AI as editor

161 views
Skip to first unread message

Tom Gally

unread,
Oct 1, 2024, 5:04:51 AM10/1/24
to hon...@googlegroups.com
A year and a half ago, when I was doing my initial feverish experiments using ChatGPT for translation, one thing that didn’t work well at all was having it check translations for accuracy and naturalness. Either the LLMs have gotten better or I have become better at prompting (probably both), because today I was able to use AI suggestions to polish a translation quite effectively.

The job is a 1000字 official statement that will be released by a Japanese organization in both Japanese and English. I prepared my first draft as usual, based on versions from several LLMs, and then polished my final version. Normally I would have been finished then, but this time I decided to ask the LLMs to comment on what I had produced.

My prompt was the following:

The following text in Japanese is an official statement that has been prepared by [organization name]. I have been asked to prepare the official English translation, which will be released to the public together with the Japanese. The translation needs to accurately reflect the meaning of the Japanese while also being correct and natural in English. The English text below is my draft of that translation. Please compare the Japanese and English carefully, identify any places where the English meaning deviates significantly from the Japanese, and point out any errors or awkward expressions in the English.

<original text> 日本語 </original text>

<translation draft> My translation </translation draft>

I gave this prompt iteratively to ChatGPT o1-preview, Claude 3.5 Sonnet, and Gemini 1.5 Pro 002, updating my translation each time with the suggestions I liked from the previous model.

The comments they gave me along the way looked like this:

組織的、包括的、効率的な管理体制を構築し: "Establish an organized, comprehensive, and efficient management system" is technically correct, but consider rephrasing it as "develop a robust, inclusive, and efficient management system."

Original Japanese: 「障害の有無」
Your Translation: "ability status"
Suggestion: The term "障害の有無" specifically refers to "presence or absence of disabilities". Therefore, "disability status" would be a more precise translation.

"Mutual respect for each other’s diversity": While grammatically correct, it might read smoother as "mutual respect for diversity".

"Competing according to the rules": The phrase "ルールに則って競う" emphasizes fair competition. For greater naturalness, consider "engaging in fair competition" or "competing fairly."

Original:「暴力、差別、あらゆるハラスメントを認めない。」
Your translation: "...we do not tolerate violence, discrimination, or any form of harassment."
Translation Issue: This phrase is translated correctly, but "any form of harassment" could be emphasized by changing to "harassment of any kind," which might read more naturally in English.

高潔性や社会性の涵養: "Fostering integrity and social skills" is accurate but slightly less formal than the Japanese. Consider "Cultivating integrity and social responsibility." "Social skills" is a bit narrow; the Japanese term implies a broader sense of social responsibility and awareness.

Not all of the suggestions were appropriate, of course. Several times an LLM would suggest changing word A to word B when word B was already used elsewhere in the same sentence and I had intentionally used a different word. But nearly all of the suggestions were worth considering, and it didn’t cost me anything extra to ask the models for suggestions again and again. And I didn’t find the LLMs’ comments nearly as annoying as I would have if human editors had been pointing out issues in my translation.

Tom Gally
Yokohama, Japan

Matthew Schlecht

unread,
Oct 1, 2024, 10:56:41 AM10/1/24
to hon...@googlegroups.com
On Tue, Oct 1, 2024 at 5:04 AM Tom Gally <tomg...@gmail.com> wrote:
A year and a half ago, when I was doing my initial feverish experiments using ChatGPT for translation, one thing that didn’t work well at all was having it check translations for accuracy and naturalness. Either the LLMs have gotten better or I have become better at prompting (probably both), because today I was able to use AI suggestions to polish a translation quite effectively.

Thanks for sharing this account. It is amazing or scary (or equal parts both) to see what AI is becoming capable of.
I read more and more about people having adapted to using AI as an office assistant, as you have done.
Most of your account is far outside my normal sphere, but my eye caught the last bit.
 
And I didn’t find the LLMs’ comments nearly as annoying as I would have if human editors had been pointing out issues in my translation.

Because, for better or worse, dealing with non-human translation/editing removes the personality factor.
It's no longer some other know-it-all human who is trying to school you through an attitude of superiority. Or, despite being in the editor/reviser role, trying to re-translate your text because they are convinced they are the better translator (been there!). It's just ones and zeroes doing what they do best.
I have had a feeling similar to what you describe about doing MT/PE (machine translation post-editing). When editing human translation, I always hold the translator's right to his/her own style as paramount. Even if I feel strongly that a phrase could be "improved", I leave it as is unless there is a flaw that is fatal to proper comprehension.
I have no such restrictions when doing MT/PE work, and feel free to make revisions without wondering "how will this make the translator feel?". Because the "translator" is just ones and zeroes.

