AI translation

[Karlin]

I published a novel in English, a novel that has potential audiences among Hebrew and Chinese readers. I am an AI skeptic, a Luddite. If I could type on this forum with a quill and homemade ink I would. My son, despite his positive qualities, is, heaven help us, an AI USER, and he said "hey, let's see if ChatGPT can translate the book into Hebrew." So we fed it the text, and in Deep Thought fashion, it said: "I estimate it will take several days to complete a ...". Better than 7.5 million years.

Well, my son asked it/she/him/they for a preview, just the first chapter. I am bilingual, though English is my mother tongue, so I can read it. It needs some editing, but far less than I expected. In fact, to my surprise, it is pretty darn good.

Maybe I should give up on the quill and invest in a typewriter after all.

I wonder how it will do in Chinese, or if I should try DeepSeek. I'll have to ask for help checking it.

Have any of you had experience with AI translations?

 

[skip.knox]

None with translation, but I'm currently working with audio narration. The AI is surprisingly good, but teeters just short of good enough. It makes odd mistakes, some of the tools are rather clumsy, and there are some features that would move it over into the Genuinely Useful column.

I figure the problems I see in narration will have their analogues in translation, arguably a more sophisticated challenge. The AI narrator cannot, for example, tell the difference between "lives" and "lives" (the one a verb, the other a plural noun). I have many other examples. Translation is many times more complex. When does one choose to translate a word as a phrase, for example. What does it do with antiquated word or phrases? Can it distinguish between irony, mockery, satire, and scorn, just to name a single example.

Then there's my favorite test, which is translating Jabberwocky. How does one translate nonsense in a sensible way?

And even after allowing for all that, the plain fact is, there are not merely good translations and poor translations, there are translations that appeal to one generation and not to another, or indeed one individual to the next. It's not at all an easy question.

 

[Prince of Spires]

Professional translators use AI. They have been for years, though they called it a computer program before AI became a hype. Source: my father was a translator at the largest translation organisation in the world (he's retired now).

Howeve, they don't use AI to create their finished product. The reason is as Skip mentions. AI has no clue about what it's actually translating. They're just words, and it guesses what the best replacement word is. The problem is that languages don't translate one-on-one. Simple things like homonyms aren't homonyms in all languages. And where is the difference between a brook and a creek and a stream and a rill and what have you?

So throwing a novel into an AI program and asking it to translate is a good first step. But then you need someone who is fluent in both the language it's being translated into and the original language to go through the whole thing and actually writing the whole translation so it's an actual novel instead of just a bunch of words that roughly resemble to original.

 

[pmmg]

Without the ability to read and understand the other language, i would never know. Id still need a human that could verify it for me.

 

[minta]

I suggest splitting the book into smaller chunks that I think an AI would understand and work with since it has a relatively small context, I believe GPT has a 128k context window (deep seek same) while Gemini has a bigger context and in my own testing with both Gemini is quite good in translating.

As the previous comments suggested, after finishing, you would give them to a native/proficient person to verify and correct any mistakes (or hallucinations) AI would make.

 

[Karlin]

Professional translators use AI. They have been for years, though they called it a computer program before AI became a hype. Source: my father was a translator at the largest translation organisation in the world (he's retired now).

Howeve, they don't use AI to create their finished product. The reason is as Skip mentions. AI has no clue about what it's actually translating. They're just words, and it guesses what the best replacement word is. The problem is that languages don't translate one-on-one. Simple things like homonyms aren't homonyms in all languages. And where is the difference between a brook and a creek and a stream and a rill and what have you?

So throwing a novel into an AI program and asking it to translate is a good first step. But then you need someone who is fluent in both the language it's being translated into and the original language to go through the whole thing and actually writing the whole translation so it's an actual novel instead of just a bunch of words that roughly resemble to original.

There is no question that much that is labelled " AI" is simply computer programs with a new label. But that isn't what is happening here.
It's not just word by word. It doesn't take several days for a computer program to look up 70,000 words in a dictionary. That would take a few minutes. It's mimicking other texts it finds. The chapter Reads Right.
Note that this is two unrelated languages.

If you know two languages, try it. Or find someone bilingual To help test it.

 

[Mad Swede]

Writing this both as a traditionally published author who's books have been translated and as someone who has used translation tools professionally.

Translating a text is as much about interpretation as it is about translating words from one language to another. To properly convey nuances you need to interpret the text you are translating, both in the context of the work itself and in the cultural context of the society/country or organisation in which it will be read. In my experience no translation software or AI tool can do this properly, and you (or whoever is doing the translation) will end up going through the machine translated text several times doing a lot of editing. The more subtle your original text is the more you have to edit the translation.

This is all perhaps best illustrated in an article published in todays Svenska Dagbladet, where Editor-in-Chief Lisa Irenius writes about the newspaper's attempts to use AI tools to quickly translate texts from abroad into Swedish. The article is titled "Vår pinsamma miss visar på problemet" (Our embarassing mistake illustrates the problem) where she describes what happened when an AI tool tried to translate an AP article about the recent Signal chat that some US government ministers were engaged in, and then goes on to discuss other translation attempts. Lisa Irenius concludes that such AI tools can sometimes produce a reasonable rough translation, but that a lot of work is then needed by journalists and editors to produce an accurate text for publication.

 

[pmmg]

I'll say...I don't think we would ever call them ministers Fools maybe, jerks.... but thats not really our term.

