Hacker News Reader

Comment by tkgally | Hacker News Reader

This is a nice essay, with an excellent example of the challenges and art of literary translation. But the author seems to assume that "computer translation" is limited to services like Google Translate or DeepL: put in a text, get a translation, with no consideration of the context or purpose. Properly prompted LLMs yield much better results and are already helping human translators produce better translations than they could on their own. Whether a completely automated workflow is possible for translating literary texts in an artful and engaging way—I don't know. But I wouldn't reject that possibility out of hand.

Addendum: The above comment has received at least two up votes and two down votes: in the hour after I posted it, I saw it go from 1 point to 3 points and back to 1. Perhaps people who disapprove of it could say why?

View on HN

I've been out of the translation game for quite a while, but I remember there being more to translating as a profession than just communicating ideas from one language to another. Sometimes precision and consistency is needed; complex scientific manuals come to mind. I recall using translation tools (e.g., translation memory) being essential to providing quality translations. TMs would help ensure that the same terms were used consistently throughout the text, and that similar phrases were translated similarly as well. The TM was an essential part of the workflow, because the first line of translators would pass their translation and the TM off to an editor, who would make suggestions in the text and TM, and the translator would update the TM based on those suggestions. It's been a while so I might have gotten a detail or two wrong, but that's the gist of the workflow.

I'm curious what sorts of prompts can replicate this. Again, there's a lot I've missed in the decade or so that I've been out of the industry. Does translation memory now become part of the context window? Does the context window self-actualize with new translations? What sorts of prompts can help produce quality translations?

Thanks for the reply. A few days ago, I posted several comments explaining my own use of LLMs in translation and related issues in this thread:

https://news.ycombinator.com/item?id=42894215#42895610

I didn’t say anything about translation memory tools, though, nor about maintaining vocabulary consistency. Those are not big issues in the type of translation I do now, though I recognize that they are important for many types of technical translation.

I actually don’t know how good LLMs are at maintaining vocabulary consistency even when they are provided with glossaries or translation histories. That deserves some testing.

Not an up or down here, but I’d be curious to see some sources for this claim:

> Properly prompted LLMs yield much better results and are already helping human translators produce better translations than they could on their own.

Thank you for the question. A few days ago, I posted several comments explaining my own use of LLMs in translation and related issues in this thread:

https://news.ycombinator.com/item?id=42894215#42895610

I didn’t mention it there, but I have had quite a few discussions with professional translators about LLMs over the last two years. Their reactions and attitudes towards AI are definitely mixed, but I do know translators other than myself who are using LLMs productively as part of their translation process.

Reply on news.ycombinator.com