Matthew Schlecht, PhD
Word Alchemy Translation, Inc.
Newark, DE, USA
wordalchemytranslation.com

Tom Gally

unread,
Oct 1, 2024, 10:04:46 PM10/1/24
to hon...@googlegroups.com
Matthew Schlecht wrote:

... for better or worse, dealing with non-human translation/editing removes the personality factor. ... Because the "translator" is just ones and zeroes.

Yes, I think that’s a really important point, not only for translation but for other applications of AI as well.

For the past couple of weeks, I’ve been experimenting with ChatGPT’s new Advanced Voice mode. Unlike previous spoken interfaces, it actually “hears” the user’s pronunciation and tone of voice, and its own speech is much more expressive than before. That makes it more useful for language learning than previous spoken interfaces. But as long as users perceive it as essentially ones and zeros and aren’t fooled into thinking that it has an identity and a personality, its role will be different from that of a human language teacher.

On the one hand, learners might be more receptive to feedback and error correction from a bot than from a human. In one test, I read aloud some Japanese text to it and asked it to comment on my pronunciation. Some excerpts from the transcript of that conversation:

Me: I'm going to read a sentence from the magazine Sekai. Do you notice anything about my pronunciation of the Japanese words that could be improved? There might be multiple things. 看板政策だった大阪都市構想が二度にわたり、住民投票で民意によって否定されたため、一時は大阪維新の会…

ChatGPT: Overall, your pronunciation was quite good! One thing to watch out for might be the pitch accent, especially in words like "住民投票" and "維新の会." Japanese pitch accent can be tricky, so focusing on the natural rise and fall of the pitch in words and phrases can help you sound more fluent. But honestly, you did well!

Me: Oh, thank you. That was a very good point about the pitch accent. That's one of my weaknesses. What about the vowel length? Let me read that part again. 看板政策だった大阪都市構想が二度にわたり、 住民投票で民意によって否定された ... [I intentionally pronounced some long vowels as short vowels.]

ChatGPT: Your vowel length was mostly good! However, in "住民投票," make sure the "u" in "住民" and "o" in "投票" are held slightly longer, as they are long vowels. It should be "じゅうみんとうひょう [read very slowly]." Keep an eye on those, and you'll be spot on!

Forty-one years in Japan, and for the past thirty-eight or -nine I don’t think anyone has corrected my pitch accent or vowel length, even though I still make mistakes and sound foreign. If a human corrected me now I would probably be annoyed. I have no problem with ChatGPT correcting me, though.

On the other hand, language learners will, I think, be less motivated to continue studying when they have only bots praising them, correcting them, and acting as models of language use. The motivational and guidance role of human teachers should continue to matter.

The lesson for translators might be to make sure that our clients know we are human and that they trust us and our human judgment more than what comes out of the much cheaper and faster AI.

Tom Gally

--
You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/honyaku/CALnYxcyBeQaM4CxxZs_5OYo93eQ0boafonEhae9Mu5g8OtmtHg%40mail.gmail.com.

li...@letstalktranslations.com

unread,
Oct 1, 2024, 11:54:09 PM10/1/24
to hon...@googlegroups.com

It’s an interesting obvious next step, isn’t it? Getting the computer to correct its output…

 

Everyone is running around trying to figure how to survive in the MTPE world, and all the while the MTPE world is working out how to remove the human from the equation altogether.

 

I came across “machine translation automatic post editing” MT-APE (?) for the first time the other day. (Of course, my first though was “why not machine translation machine post editing?” but I suspect the choice of “automatic” is a semantic game designed to redirect rather than inform.)

 

https://machinetranslate.org/automatic-post-editing

 

I really don’t see a future for translation at all unfortunately. And the computer programmers writing all the code? They are writing the code that will replace them. It’s like getting fired and one of your last jobs is to train your successor. Ha!

 

Michae Hendry, in Nagano Japan

 


Virus-free.www.avast.com

Geoffrey Trousselot

unread,
Oct 2, 2024, 2:05:25 AM10/2/24
to hon...@googlegroups.com
I posted last year that my employer has a vacancy and we have pretty much struggled along without it getting filled. The company’s current goal is simultaneous bilingual disclosure for listed companies. Still working on a mix of jobs including MT and HT. 

I personally think in-house is better than freelance. If you apply, I recommend haggling to get a salary that is commensurate with your skills. 
By the way, I don’t think I’m leaving and don’t go into the office enough to participate much in training. 

Also, book 5 of the time travel coffee series I’m translating is out or will soon be out. 

cpta...@ozemail.com.au

unread,
Oct 2, 2024, 2:45:38 AM10/2/24
to hon...@googlegroups.com

This was all very interesting.

But this was the most very interesting:

 

And I didn’t find the LLMs’ comments nearly as annoying as I would have if human editors had been pointing out issues in my translation.

