>>1015459 (OP)
Well, the thing with languages other than autistic stuff like Lojban is that regardless of how much data you have it's almost impossible to get subtlety and double meaning just with plain text and no context, so anything other than academic texts or news articles seem to be really difficult.
What I've been thinking could somehow 'fix' the problem (at least until super-AI comes to save the day), is have some kind of 'bridge' language, so you would translate Lang1->BridgeLang->Lang2.
This hypothetical language would have the necessary features to write down all aspects of speech, and people would then try to translate something to the bridge language writing it in a way that the whole meaning stays intact, and then from a database of premade translations (maybe something like Tatoeba), ensamble the phrase/text in the target language, and since you already have the bridge language script written, it could be used to then automatically translate to any other language (supposing you have a rich enough database of translations).