We’ve developed an open-source, high quality isiZulu parallel corpus that comes from a mixture of domains, taking into account both Southern African context and international English context, by using professional translators. We sourced 5000 English sentences, sampled from News Crawl datasets that were translated into isiZulu. Additionally, we translated 5000 isiZulu sentences, sampled from both the NCHLT monolingual corpus and the open-source documents of the UKZN isiZulu National monolingual corpus, into English. From each set, we separated out 1000 patterns as the evaluation dataset. Since isiZulu is highly morphologically complex, we believe that the English-to-isiZulu evaluation set should be translated at least twice, by different translators which will allow us to calculate human-level BLEU score for the dataset. More details in the provided Data Statement

Thank you to Facebook Research for funding the creation of this dataset