I tried it on a couple Armenian lessons today, but because, as you said above, it’s a less common language, it was only useful for words that were already easy to translate. I was hoping it could help with some of the more obscure words I’ve run across, but the more obscure the word, the more wrong it got.
I do like the idea in theory, though, and I think the UX execution is pretty good. I like, too, the ideas mentioned above of attempting to fine-tune the training with Wiktionary or other open-source options if you determine it’s a useful feature to move forward with. That seems pretty smart.