Abstract
Low-resource African languages have not fully benefited from the progress inneural machine translation because of a lack of data. Motivated by thischallenge we compare zero-shot learning, transfer learning and multilinguallearning on three Bantu languages (Shona, isiXhosa and isiZulu) and English.Our main target is English-to-isiZulu translation for which we have just 30,000sentence pairs, 28% of the average size of our other corpora. We show theimportance of language similarity on the performance of English-to-isiZulutransfer learning based on English-to-isiXhosa and English-to-Shona parentmodels whose BLEU scores differ by 5.2. We then demonstrate that multilinguallearning surpasses both transfer learning and zero-shot learning on ourdataset, with BLEU score improvements relative to the baselineEnglish-to-isiZulu model of 9.9, 6.1 and 2.0 respectively. Our best model alsoimproves the previous SOTA BLEU score by more than 10.