---
language: 
- eng
- afr
- nbl
- xho
- zul
- nso
- sep
- tsn
- ssw
- ven
- tso
pretty_name: "The Vuk'uzenzele South African Multilingual Corpus"
tags:
- multilingual
- government
license: "cc-by-4.0"
task_categories:
- sentence-similarity
- translation
arxiv: 2303.03750
---

# The Vuk'uzenzele South African Multilingual Corpus

Github: [https://github.com/dsfsi/vukuzenzele-nlp/](https://github.com/dsfsi/vukuzenzele-nlp/)

Zenodo: [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.7598539.svg)](https://doi.org/10.5281/zenodo.7598539)

Arxiv Preprint: [![arXiv](https://img.shields.io/badge/arXiv-2303.03750-b31b1b.svg)](https://arxiv.org/abs/2303.03750)

# About
The dataset was obtained from the South African government magazine Vuk'uzenzele, created by the [Government Communication and Information System (GCIS)](https://www.gcis.gov.za/). 
The original raw PDFS were obtatined from the [Vuk'uzenzele website](https://www.vukuzenzele.gov.za/).

The datasets contain government magazine editions in 11 languages, namely:

|  Language  | Code  |  Language  | Code |
|------------|-------|------------|-------|
| English    | (eng) | Sepedi     | (sep) |
| Afrikaans  | (afr) | Setswana   | (tsn) |
| isiNdebele | (nbl) | Siswati    | (ssw) |
| isiXhosa   | (xho) | Tshivenda  | (ven) |
| isiZulu    | (zul) | Xitstonga  | (tso) |
| Sesotho    | (nso) |


# Datasets
The datasets consist of pairwise sentence aligned data. There are 55 distinct datasets of paired sentences.
The data is obtained by comparing [LASER](https://github.com/facebookresearch/LASER) embeddings of sentence tokens between 2 languages. If the similarity is high, the sentences are deemed semantic equivalents of one another and the observation is outputted. 

Naming convention:  
The naming structure of the pairwise_sentence_aligned folder is `aligned-{src_lang_code}-{tgt_lang_code}.csv`.  
For example, `aligned-afr-zul.csv` is the aligned sentences between Afrikaans and isiZulu. 

The data is in .csv format and the columns are `src_text`,`tgt_text`,`cosine_score` where:  
- `src_text` is the source sentence
- `tgt_text` is the target sentence
- `cosine_score` is the cosine similarity score obtained by comparing the sentence embeddings, it ranges from 0 to 1

**Note:** The notion of source (src) and target (tgt) are only necessary for distinction between the languages used in the aligned pair, as the sentence semantics should be bidirectional. (hallo <-> sawubona)

# Citation
Vukosi Marivate, Andani Madodonga, Daniel Njini, Richard Lastrucci, Isheanesu Dzingirai, Jenalea Rajab. **The Vuk'uzenzele South African Multilingual Corpus**, 2023

> @dataset{marivate_vukosi_2023_7598540,
  author       = {Marivate, Vukosi and
                  Njini, Daniel and
                  Madodonga, Andani and
                  Lastrucci, Richard and
                  Dzingirai, Isheanesu
                  Rajab, Jenalea},
  title        = {The Vuk'uzenzele South African Multilingual Corpus},
  month        = feb,
  year         = 2023,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.7598539},
  url          = {https://doi.org/10.5281/zenodo.7598539}
}

### Licence
* Licence for Data - [CC 4.0 BY](LICENSE.md)