Chris

--

You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

cpta...@ozemail.com.au

unread,
Oct 2, 2024, 3:17:02 AM10/2/24
to hon...@googlegroups.com

Hey Geoff,

I have been meaning to catch up, here in front of everyone, to laud your work translating the books of Kawaguchi Toshikazu. I interpreted for him when he appeared at the Melbourne Writer’s Festival this year and he was a phenomenon! The queue for the book signing was hundreds of metres and it took a lot longer than his actual time on stage. The next day we went around Melbourne seeing the sights, and literally walking past a woman with her nose buried in one of his (and your!) books at an outdoor table while she had lunch and he was “hiee!” throwing peace signs and she was all omg and he was signing the book and taking photos, it was mayhem. Same thing happened at a couple of other bookshops not on the schedule. He signed every copy they had! He never tires!

I read the first book in both languages before the job, I am embarrassed to admit that it is the first Japanese novel I have ever finished.

Two astonishing things I learnt:

He drafts all his books on an iphone, sitting in a café. (He edits on a computer at home) and,

He’s never met you.

 

How the HELL? I so couldn’t operate that way.

Anyway, mad respect for your success in this.

 

Chris

 

 

 

From: hon...@googlegroups.com <hon...@googlegroups.com> On Behalf Of Geoffrey Trousselot


Sent: Wednesday, 2 October 2024 4:05 PM
To: hon...@googlegroups.com

--

You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

image001.jpg

cpta...@ozemail.com.au

unread,
Oct 2, 2024, 3:17:50 AM10/2/24
to hon...@googlegroups.com

How do we even know this is Tom posting?...

 

From: hon...@googlegroups.com <hon...@googlegroups.com> On Behalf Of Tom Gally
Sent: Wednesday, 2 October 2024 12:05 PM
To: hon...@googlegroups.com
Subject: Re: AI as editor

 

Matthew Schlecht wrote:

Dan Lucas

unread,
Oct 2, 2024, 3:52:14 AM10/2/24
to Honyaku E<>J translation list
正社員登用後 想定年収500万円

No offense to the company Geoff, but I'm not surprised they haven't managed to fill the position. ¥500,000 a month is less than ¥3,000 an hour, assuming that there is no overtime, which I think highly unlikely. I'd be wanting a lot more than that for an on-site position in Tokyo if I had good bilingual skills and a good knowledge of specific areas of industry. Working from home would make it slightly more appealing, but only slightly.

Regards,
Dan
--
You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

Geoffrey Trousselot

unread,
Oct 2, 2024, 4:25:02 AM10/2/24
to hon...@googlegroups.com
Chris, thanks for the kind words and reporting on your meeting with Kawaguchi-san. I’m sure I’ll meet him eventually. I’d just end up complaining about the impossible challenges he leaves me. When the UK and US editors query me about things, I can just shrug my shoulders rather that passing on the query. 

Geoffrey Trousselot

unread,
Oct 2, 2024, 4:38:05 AM10/2/24
to hon...@googlegroups.com
Dan, 
I have mentioned it. And I personally think the right candidate could negotiate a higher salary. Overtime and bonuses are on top of the stated amount. Middle management don’t really understand the skill required.  They recently tried a temp agency and after three months of training, the temp was never given real work. I was tempted to suggest they should seek a temp position for middle management, but I bit my tongue. The other option would be some kind of outsourcing arrangement. If someone is interested, I suggest seeing the stated amount as a guide. 

On Wed, 2 Oct 2024 at 16:52, Dan Lucas <dan....@carninglipartners.com> wrote:
正社員登用後 想定年収500万円

Dan Lucas

unread,
Oct 2, 2024, 4:43:51 AM10/2/24
to Honyaku E<>J translation list
Thanks for the comments, Geoff. I suspected you would have similar views.

I wouldn't mind this attitude, but pretty much every financial report we translate contains references to labor shortages and the need for flexible working and so on. You would have thought that management would have realized that the same forces affect their own industry as well.

As for outsourcing - presumably that would be something less flexible that freelancing, but more stable? (E.g. guaranteeing the company a certain number of hours per week, and being paid for such.) If so, that might appeal to some...

Regards,
Dan
--
You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

Tom Gally

unread,
Oct 2, 2024, 9:37:59 AM10/2/24
to hon...@googlegroups.com
Chris asked:


How do we even know this is Tom posting?...

I was thinking the same thing. The post seemed much more intelligent than anything I could write.

Tom Gally

Leann Wolley

unread,
Oct 2, 2024, 11:32:19 AM10/2/24
to hon...@googlegroups.com
Geoff,

Pardon my sudden entry into the discussion, but have been following along with interest.

The position looks a good option, but I don’t suppose middle management’s flexibility would stretch to someone in a different timezone? I am an American based in the UK for the foreseeable future, so hours in the Tokyo office would not be the most realistic for me. 

