataset for Named Entity Recognition for 282 languages, including several African languages.

• Yoruba GV NER: Yoruba Named Entity Recognition dataset.

• Hausa VOA NER: Hausa Named Entity Recognition dataset

Automated Speech Recognition (ASR)

• ALFFA: An ASR dataset for Amharic, Hausa, Swahili, and Wolof.

• AMMI ASR dataset: An ASR dataset for 19 Languages, including 16 African Languages.

• CommonVoice: An ongoing ASR dataset project for 60 languages (as of May, 2021), including Kinyarwanda, Kabyle, and Luganda.

• Fon: An ASR dataset for Fon.

• Swahili: A Swahili speech dataset (Free registration required).

• Congolese Swahili: A Congolese Swahili speech dataset (Free registration required).

Speech Translation

• Mboshi: Mboshi-French parallel speech dataset.

• IWSLT 2021 Speech Translation: Speech translation datasets for Swahili - English and Congolese Swahili-French.

Monolingual Data

• Swahili Language Modeling: A Swahili dataset for language modeling and additional datasets for Swahili Syllabic Alphabet and Swahili Word Analogy.

• OSCAR: A multilingual dataset for 166 languages, including Amharic, Somalia, Yoruba, Egyptian Arabic, Malagasy, Swahili, and Afrikaans.

Contributions

This is a growing list of NLP datasets for African languages. Please, if there is any publicly available dataset I missed out, kindly feel free to do a pull request or email me at [email protected] to add it.