top | item 44617602

(no title)

maksimur | 7 months ago

I wouldn't mind and I would actually prefer if things were translated correctly, adapting to local slang, cultural references and the likes. As of now translations are weird if not outright cringey. This is a problem with LLMs in general (probably because they're trained mostly on English data), although to a lesser degree, but enough that I have to correct them daily.

discuss

order

xdennis|7 months ago

My experience with LLMs (OpenAI) is that it translates Romanian very well (including local slang), but that Google Translate (which I assume is not LLM-based) is very bad. Take for example this video https://www.youtube.com/watch?v=azwUg6cP9BQ .

Original title: "Swedish warm foundation for the house. Full construction process"

Actual title on YouTube in Romanian: "Fond de ten suedez cald pentru casa. Proces complet de construcție"

OpenAI translation: "Fundație suedeză caldă pentru casă. Procesul complet de construcție."

The OpenAI translation is perfect. The Google one is trash. It doesn't just omit diacritics, but translates "foundation" as "fond de ten" (literally "background of tan") a term used in makeup instead of "fundație".

LLMs have enough context here ("house", "construction") to know that "foundation" refers to the term used in construction not the one in makeup.