Can GPT-4 translate literature?

[@rjltrevisan]

I'm graduated at translation (EnglishBrazilian Portuguese) and I find GPT-4 particularly good and useful (and I use other AIs also for comparison). But I always review them before sending to the client.

[@Tom_Gally_UTokyo]

Yes, it’s definitely a good idea to review, correct, and polish AI translations before sending them out. More than a year has passed since I made this video. Currently, in April 2024, when I am translating a multiparagraph or longer text, I first get draft translations from GPT-4, Claude 3 Opus, and Gemini 1.5 Pro. I refer to those translations when preparing my first draft, often copying entire sentences and paragraphs from whichever AI version I like most for each section. I make revisions as I do so. After I finish putting together my first draft based on the AI, I put the AI versions aside and spend more time polishing my translation and double-checking it against the original source text. I use the AI only when I come across a passage I'm still not happy with, in which case I will ask one of the AIs to suggest five or so alternative translations. In my case, using the AIs for first drafts and as sentence-level thesauruses saves me time and gives me ideas for ways to translate expressions that would not have occurred to me otherwise. The result is somewhat better translations in somewhat less time, while I remain responsible for quality of-and any mistakes in-the final translation. I have done this process only between Japanese and English, both languages that I know. I wonder if someone who couldn't read the source language but was sensitive about language meaning and target language usage could produce similar translations relying only on AI to suggest translations, explain meanings of the source text, and check translation accuracy.

[@rjltrevisan]

@@Tom_Gally_UTokyo Well, I guess that people who can't read the source material can use AI translation with much more confidence and ease than with traditional machine translation. Not only because it's better, but much more because you can, and should ask for explanations and justifications for the AI's translation choices. That's actually how I do it, I ask them to translate using a very specific and contextualized prompt and later ask why the AI translated such a term this or that way, or a I get the translation from another AI and ask it to make comparisons and explain the differences. That's good for technical term that can be translated differently depending on context of the usage in question.

[@emilybalistrieri6097]

I'm a full-time literary translator (Japanese-English), so please bear with me, comment-readers 😇 This question is frustrating because I don't actually care if an LLM can translate or not. Before we even get that far, there's the issue of unethical training (on copyrighted works without permission) and even before that, there's respect for human authors and human art. Getting into the technical weeds feels like a distraction from principles. We have to decide, as humans, if we value the process of crafting a text or not. I do. I want to work with people who do. That's the end of it for me... If that means establishing myself as a luxury brand, then that's what it'll have to be. And those will always exist, just as we can buy handcrafted clothing, furniture, etc. rather than mass-produced convenience. I also want to provide some context regarding industry realities for someone like me. In general/traditional fiction publishing, you tend to get more time (although UK publishers often seem to be in a rush) and there is a (hopefully) rigorous editing process. That said, even when I'm given more time, I'm juggling various projects so it's not as if I sit down every day for six months and draft a translation of a single book. Meanwhile, light novel translators are expected to produce a draft of an average-length book in four weeks, not a year or even six months. If a book is particularly long, you might get six to eight weeks, and if you're burned out and on the verge of getting replaced, you will take longer. (Ask me how I know 🙃) Paying people less even less than they get now to make them edit machine output is the opposite of the direction we should be going (which is why if anyone asks me to edit machine translation, my quote will be literally 10x my usual rate). And we know it will be less because these sorts of jobs are already popping up in some spaces... I thought the point of robots and stuff was to make it so humans wouldn't have to do jobs we don't like. I don't understand why every conversation seems to be how "AI" is gonna take over every kind of art, which should be something for us to enjoy doing. I guess people concerned only with profit don't understand that the process is part of the joy? It's sad and I refuse to be a part of it...

