Is there a #greenlandic speaker here? Google translate just launched #Kalaallisut, I'm trying it out but the answers look wild? I *think* it is translating danish via english to Kalaallisut, but that is a bizarre decision because there must be many more training data from danish.

I'm now using
https://nutserut.gl/machine from the official #Greenland govt and it seems to work (ish). The problem is I only have a few words of Greenlandic, so it's hard to be sure.

#askFedi

Nutserut - Greenlandic-Danish Machine Translation

@Ruth_Mottram "Pivot through English" is pretty common for machine translation. This is partly because it's often hard to find enough text for most language pairs and partly because translation models are *huge*. If you could manage to keep models for all language pairs you'd need N^2 models, most of which would be very rarely used, but if you pivot through English you only need 2N models.

Non-english models are built and kept, but only for pairs that see enough use to justify the cost.

@Ruth_Mottram So the net result is that even in cases where there's a large enough corpus to build a good translation model (and there may be for Danish/Greenlandic) the model would see little enough use that you couldn't justify the machine resources to build, store, and maintain it. Which sucks, frankly, but is what it is.