The largest commercial translation engines target about 130 languages. These engines perform well for some languages; other languages struggle to produce meaningful translations of even general texts. And then there are languages for which MT capability does not yet exist.
In these instances, VoxCroft set out to build MT on models trained on domain-specific bilingual corpora. We also leveraged the similarities between related languages by building multilingual models. When we tested these in a real-world environment, they outperformed the MT engines which are typically trained on much larger corpora.
VoxCroft’s training data is human-translated and verified. For example, a commercial translation engine trained on 4 million pairs received a BLEU score – standard metric of translation performance ranging from 0 to 100 – of 22.6 on English-Tigrinya translations of news content. VoxCroft’s model, trained on only 80,000 sentences from online news content, received a BLEU score of 28.5.