It may even be the best approach, as odd as that seems. It's the data. Or rather, lack of it. Machine learning is best trained with extensive, accurately tra

It may even be the best approach, as odd as that seems. It's the data. Or rather, lack of it. Machine learning is best trained with extensive, accurately translated, aligned bilingual texts. By far, the largest corpus for that is English-French. Unsurprisingly, English-French is the best language pairing for most machine translation systems. Closely followed by other major languages spoken in the EU, especially Spanish and German. The multilingual publications of the EU, as well as the UN, and companies and governments in countries like Switzerland and Canada, are the largest sources of such documents, as I understand. Without that kind of resource to drawn on, the quality of a direct pairing may end up worse than running it through a pivot language. That said, with German-French, for example, there would likely be enough sources out there to enable that directly. The article notes Japanese/Korean is a direct pairing, which makes sense given lexical similarity and also extensive bilingual texts. I'm kind of surprised Spanish/Catalan would get that priority before French/Spanish or Japanese/Chinese or German/French, though.