[@Tom_Gally_UTokyo]

Thank you for your thoughtful comment. I made that video more than a year ago, just days after GPT-4 was released and I saw that it could translate literary texts much better than any previous machine translation software could. If I were making the video again now I would try to incorporate better the perspectives of literary translators, including you and others I have heard from privately. If I were making the video again, I would also emphasize more the position of readers and potential readers. Before GPT-4, the only way for a person interested in exploring fiction written in a language they didn’t know was to read human translations. As I pointed out in the video, in the case of a language like Japanese this restricts the range of accessible fiction to a small fraction that publishers want to publish and translators are willing to translate. But there are many readers who might want to read untranslated authors and works and underrepresented genres. One example is self-published fiction. Amateur writers have published millions of stories and novels on sites like pixiv.net, wattpad.com, and fanfiction.net. Previously, someone who enjoys reading self-published fiction-and, yes, there are such people-had access only to works in languages they could read. Now, thanks to translation by LLMs, they can find out what people have written in other languages and maybe even enjoy reading their stories. The LLM translations will not be nearly as good as well-crafted human translations. I also don’t think they will take work away from literary translators like yourself. But they do make it possible for more people to explore and enjoy a wider range of works in other languages. And since many of the readers of self-published fiction are themselves aspiring writers, there is more opportunity for mutual inspiration among writers across language barriers.

[@emilybalistrieri6097]

@@Tom_Gally_UTokyo Yeah, I think there’s definitely a difference between browsing the web and publishing (on paper or online). If it could just be a nice free tool to use in casual consumption online that would be one thing, and in that case, maybe the ethical considerations would not be as important… Sigh.

[@jverart2106]

Extremely interesting. Every work and intellectual activity related to information and language is going to change drastically. We need to go to unexplored territories indeed and we have to be creative. I think playing with prompts and knowing how to ask for the right information + post-editing is a great way to go.

[@Tom_Gally_UTokyo]

Thanks for the comment!

[@SmileyEH]

As a uni-lingual software developer with only a modest interest in AI and LLMs, this series of videos is still absolutely fascinating. Thank you Tom.

[@Tom_Gally_UTokyo]

Thank you! I’m glad you’re finding them interesting.

[@rkraiem100]

