Last November, The Bookseller reported Dutch publisher Veen Bosch & Keuning, owned by publishing titan Simon & Schuster, was testing the use of artificial intelligence to help translate several of its books to English.
Last November, The Bookseller reported Dutch publisher Veen Bosch & Keuning, owned by publishing titan Simon & Schuster, was testing the use of artificial intelligence to help translate several of its books to English.
I use local instances of Aya 32B (and sometimes Deepseek, Qwen, LG Exaone, Japanese finetunes, others depending on the language) to translate stuff, and it is quite different than Google Translate or any machine translation you find online. They get the “meaning” of text instead of transcribing it robotically like Google, and are actually pretty loose with interpretation.
It has soul… sometimes too much. That’s the problem: It’s great for personal use where it can ocassionally be wrong or flowery, but not good enough for publishing and selling, as the reader isn’t necessarily cognisant of errors.
In other words, AI translation should be a tool the reader understands how to use, not something to save greedy publishers a buck.
EDIT: Also, if you train an LLM for some job/concept in pure Chinese, a surprising amount of that new ability will work in English, as if the LLM abstracts language internally. Hence they really (sorta) do a “meaning” translation rather than a strict definitional one… Even when they shouldn’t.
Another thing you can do is translate with one local LLM, then load another for a reflection/correction check. This is another point for “open” and local inference, as corporate AI goes for cheapness, and generally tries to restrict you from competitors.
Or a tool for the translator to save time?
These language models don’t get the meaning of anything. They predict the next cluster of letters based on the clusters of letters that have come before. Sorry, but if it feels to you like they’re captured the meaning of something, you’re being bamboozled
It’s a metaphor.
They’re translating the input tokens to intent in the model’s middle layers, which is a bit more precise.
Actually, as to your edit, the it sounds like you’re fine-tuning the model for your data, not training it from scratch. So the llm has seen english and chinese before during the initial training. Also, they represent words as vectors and what usually happens is that similiar words’ vectors are close together. So subtituting e.g. Dad for Papa looks almost the same to an llm. Same across languages. But that’s not understanding, that’s behavior that way simpler models also have.
True! Models not trained on a specific language are generally bad at that language.
However, there are some exceptions, like a Japanese tune of Qwen 32B which dramatically enhances it Japanese, but the training has to be pretty extensive.
And even that aside… the effect is still there. The point it to illustrate that LLMs are sort of “language independent” internally, like you said.