Example of Claude-designed translation workflow

52 views
Skip to first unread message

Tom Gally

unread,
May 14, 2026, 11:16:36 AM (7 days ago) May 14
to hon...@googlegroups.com
Some years ago, I translated a dozen or so essays by 寺田寅彦 for Iwanami Shoten’s 科学 magazine. They were later published in two books. Yesterday, an editor at another publishing company contacted me, asking me to write an essay about my translation process. I told him that I would prefer to write about the issues of AI and translation, and I probably will. The deadline is several months off, so I have some time to mull it over.

In the meantime, I had Claude start creating a knowledge wiki focused on the translation of such essays. After it worked on that wiki for a couple of hours, I had it design and run a workflow for translating one essay as a test. I haven't checked the translation, but I thought its detailed explanation of the workflow might be of interest:


Tom Gally

Tom Gally

unread,
May 15, 2026, 9:01:28 PM (5 days ago) May 15
to hon...@googlegroups.com
Yesterday morning, in the shower, I wondered if it would be possible to create a system in which various elements of the translated text—register, literalness, etc.—could be varied along different axes. I tossed the idea at Claude and it spent the day extending and testing it, with a focus on translating essays by Terada Torahiko. Here are the results of the first test:


Presumably similar systems could be built for other, more commercial purposes. That would make it easier to create translations tailored to the particular needs of a client.

Tom Gally

Kevin Kirton

unread,
May 16, 2026, 1:21:55 AM (5 days ago) May 16
to honyaku
Fascinating work, Tom, thanks for sharing.
The sudden quality improvements in machine translation brought about by LLMs was initially very dispiriting for me. I remember long threads on honyaku in the early 2000s about whether machine translation would ever reach human quality, but I'm assuming few translators would argue that it hasn't at least come close now. And when you factor in speed and cost, it's a harsh competitor for humans. But your experiments show how important and creative the human in the loop can be. Very interesting time to be alive.
Kevin Kirton 
Canberra, Australia

--
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 visit https://groups.google.com/d/msgid/honyaku/CACnS3Nfh2UuhJ85Tw8SxFN0a5NZu33v%3DftOhUOhpjKBbtQWpwg%40mail.gmail.com.

Tom Gally

unread,
May 16, 2026, 10:35:25 PM (4 days ago) May 16
to hon...@googlegroups.com
Thanks for the kind words, Kevin.

I should point out, though, that my creative contribution to that project has been minimal. Yes, it was my idea to try varying translations across multiple axes, but it was Claude that actually came up with the specific axes for Terada Torahiko essays (other than two or three that I suggested). It also devised and implemented the idea of testing the validity of the parameters by having other LLMs evaluate sample translations at various settings. That would not have occurred to me. 

Regarding the issue of human vs. LLM-driven machine translation: At least for the sorts of translations I do and have been testing with LLMs, the quality differences had already entered the range of individual preference a year or more ago. Once it became possible to have LLMs check each other’s translations for omissions and mistakes, the differences from human translation came down to subjective judgment and taste. When I use a multi-LLM workflow to translate a speech now, I still make a lot of changes to their final draft before submitting it. But the changes I make are no longer corrections of the text as a translation. Instead, they reflect my own preferences for English writing and speaking style; other human translators and editors might very well prefer the machines’ versions.

And regarding that experiment I did with having the same text translated sixteen different ways with various combinations of parameter settings: Are we safe in saying that it would be impossible to get such a set of translations from human translators? At least, I myself would not be able to create a set of translations with the variations controlled to certain specified parameters like that, and it would seem difficult to assemble a team of human translators to do so, too.

Tom Gally

Reply all
Reply to author
Forward
0 new messages