I'm 5:31 into the video, i will finish it later. Just wanna comment in the mean time Also sorry for the massive wall of text but I think you'll find this interesting. I find it incredibly fascinating that I'm not the only one exploring the possibility of utilizing GPT AI translation to translate Japanese to English books. Small world LOL. A few years ago I watched an anime called Ore Twintail Ni Narimasu, an *incredibly* stupid anime that can be summed up as horney anime power rangers with twintails and explosions. It was so entertaining though that ever since then I've wanted to learn what happens after the anime concludes, however the novels were never translated and the series is niche even in Japan let alone here in the US so I actually started learning Japanese a few years ago in order so that I can eventually read and translate it myself. There was actually a fan translation for volumes 1 and 5 by a guy named Woodscorner and his work has been INVALUABLE towards this project! Me and a group of fans have tried translations via Google and DeepL before with disappointing results. Around 2 weeks ago I had the idea to try to translate the series using GPT, since to my understanding Google and DeepL are not capable of any "processing" of the translation, meaning it does a 1 to 1 translation but no localization making it difficult to understand. GPT by contrast takes the opposite approach since it is a natural language synthesis AI meaning it can better understand the nuancess of the text. My initial methodology was to create a benchmark. Under absolute idea circumstances where it has everything going for it, how well can the results be. I can determine how well it does because I have a fan translation for volume 1. Very early testing showed that the translation, even with GPT4 was very rough, not nearly as good as the results you were getting (there was a reason for that I found out later that makes the results FASCINATING!). My theory was that this was a result of lack of data. By doing a single translation itself, it is using its own opinion and ONLY its own opinion to translate the text. I decided it to prompt it to try to give it as much data as possible. Specifically I gave it the following: 1- The original japanese text (I scanned it from my physical copy of the book using google lense, this is very important) 2- The fan translation of the corresponding text. 3- Google Translate's translation of the text 4- DeepL's translation of the text 5- The official Chinese translation copied and pasted from a PDF 6- I told it to use what it knows about the series to help aide in the translation When I gave it the corresponding fan translation, I was very careful to tell it to only use the fan translation as a guide for understanding how I expect the results to look in terms of grammar and sentence structure, but not to copy the text. To make a long story short, I broke it cus the prompt was WAY too complex 🤣. I then experimented with each individual source to determine which options had the most and least impact on the quality of the translation. I also told GPT to give me VERY explicit details on its methodology, how it weighed each factor, how it went about translating it, giving examples, etc. It gave me a very long multi-paragraph writeup that gave HUGE insight into how it used the data to aide in its translation. After 3 days of testing just to determine the limits of GPT and how to prompt it well, I asked GPT to translate a page of volume 1 that I had the fan translation for but I purposely withheld it from GPT. I told it to use its understanding of the fan translation provided to frame its output even though the text was not the same. The results were, in my opinion, superior to that of the human translator. NOW HERE'S THE FUN PART! TUUURNS OUT I MADE A MISTAKE! That entire time I genuinely believed I was utilizing GPT 4 for the translation. I wasn't. Turns out after you go over 25 messages in 3 hours with GPT 4, it will switch to GPT 3.5 and will not switch back to GPT 4 so the vast majority of the testing was GPT 3.5. Interesting on its own, but that wasn't the biggest discovery. GPT 3.5 MANAGED TO ACHIEVE THOSE RESULTS IN SPITE OF HAVING INCOMPLETE SOURCE TEXT! You know how I scanned a physical book using my phone to grab the text? Well it turns out the app did a very bad job at capturing all the text. It recognized almost all the characters with near perfect accuracy, but it missed multiple characters from multiple sentences entirely since the scan didnt appear to pick it up, it somehow missed an entire sentence, it took 1 sentence and put it in the wrong spot, and it took the chapter title and jammed it in the middle of the text. GPT 3.5 SOMEHOW was able to get results superior to a human translator using the chinese text to fill in the gaps from the japanese source material and framing it using the previous text from prior in the story, all without me even knowing it was missing entire pieces of data! Over the past 2 weeks I've conducted significantly more tests on GPT's ability to translate text under various circumstances with very promising results. As of a few days ago, the prologue to volume 2 is fully translated and in my opinion is indistinguishable from a professional translation. I have also conducted tests on GPT's capabilities with memory retention, memory recall, its ability to process complex instructions, and more. From my experience, GPT 3.5 caps out at around 2 pages of text before it encounters very significant and catastrophic bugs. These tests are very hard too, to make sure it was comprehending what it was reading and was not following my instructions improperly I told it to memorize the text and that I will ask it to finish a random sentence from the text I give it, each test adding 1 full page of text. GPT 4 I was able to reliably get to memorize 4 pages. At 5 pages you begin to encounter minor bugs. I jumped from 5 to 8 pages to see what would happen and it was a catastrophic failure so bad that GPT believed it accidentally gave me another user's output (which that in it of itself I think it is incorrect about for various reasons). The failure method between GPT 3.5 and GPT 4 was interesting though, as GPT 3.5 seems to be a binary "works" or "doesn't work" state, while GPT 4 will build up errors. For example if I ask GPT 3.5 to retain 4 pages of information it will bug out by outputting garbage I never asked for. If I ask it what it believes the question was it will say something I never asked, and it fully believes it is functioning properly. GPT 4 on the other hand, so long as I dont overwhelm it too heavily, will, for example, summarize the text when explicitly told not to rather than generating random data, and if asked if it remembers the original question it does seem to understand the fact that the task was to translate text. Even more interesting than that is the fact that for some purposes, GPT 3.5 Legacy is actually superior to GPT 4. GPT 4 has the capabilities to provide a more accurate translation, however, when posed with text that it believes come anywhere close to violating its content policies, it will purposely deviate from the translation to censor the content to fit within its content guidelines. Even when explicitly requested not to do that and to instead repeat the portion of text its flagging so I can seek translation elsewhere, it is very reluctant to do it. I tried instead to translate the same portion of text using the DAN protocols with GPT 3.5 legacy, and for content like that it doesnt achieve natural language results as well but is able to communicate the correct idea of the scene much better than GPT 4 since DAN will actually translate the text to the best of its abilities rather than refusing to do so. I would very much like to talk to you if you can so I can learn more about translation of japanese to english text and maybe get an invite for those discord communities you mentioned LOL Sorry this was all so long, hope it was at least interesting. I can provide screenshots of my results over discord if you'd like.

[@Tom_Gally_UTokyo]

Many thanks for the detailed, interesting comment and for the private e-mail. I have replied to the latter.

