Comment by pjc50
2 days ago
My wife has been working on translation recently, and the LLM hit rate for novels is highly variable. It's capable of just dropping out entire paragraphs. You still need a final pass from a human native speaker editor to check that it makes sense. Which is what's happening in this article, the news site cares enough about their brand and quality to check the output.
I agree that bidirectional communication is probably going to work a lot better, because people are more likely to be alert to the possibility of translation issues and can confirm understanding interactively.
Definitely contextually dependent: things like news articles are probably the gold standard for machine translation, but creative works and particularly dense reading material like novels seem like they will remain a tougher nut to crack for even LLMs. It's not hopeless, but it's definitely way too early for anyone to be firing all of their translators.
> It's capable of just dropping out entire paragraphs.
I suspect, though, that issues like this can be fixed by improving how we interface with the LLM for the purposes of translation. (Closed-loop systems that use full LLMs under the hood but output a translation directly as-if they are just translation models probably already have solved this kind of problem by structuring the prompt carefully and possibly incrementally.)