Ask HN: If you translate with LLMs, GT or DeepL–what features are missing? | Hacker News Reader

Ask HN: If you translate with LLMs, GT or DeepL–what features are missing?

5 pointsorencoda3 comments

What is still the single biggest headache when you translate with GPT, DeepL or Google?

If you use DeepL – what would you need to finally pay for GPT translation?

Which features are still missing in current classical or LLM based translators.

How do you prefer pricing: SaaS, per-token, seats, or a flat plan?

Context: I’m exploring a side-project and don’t want to build the wrong thing. Not looking for promotion—just war-stories and must-have features (LLM model selection? scanned-PDF OCR? bilingual output? segment-level edits? glossary? missing format support?).

View on HN

> (LLM model selection? scanned-PDF OCR? bilingual output? segment-level edits? glossary? missing format support?).

Yeah, none of that matters.

I've worked with professional document translators for years, and compared their translation services with Google translate and Deep L.

Translations are still quite bad with LLM's for niche or technical topics. There are borrowed words, words that only make sense to specific professions or used in the right context etc.

Most translations suffer from overly literal translations between the English word and the target language word or phrase. Sentence structure, grammar, word choice, and industry research to find the common terms a 'topic-native' person would understand are all still very hard for LLM's to get consistently correct (and are the things that good professional translators do well).

There's a chance for common topics that the general public would be familiar with, that LLM's can now translate well enough. For business or medical or technical writing it's still a bit varied in quality.

It's hard to say if you don't speak the language, but I think it translates to the wrong level of formality. Often too formal when asked to translate and too intimate when talking to a user in that language.

They also get the tone all wrong. If you tell them what the tone should be like, they nail it. Often better than official human translations. Especially say, religious text. Some parts are meant to be heavy and grim, some are humor/sarcasm. But religious text often has that monotone thee/thou tone, especially in English. Which I feel is based out of a culture that places it distant, as something to be quoted instead of a proper story.

I think it often misses out poetic bits. The order of language in German, like "here comes me" has a stronger and more poetic feel than the English "here I come". Haikus are supposed to bring a wabi-sabi feel of imperfection and impermanence, but in English they rarely carry over.

Always prefer token pricing. None of the other forms of pricing really make sense to me; I guess they're just easiest for the VCs.

Sadly, OP, your reply was killed by the sockpuppet algorithm, so I'll respond here.

"...LLMs are good if you tell them the specifics (about the tone / context). Is it what you meant?"

They're inconsistent. I believe they were better 3 years ago. Today, they seem to be combining multiple sources and sometimes picking the wrong one. They are better at responding but translation took a slight decline.

You do need to give specific instructions on how to respond. Some of the cheaper ones seem to ignore instructions... GPT, Claude, Gemini, all three make this mistake.

I also mean there's a lot of room for improved quality and nobody is tackling that either.

"And any idea why people still use GT/DeepL? Is it the speed and cost? Risk of hallucinations? Or some features LLMs miss?"

I guess for the reasons people use Wikipedia. GT is verified by humans, including some experts. DeepL appears to be an AI and it does some languages well. It's a safe choice and popular brand.

Reply on news.ycombinator.com