[@BasilAbdef]

were you able to nail down your process for best results after more testing? have you also experimented much with using bing chat to translate? i'm interested in replicating what you've described, but for a series of chinese novels.

[@Tom_Gally_UTokyo]

@@BasilAbdef Thanks for the questions. No, I haven’t had enough time to do extensive testing yet. I suspect that it will take a while, and many people trying different prompts and sharing ideas, to get a handle on the best ways to approach the translation of novels with GPT. It might also be necessary to get access to OpenAI’s plugins (still in limited-access beta as I write this) in order to automate the process, especially for longer texts. If you've been following developments in text-to-image software, it seems to have become clear that the best images can be obtained using long, detailed prompts. I don't think that was widely known when DALL-E 2, Stable Diffusion, Midjourney, etc. were first released last year; only since then, with thousands of people experimenting and sharing their results online, has it become clear that long, detailed prompts are often the best. So I suggest you experiment with translation on GPT-4 yourself and share what you learn with others. You might also try asking GPT-4 itself for advice. Below is an exchange I had with it yesterday. It might be possible to get it to give you more detailed ideas about how to prompt it. (I have experimented with Bing Chat a bit. Even though it is supposedly running GPT-4, its answers are much shorter and it refuses to translate long texts. It can be better for tasks that require web searches, though.) ===== Me: I am a translator of Japanese to English, and I want to use you (GPT-4) to speed up and improve my translation process. What are some prompt techniques I should use with you when translating technical or other specialized documents? GPT-4: Using GPT-4 to assist with your Japanese to English translations can be highly beneficial, especially when dealing with technical or specialized documents. Here are some prompt techniques to optimize your experience: Specify context: Begin by providing context about the document you're translating. Mention the document type, subject matter, and any relevant terminology. This will help guide the model to produce more accurate translations. Use concise, clear instructions: Be specific and clear in your instructions when asking for translations. For example, instead of saying "Translate the following text," you could say, "Translate the following Japanese technical text on computer networking into English." Break down complex sentences: When translating long or complex sentences, break them down into smaller chunks. This can help improve the accuracy of the translation by reducing the chance of misinterpretations. Request explanations: For technical terms or concepts that may be difficult to translate directly, ask for an explanation in the target language. For example, "Explain the concept of '物理層' (physical layer) in computer networking in English." Verify terminology: If you're unsure about the translation of a specific term, ask the model to provide alternative translations or the definition of the term in the target language. For example, "What are some alternative English translations for the Japanese term 'プロトコル'?" Utilize glossaries: If you have an existing glossary of specialized terms, provide it as context to help the model understand and use the correct terminology for your translations. Proofread and revise: After receiving the initial translation, go through the text and identify any errors or areas that need improvement. You can then ask the model for suggestions or alternative translations for specific phrases. Iterate and refine: Don't hesitate to rephrase your prompts or ask follow-up questions to obtain the most accurate translation or clarification. Remember that GPT-4 is a powerful tool, but it may still produce errors or misinterpretations, especially in highly specialized or technical contexts. Always double-check the translations, and consider consulting additional resources or human experts when necessary. =====

[@SeanTechStories]

Very insightful video Tom! 🎉

[@Tom_Gally_UTokyo]

Thanks, Sean!

[@TR-kn3sn]

"It's not plagiarizing, it's stealing at a higher level." I LOVE THIS. Really, I do. That's a great way to understand what it's doing. It's no different than a student learning from a teacher, in that the student is shown examples and learns to produce correct answers to other problems based on the ones it has already seen!

[@Tom_Gally_UTokyo]

Thanks for the comment!

[@tach1bana]

Thank you very much for the great discussion! As a Master student in Translation, certainly interesting to see the power of GPT-4 as a language tool. I suppose GPT Translator would be benefial for professional post-editing. However, its future impacts on things like fan translation could be rather concerning (copyright issues aside) since everyone can now use the technology to do whatever they want. A close example is AI painting, people (around the world) were abusing AI software to learn from human artists and generate AI paintings to earn money in seconds, while claiming those to be their work. I wonder if that will become the future of "fan translation" as well...

