AI-Powered Translation Technique: Translate Fuzzy Matches In ChatGPT

Greetings in 2023!

Most commentators on GPT-3 (aka ChatGPT) seem fixated on how the GPT-3 model can help writers quickly create lots of new content.

But as a translator, I spend my days transforming content, not creating more of it. So I’m interested in other use cases.

Stanislav Okhvet (the guy behind TransTools) and I are working on a translator app that plugs into the GPT-3 API. I’ll have more to say on that soon.

But in the meantime, I’ve got something else exciting to share.

Try the following experiment in the ChatGPT interface at (You’ll need to set up a free account if you don’t have one already.)

Step 1: Find a matching source segment (“Source text 1”) and target segment (“Target text 1”) in any language pair you wish.

Step 2: Find a segment that is a low-value (say, 70% or so) fuzzy match (“Source text 2”) for “Source text 1”.

Step 3: Plug those segments into the following colored-text instructions (replacing the bracketed text; doing this in Notepad as an intermediate step is convenient; leave the space after “Target text 2” empty) and paste it all into the ChatGPT input box at once:

The following “Source text 1” is the translation of “Target text 1”:

Source text 1: [Paste “Source text 1” here]

Target text 1: [Paste “Target text 1” here]

The following “Source text 2” is somewhat similar (but not identical) to “Source text 1”.

Source text 2: [Paste “Source text 2” here]

I want you to translate “Source text 2” to into the language of “Target text 1” and write the new “Target text 2” translation below. The new translation should be phrased similarly to “Target text 1” but be an accurate translation of “Source text 2”:

Target text 2:

How did ChatGPT do? Did it work?

If the experiment worked for you (or even if it didn’t), will you email me with your results so I can collect them?

Based on my experience with English and Korean, this prompt has about a 65% chance of producing a perfect translation that reads naturally and maintains consistency with the original translation. Current fuzzy matching in the CAT tools can’t touch this.

Also, if you didn’t get a good result, perhaps it was your language pair? Or try again with another set of segments. Sometimes fiddling with the prompt wording can help in unexpected ways as well.

Anyway, remember that we’re still on GPT-3. Wait until GPT-4 comes out later this year; that’s going to be incredible!

I hope you found this little glimpse into the potential of the latest AI tool insightful.


PS – So far, we’ve frameworked dozens of translator use cases into the app based on the following categories: 1) machine translation, 2) style improvements, 3) technical editing, 4) grammatical transformations, 5) entity transformations, 6) workflow management, 7) terminology management, 8) corpus maintenance, 9) communications, 10) other things that work, 11) things that don’t work.

Steven Bammel

Steven S. Bammel is president and chief translator/consultant at Korean Consulting & Translation Service, Inc. A graduate of the University of Texas at Arlington (B.B.A. Economics) and Hanyang University (M.S. Management Strategy), Steven has worked for over twenty years in Korean business and translation. | more about Steven

You may also like...