Looking back at your comment I take your point that it's surprising that a low speaker-population language might have a decent corpus to work from... unless that corpus is the entire legislature and discussion archive of a pan-European institution.

The thing with the machines, as far as I see it to date, is that they rely on good, professional human input, which is limited. So until everything that needs to be said is said, and translated professionally, current models are stuck just behind in the creativity stakes.

But when it comes to boiler-plate and repetitious texts, computers can absolutely automate that. Particularly if overseen by competent homo sapiens. It's just a crying shame that the dream of the babelfish dissuades people from translating their literature and business offers today, as we all wait on and on for that point in time. A wait that could be a very long time, while human translators sit waiting for the middle-man agency to send another highly marked up project. (You may have guessed; I'm working on a related thing.)

Now, neural networks and massively parallel machine learning, that's a different story... If the machine manages to infer meaning and work within the confines of acceptable grammar use, it's game, set and match to the singularity.

That's not correct. The Council of Ministers, acting unanimously, decide on the rules governing the use of languages by the European institutions. Ireland (and Malta) would veto removing English as an official language.

https://ec.europa.eu/ireland/news/statement-on-behalf-of-the...