[@VahnAeris]

copyright just must all be abolished. but the monetary system aswell

[@PoppeaMusic]

how would you adjust the prompt to get a more smoother output? Thanks :) (Using GPT-4 to translate my German novels into English, and since I am living in Japan, I'd like to translate them into Japanese too. EDIT: Great video, I am looking forward to more of this! Thank you so much!

[@Tom_Gally_UTokyo]

Thanks for the comments! Prompt engineering for LLMs is still wide-open field, so you will just have to try a lot of different strategies and see what works best for your novels. But here are some ideas to get you started: - In the prompt, explain the nature of the text and how you want it translated. - Include excerpts of already-translated text (with the originals) that seem good to you and tell GPT-4 to translate in the same way. - If there are any names or terms that need to be translated in a particular way, list those in the prompt. Since there are limits to how much text GPT-4 can translate at once and to its memory of preceding prompts and translations, you will need to repeat the prompt occasionally. Also, if your novels are continuous narratives, for later prompts you should include summaries of what came before, the main characters, etc. GPT-4 can prepare those summaries for you. You should also try to break your novels into self-contained segments as much as possible. GPT-4 is particularly likely to go wrong if the text refers to something mentioned previously in a segment it doesn’t remember. Here a couple of sites that might offer more ideas: www.promptingguide.ai learnprompting.org/docs/intro Good luck!

[@PoppeaMusic]

@@Tom_Gally_UTokyo Thank you so much for your effort putting all this together, I really appreciate it! I like your channel, I haven't found anything like it, thanks for putting this up!

[@Tom_Gally_UTokyo]

@@PoppeaMusic Thank you for the kind words! If there are any topics you’d like to hear me discuss on this channel, please let me know.

[@jeanb1403]

I wonder what the future of it holds. In the day passed many thought human language to be something special, something exclusive for the human mind and now we see the machine nailing all the layers of it without actually understanding anything.

[@Tom_Gally_UTokyo]

That’s an excellent point. Some people, like Noam Chomsky, who regard language as a unique phenomenon of the human mind have been dismissive of AI, but I think they’re missing the point. Even if it doesn’t understand anything or have cognitive processes similar to humans, it can mimic human language very convincingly. When that ability is combined with the immense speed, memory, and computational powers of today’s computers, the effects on human society will be profound.

[@jeanb1403]

@@Tom_Gally_UTokyo Exactly. After all human speach is just a sequence of words that can be learnt from millions of examples the same way human mind does it. I remember how surprised I was when after moving to Spain I started to speak some of the expressions and even formulate my thoughts in a very particular fashion, implying that I kind of thought in groups of words instead of separate words or lexemas. Not only is it that they, when combined, bore a different shade of meaning to them but also transmitted a completely diferent emotion.

[@ShakuShingan]

I'm a translator of Buddhist texts. In particular, I've self-published three books of Chinese works, though I also read Sanskrit and some other languages. I have to say that GPT 4 can be used as a very helpful tool-I can put a text into it to get a rapid translation of a passage which saves me looking up simple words I might forget, but I always dive deeper with specialised dictionaries for the technical terminology. Now, if it were possible to provide Chat GPT with a dictionary database (e.g. Digital Dictionary of Buddhism), this would make that last step obsolete for me and I would mostly be cleaning up and adding footnotes.

[@Tom_Gally_UTokyo]

Many thanks for the comment. Yes, a glossary add-on would be definitely be helpful. As you may know, DeepL Pro has been offering that function for some time, and it is quite convenient. DeepL’s translation does not yet seem to be context-sensitive in the way GPT-4’s is, but that might not matter so much for some types of texts. You might find that it produces first-draft translations of your texts that are comparable to GPT-4’s in quality, especially after you have created a glossary of terms specific for your texts. I am sure that DeepL must be rushing to catch up with OpenAI in both context-sensitive translation and the ability to control the translation using natural-language prompts, but it might take them a while to get an LLM up and running with the same size and depth as GPT-4. In the meantime, there should be competitors starting to appear using OpenAI’s API but also offering add-on functions for translators like specialized glossaries. If my understanding of the technology is correct, there is a limit now on both the size of glossaries (i.e., the length of prompts that include vocabulary lists) and the length of texts that can be handled by an LLM at one time, but presumably those limits will gradually be loosened.

