Yeah, no great answer to that one. Localization of values is a big challenge, largely because many of the values you'd want to translate are sentence fragments and this is difficult for translation models to handle because they rely on word embeddings--context.
There are lots of ways this is a problem for translation models, but just consider the simplest one: homonyms. English is particularly egregious here, but other languages have plenty of homonyms of their own and distinguishing them is incredibly difficult from a sentence fragment, even if a human can look at the fragment and know immediately which is the correct interpretation.
Example: you build an app for a cryptocurrency expert and add a field "Notes" to one of their tables. They type "mined target" into their notes. Does that mean that they:
- Generated encryption keys for new bitcoin up to the threshold they intended,
- Dug into the earth to find a substance called "target", or
- Buried explosive ordinance around a giant bullseye pained on the ground
Sadly, while to a human this is obvious, to a ML model that only has those two words to go on, the meaning could be any of the above, so when translated the result becomes arbitrary.