 

[dollyt8]

I published a novel in English, a novel that has potential audiences among Hebrew and Chinese readers. I am an AI skeptic, a Luddite. If I could type on this forum with a quill and homemade ink I would. My son, despite his positive qualities, is, heaven help us, an AI USER, and he said "hey, let's see if ChatGPT can translate the book into Hebrew." So we fed it the text, and in Deep Thought fashion, it said: "I estimate it will take several days to complete a ...". Better than 7.5 million years.

Well, my son asked it/she/him/they for a preview, just the first chapter. I am bilingual, though English is my mother tongue, so I can read it. It needs some editing, but far less than I expected. In fact, to my surprise, it is pretty darn good.

Maybe I should give up on the quill and invest in a typewriter after all.

I wonder how it will do in Chinese, or if I should try DeepSeek. I'll have to ask for help checking it.

Have any of you had experience with AI translations?

Don't try DeepSeek. It's not reliable at all in about a hundred ways.

I have used AI regularly in a professional context (I'm required to by the company I work for) and I can say that it's surprisingly good at translation, specifically ChatGPT. I'm fluent in Spanish and it's still made things so much faster to translate into Spanish, it's very accurate, and even native Spanish speakers have been surprised by it. That said, it will do better with the most common languages (including Chinese), whereas with Hebrew it will probably struggle somewhat more. I'd say as long as you are comfortable with it, divide your story up into smaller chunks that ChatGPT can handle, ask it to translate, and then have a native speaker correct. It will be much faster than a native speaker trying to translate it themselves. I will say though that translating fiction with AI is always a tricky business. I do articles on tech, and it still doesn't perfectly get the "feeling" of the original. I can only imagine it's much more so with fiction.

 

[Karlin]

Don't try DeepSeek. It's not reliable at all in about a hundred ways.

I have used AI regularly in a professional context (I'm required to by the company I work for) and I can say that it's surprisingly good at translation, specifically ChatGPT. I'm fluent in Spanish and it's still made things so much faster to translate into Spanish, it's very accurate, and even native Spanish speakers have been surprised by it. That said, it will do better with the most common languages (including Chinese), whereas with Hebrew it will probably struggle somewhat more. I'd say as long as you are comfortable with it, divide your story up into smaller chunks that ChatGPT can handle, ask it to translate, and then have a native speaker correct. It will be much faster than a native speaker trying to translate it themselves. I will say though that translating fiction with AI is always a tricky business. I do articles on tech, and it still doesn't perfectly get the "feeling" of the original. I can only imagine it's much more so with fiction.

My publisher has had similar experience with AI translation. Really helpful, but need to go over it.

we've already fed the entire novel into chatGPT, we'll see soon (hopefully) what it managed to do with it.

 

[A. E. Lowan]

I'll say...I don't think we would ever call them ministers Fools maybe, jerks.... but thats not really our term.

In the United States we don't call our legislators and lawmakers and leadership "ministers," no, but other countries do. Even in the US we have a lot of variation when it comes to government titles. What most of the country would call a "detective," is in San Francisco an "inspector," like a lot of Europe.

When you live on the top of the hill and all you see around you is yours or someplace you've deeply influenced - read: Colonialism - it can get difficult to remember that there's a few hundred other countries, many more cultures, and tens of thousands of years of the push and pull of historical events. It's so very easy to see only the familiar and assume that its like that, everywhere.

None with translation, but I'm currently working with audio narration. The AI is surprisingly good, but teeters just short of good enough. It makes odd mistakes, some of the tools are rather clumsy, and there are some features that would move it over into the Genuinely Useful column.

I figure the problems I see in narration will have their analogues in translation, arguably a more sophisticated challenge. The AI narrator cannot, for example, tell the difference between "lives" and "lives" (the one a verb, the other a plural noun). I have many other examples. Translation is many times more complex. When does one choose to translate a word as a phrase, for example. What does it do with antiquated word or phrases? Can it distinguish between irony, mockery, satire, and scorn, just to name a single example.

Then there's my favorite test, which is translating Jabberwocky. How does one translate nonsense in a sensible way?

And even after allowing for all that, the plain fact is, there are not merely good translations and poor translations, there are translations that appeal to one generation and not to another, or indeed one individual to the next. It's not at all an easy question.

We're looking at AI for audiobooks, since Amazon and the Baby 'Zons have lifted the restrictions. My wife is impressed. It's gotten a lot better over the past couple of years. It's tempting, especially since we've had narrators ghost on us more than once, like so many others have had. But I've got one thing that makes me hesitate, and it's that I can still hear the AI in the audio. If you can hear it, you know what I'm talking about. I gatekeeper my in-laws' phone a lot. They're both older and now here come the scammers. I refuse to talk to AI on the phone. I can always spot them, and it's because I'm a wiseass by blood and absolutely will start cracking jokes right out the gate. A human staggers. They aren't sure how to process sudden, rapid fire joking, so the hesitation is audible. An AI won't do more than pause for you to speak. That's all. No change in tone, no confusion, no tripping. And that's when I cut the connection.

I think I want her to keep looking at it, but until I can't hear the AI, I'm not sure I want to lay that at the feet of our readers. They're carnivorous and they are rabid and most of them are smarter than I am. Send help!

 

[pmmg]

Oh. Im sorry. I must have forgot that other countries do things differently