Hacker News Reader

Comment by dylan604 | Hacker News Reader

If you were willing to do that, why not just start there in the first place?

Sometimes they really mess things up when there’s little context.

A few basic examples: “file” in English is both a verb and a noun, but other languages have different words for each. The machine translator will just guess at presenting you the noun or verb.

Other languages have genders for everything, so if a computer is male, but a standalone sentence doesn’t know that, the translator has to take a guess. If it guesses wrong, the sentence will be confusing to the user that thinks it’s about something other than a computer.

That’s just scratching the surface.

This recently was an issue for zip files in UK Windows (or something) being translated as "post code files"

Reminds me of a mistake in Spanish in a TomTom device I have around: "AP" stands for autopista (motorway) but it reads "AP" aloud as apartado postal (PO box).

Have had some satnav apps when set in French translate “dr” in an Anglo street name as “docteur” instead of reading it as “drive”.

Just because I set the language to something foreign, it doesn’t mean I’m in that country with its street naming conventions.

Have the machine translation more as a sanity check than the actual translation: you don't want to blindly trust what the machine produces if it's in a language you don't understand. But going the other way for quality control can eliminate the most egregious cases like this.

Sounds like more fine tuning is necessary. At this point in time with LLMs, I'm really surprised at how auto-translation is still so not natural. I mean, I get why it is that way, but at the same time, it's no less shocking how bad it is.

Subpar results; it's harder to translate than to check a translation.

Reply on news.ycombinator.com