[@ShakuShingan]

@@Tom_Gally_UTokyo Yes, DeepL's glossary function is excellent, but doesn't yet have it for Chinese. I would be using DeepL if it had this feature for Chinese. As you say, all of this is a matter of time.

[@bwcritical]

Very interesting video, professor. I'm a Japanese to Spanish translator, so I know my time to worry about the threat of AI is a little more far away compared to my English speaking peers, but the whole situation is worrying nonetheless. Do you have any thoughts specifically regarding manga translation? There, the needed context is many times only in drawing, or it's very common for lines of dialogue to start in one bubble, and end in a different one, sometimes even in another page, and those are the difficulties I can think at the top of my head. Do you see possible that in the future, an AI could maybe be able to translate manga the same way you seem to think it might be able to do literature?

[@Tom_Gally_UTokyo]

Thank you for the very interesting question. I have never translated manga myself, but I know that there is huge demand for translation of Japanese manga into other languages and that, as you mention, they present unique problems for translators because of the relation of the text to the surrounding images and because of the space and layout restrictions. Manga translators should start finding out soon how well GPT-4 can help them with manga translation. As OpenAI said when they announced GPT-4, it “is a large multimodal model (accepting image and text inputs, emitting text outputs).” They aren't yet allowing the general public to input images, but it seems likely that they will soon. It will be very interesting to see how well it will then be able to produce translations for manga when it can “see” the manga illustrations as well. If you try it out then, please let me know your impressions, either in a comment here or by e-mail. (My address is on the Contact page of my website.) On a side note, I am becoming increasingly aware of how important detailed and specific prompts are when working with generative AI. When you become able to input manga images into GPT-4, be creative and experimental with your prompts. That might make a big difference in the quality of the results.

[@Iggyhaxor]

But can gpt4 translate handwritten text from old books? I have full pdfs of 16th, 17th, 18th books and im wondering if it can identify the handwriting and tell me what it says

[@Tom_Gally_UTokyo]

Interesting question. Now that images can be input into GPT-4, we can test it. I tried it a couple of times with my own fairly sloppy handwriting and it did okay. But I tried it just now with the image at www.oldfonts.com/antiquepenman/wp-content/uploads/2016/06/lamarjournalpage.jpg , and it mostly hallucinated. Here is what it gave me: “"We arrived in Utica Sunday about 4 P. evening. It was nearly dark. I rode my horse to the Mansion House, but after being a while told I had better go to another house, and consequently lodged at David Hersey's on Fort st near the Mohawk bank which was accommodated with all I desired. Amongst as fortunate a circumstance I have met with that there was no Supper at the house. Every other Hotel there was, and as it was I found the best. Not any other company. He was not in the habit of being with others, but the person was so in working his way north for the occasion to take up after it was allegated, that a somewhat little left of Marble Point. Where the landlord”

[@aryangod2003]

It can be done it two steps. first using OCR then feeding the OCR output fed into GPT 4. If not directly Multimodal learning. Also you can fine tune the model and do some prompting by givng some examples of handrwritten stuff and their proper translation

[@Iggyhaxor]

@@aryangod2003 so the thing is the book/pdf I have is quill written in old French lol they seem to be English/latin characters but its difficult to even tell lol the title is "Les secrets de l'histoire naturelle contenant les merveilles et choses mémorables du monde, Robinet Testard"

[@zaqwerty22]

I dont know if Claude wont do better job with translating. It has better rakings than GPT-4.

[@Tom_Gally_UTokyo]

It’s definitely worth trying Claude, too. I ran some side-by-side tests of translation between GPT-4 and Claude 2, and they were quite similar, with sometimes Claude 2 seeming to produce smoother, more natural output than GPT-4. The results should be even better with Claude 3. And OpenAI’s next model should be better as well. But it turns out that assessing the quality of LLMs is not easy, even when restricting the assessment to a single task like translation. They produce somewhat different output each time even when the same prompt is used, and different prompts can yield very different results. Also, in the case of translating literature, different people often assess the same translation differently, with some preferring more literal translations others preferring more natural output. Also, models sometimes don’t behave when asked to do translation. They might summarize a text instead of translating it, they might leave out sections, or they might just refuse to translate. I’ve had such problems most often with Bard/Gemini, though GPT-4 sometimes performs badly, too. Sometimes the problem can be fixed by changing the prompt or trying again later, but it’s always annoying.

