Dharmamitra.org, machine translation for Classical Asian languages

[Khemarato.bhikkhu]

@anon5008928

we are currently lacking the MN in our training data

I think I’ll quote our own @cdpatton at length here, who got ignored in the other thread:

@cdpatton

It’s not actually a translation. It’s a mishmash, arbitrarily put together, of existing translations. This is why when it’s compared to a text that has already been translated one or more times in the public domain, it’s output seems accurate. It’s based on the existing translations, reworded fairly freely to avoid copyright infringement, so it will look alot like them.

I feel like people may be misleading themselves when they check the accuracy of AI translation tools against existing translations. The output is much better in that situation. It’s when they’re used to “translate” something that hasn’t been translated by anyone that they usually go off the rails. Because they doesn’t have an existing translation to mimic. Instead, in that situation, they assemble something coherent from smaller words and phrases from other translations. And then it has lost the context that a complete human translation would provide. Sometimes they get borked or time-out internally, fail to translate a chunk of text, then restart and continue.

Basically, LLMs are plagiarism machines. They reword existing material.