Life circumstances make it challenging for me to go out hunting freelance opportunities as often as I’d like, so doing a regular stream of translation work sounds quite appealing. A girl can dream.

--Leann


2024年10月2日(水) 9:38 Geoffrey Trousselot <geoff.tr...@gmail.com>:
--
You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

cpta...@ozemail.com.au

unread,
Oct 2, 2024, 4:08:34 PM10/2/24
to hon...@googlegroups.com

lol

 

From: hon...@googlegroups.com <hon...@googlegroups.com> On Behalf Of Tom Gally
Sent: Wednesday, 2 October 2024 11:38 PM
To: hon...@googlegroups.com
Subject: Re: AI as editor

 

Chris asked:

--

You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

Geoffrey Trousselot

unread,
Oct 2, 2024, 8:59:20 PM10/2/24
to hon...@googlegroups.com
Hi Leann, the advertised position requires multiple small jobs, participation in other company activities etc that requires the same time zone. However, if you like, I can help you through the application process (including trial) for normal freelance work. Whether it would turn into a regular stream of work would depend on compatibility of skills, etc. Please contact me off list if you’re interested. 

Geoff 

JamesALockhart

unread,
Oct 4, 2024, 12:38:36 AM10/4/24
to Honyaku E<>J translation list
Hi Tom,
Your AI reviews are very interesting and informative, as usual for your posts.
This is slightly off-thread, so forgive me if it's too out of place.
One thing I was wondering about (at the lexical level): 管理体制.
I frequently see 体制 rendered as system, but it's my least-favorite English for it; so I was wondering about your take on how to bring out its full (and multifaceted) meaning(s) in English. My sense for it is that it refers a comprehensive infrastructure entail human actors/roles, organized structures, protocols (for communicating and doing things), and other similar physical and non-physical components set up to achieve a goal or run an operation. That said, I do like the “develop a robust, inclusive, and efficient management” part of the AI's suggestion, as I sometimes I wonder whether 体制 even needs to appear in the English if another word that implies it is used. I get this feeling about a lot of formal suffix-like markers that slightly clarify the meaning of a noun they’re attached to to signal how it's being used, and that these markers frequently show up in translated English in ways that make the English stilted or misleading, or even cause a shift in register between the original and the translation. Any thoughts?
—Jim Lockhart
Yamaguchi, Yamaguchi

li...@letstalktranslations.com

unread,
Oct 4, 2024, 1:35:39 AM10/4/24
to hon...@googlegroups.com

> I frequently see 体制 rendered as system, but

 

Not addressed to me, but FWIW…

 

I usually translate 体制 as “structure” these days, unless “system” or “NOTHING, AS YOU ALSO SUGGEST” seems more appropriate.

 

> I do like the “develop a robust, inclusive, and efficient management” part of the AI's suggestion, as I sometimes I wonder whether 体制 even needs to appear in the English if another word that implies it is used    

 

I’m not sure whether you are suggesting “develop a robust, inclusive, and efficient management” WITHOUT adding structure or system is suitable, but if you are, then I see it as problematic. It should still make sense if the adjectives are removed. In this case, “develop a management” is what you’d be left with, and at least to me, and SANS context, that seems strange.

 

Michael Hendry, in Nagano Japan


Virus-free.www.avast.com

Tom Gally

unread,
Oct 4, 2024, 3:36:50 AM10/4/24
to hon...@googlegroups.com
I don’t have an answer to Jim’s excellent question, but OpenAI released a new feature to ChatGPT earlier this morning called Canvas. I decided to give it a try to see if it would be useful for thinking about such problems:


It looks like it might be.

Tom Gally

Tom Gally

unread,
Oct 5, 2024, 4:54:30 AM10/5/24
to hon...@googlegroups.com
A followup to my initial post to this thread.

I just translated a short speech to be given at an international meeting. The original Japanese was dense and formal, and I wasn’t happy with my translation at first—it was accurate and correct, but I was afraid it might be difficult for the speaker to read aloud and for the listeners to understand. So I went to my AI editors again, running the original and my translation successively through ChatGPT, Claude, and Gemini and asking for their comments and suggestions. My prompt was similar to the one I used before. Once again, some of their suggestions were quite good, and the final version was better than what I had produced on my own.

As noted earlier in this thread, one advantage of AI editors is that I don’t get annoyed by their suggestions and corrections. Another, I realized now, is that I don’t have to worry about annoying them when I have them comment on slightly modified versions of the same text again and again.

Another thing to note is that, while these AI editors can provide good feedback on accuracy, meaning, and style, they are not the best proofreaders—they sometimes failed to catch my typos and dropped prepositions. I spotted those mistakes either when I pasted the translation back into Word and it underlined the errors or when I had one of OpenAI’s voices read aloud the translation for my final check.

Tom Gally
Reply all
Reply to author
Forward
0 new messages