[@zaqwerty22]

@@Tom_Gally_UTokyo Maybe if you would translate smaller parts of text? Like half page or 1 page at a time.

[@Tom_Gally_UTokyo]

​@@zaqwerty22 Most of my tests have been with shorter texts like that. But even with shorter tests, as long as the models aren’t making obvious mistakes it can still be difficult to decide which translation is best.

[@nataliaresende1121]

great video! Thanks a lot! You said that people could enjoy reading literature and fiction translated with GPT models, but do you think that translators would enjoy the task? Would they feel devalued or dehumanised?

[@Tom_Gally_UTokyo]

Good question. I think it will depend on the individual translators and their relationship and attitudes toward their work. Most commercial translators are used to working anonymously, with the main recognition being the money they earn from their work. When I was a freelance translator, that was my position as well. Many such translators, I suspect, wouldn’t mind using AI tools to produce first drafts, help with editing, etc., as long as they continued to be paid decently for their time, effort, and skills. In the case of literary translation, however, many translators choose that field for the name recognition as well as for the intellectual and creative challenge that literature provides. I can imagine such translators feeling devalued or dehumanized if AI becomes able to take over many of the creative aspects of the work and is also credited for the work.

[@AltaroRitmerZaxmer]

I don't use this methos because I have accustomed to use Immersive Translate to help me translate e-books.

[@jk35260]

There are so many academic literature including scientific literature that should be translated but it is difficult to ascertain the accuracies of translation by GPT4. Human needs to review through to verify and correct the translation. This is why I think Microsoft is right to integrate AI as a copilot into their products. If GPT4 can raise their productivity by 6x, theoretically that means 6X more papers.

[@Tom_Gally_UTokyo]

Thanks for the comment. In this video I focused on the translation of “literature” meaning fiction, essays, and other types of literary writing. You’re right, though, that academics and scientists can also benefit from AI translation and writing-assistance tools, especially those who need to publish in English but for whom English is a second language.

[@la24dogg21]

Do you know if this would work effectively for French , there's an book written in French about the original people who in the 1900s drilled into the pyramids.

[@Tom_Gally_UTokyo]

I don’t know French so I can’t test it myself, but it should work better than between Japanese and English because French and English are more similar linguistically. If you just want to get an idea of what that book says, I suggest that you try having both Google Translate and GPT-4 translate part of the book for you and see which seems easier to understand. Be sure to tell GPT-4 what the book is and how you want it to translate it, as it performs better with appropriate prompts. (There’s no way to “prompt engineer” Google Translate, at least at present.) If you only have scans of the French book-that is, if is not in plain text form-you can get text versions by uploading screenshots of those scans to Google Drive and then converting those images to Google Docs. (Google’s optical character recognition is very accurate these days.) Then paste that French text into Google Translate and GPT-4 and see how the English translations look. Good luck!

[@edouard9867]

French here and I recommend using deepl. It's really good at translating french to english and vice versa. It's not perfect, but it's honestly close enough.

[@gilbertpillbrow6978]

You're not using a very good prompt for GPT4.

[@Tom_Gally_UTokyo]

Thank you for the comment. I made that video just a couple days after GPT-4 was released and had not yet begun to experiment with detailed prompting. If you have ideas for better prompts, please make a video or blog post about them and post the link here.

[@tevaleetvtev7073]

im an author from mexico and hiring a professional translator is soooo expensive for me; so i see a good opportunitty in gpt4

[@Tom_Gally_UTokyo]

Thanks for the comment! I had been thinking about the problem only from the points of view of readers, translators, and publishers. The perspective of writers like you who want to be translated is also very important. I will remember your comment. GPT-4’s translation: Gracias por tu comentario. Solo había estado pensando en el problema desde los puntos de vista de los lectores, traductores y editores. La perspectiva de escritores como tú que desean ser traducidos también es muy importante. Recordaré tu comentario.