Find-and-replace all in target text, how to do it?

Hello,

After visiting the Trados booth at the ATA conference this year, I was sold on the AI features of Studio 2024. I now have got AI Assistant working in Studio 2024. At the booth, I was told that I could create prompts that would find-and-replace target text based on instructions, e.g. "Wherever the subscripts in the target text for chemical formulas don't match the source text placement, change the subscripts in the TT to match the ST." However I cannot figure out how to do this. Even simpler queries like "Ensure that the verbs 'comprendre' and 'comporter' are always translated by the grammatically appropriate form of 'comprise'" seems beyond my reach. Can anyone help me out with figuring out how to run these prompts on the text? Thanks!

Regards,

Andrew

emoji
Parents
  •  

    It will work easily one at a time.  Your question title says "all"  but the content of the thread suggests it won't work at all.  So first the working at all:

    If that was your problem I hope it helps.  If you actually want to use AI to find and replace "All" then I'm afraid that isn't possible yet.  We'd need to have a batch task capability to do this which we don't have yet, and as you saw in the video it might be a problem anyway as it's not always right!

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji
  • Hello Paul,

    Thanks for the bespoke video to show off the capability. It does look like this will be a useful prompt but it's unfortunate that we cannot do a batch task to fix a recurring error. Hopefully in some future version.

    -Andrew

    emoji
  •   

    You can do a batch process! "Pretranslate with OpenAI provider for Trados". AFAIK you can't make these changes tracked - that would be a major advantage!

    emoji
  •  

    I didn't test this, but I think that may not be the same thing.  That approach would replace the entire translation with the OpenAI version of it as opposed to just correcting the subscript.  If you tested it and found it only corrected the subscripts then that's a major plus I reckon we got "not by design!"

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji
Reply Children
  •  

    I tested it in connection with the AI terminology question here:  Terminology-aware machine translation in Trados 2024

    I can translate something, then ask AI to sort out all the terminology without changing anything else in the translation - as a batch task. It works great. The more synonyms you have, the less glossary-style your TB is the easier it is for AI to select the wrong term. Might be my bad prompting, but it is certailnly partly because the AI Assistant does not convey any metadata to OpenAI.

    In this test, I have terms with synonyms, and I ask AI to insert the term that is most likely to fit, depending on context, length and specificity. Alphanumerics are to be handled separately.

    So if you have a suitable prompt, you could let AI edit your translation. A very useful feature would be to let AI do this in tracked changes mode. That would be incredible. Imagine telling AI to check for incomplete sentences and complete them - with tracked changes. And maybe even add the comments that it can already produce now to the change (a bit tricky) or at least to the segment containing the change(s).

    Daniel

    emoji
  •  

    I tested it in connection with the AI terminology question here:  Terminology-aware machine translation in Trados 2024

    Indeed... but this is a different usecase.  Andrew wants the prompt to make a change to an existing translation unrelated to terminology, rather based on the prompt used for interactive work to adapt or provide a translation based on AI alone.

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji
  •  

    Yes, but there is no reason to believe that this would not work as a batch process.

    emoji
  •  

    I love your optimism ;-)  But I would be surprised.  It certainly wasn't designed to work that way.

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji
  • I love your optimism ;-) 

      I am normally likened to Eeyore.

    That aside, could you attach your sample file and a text version of your prompt from the above video? Let's find out.

    emoji
  •  

    Sure... file:

    subscripts.md_en-US_fr-FR.zip

    Prompt:

    Compare the use of subscript numbers in the source with the use in the already translated target.  If there is a difference report and offer a correction.

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji
  • ,  

    I got it to work with your AI prompt, Paul:

    emoji
  •  

    Well... hats off to you!  I still cannot make this work.  One interesting thing I noted in your demo is the before and after.

    Before:

    Screenshot of a text editor with chemical equations and status labels. All lines are marked with 'NMT' indicating a status change, with a software version 'Markdown v1.0.0' at the bottom.

    After:

    Second screenshot similar to the first, showing a text editor with chemical equations. Each line is labeled 'NMT' next to a blue vertical bar, and 'Markdown v1.0.0' is visible at the bottom.

    The statuses are now all NMT, including the ones that had nothing wrong with them.  I am a bit suspicious of that.  I do think it needs more testing, and I will discuss with the developer.  I don't think we'll get back to this for a while as we don't really want to touch this anymore until the AI Assistant part has been separated and is part of the core product.  But the batch task requirement is noted.

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji


    Generated Image Alt-Text
    [edited by: RWS Community AI at 9:15 AM (GMT 0) on 14 Nov 2024]
  •  

    I should have told you that I edited your prompt. I found yours was not directive enough. Mine is:

    “Contrary to what I told you before, you are not a translator, but an editor. Compare the use of subscript numbers in the source with the use in the already translated target. If there is a difference, correct the target according to the source.”

    Why are all segments now marked NMT? Well, I'd say OpenAI touches each one of them and this causes the different origin. I think it reads each segment, applies AI, then writes the result. Even if the result is “don't change anything”, it still writes the unchanged segment. Your developers would be able to confirm that (or refute it).

    What I am struggling with is a way to make it obvious to the translator what was changed. I am using Post-Edit Compare, which works but is a bit cumbersome to set up. If I could manage to get Trados to do its replacements in the review mode, that would be a dream come true. But thinking about what you noticed, it might well be that the AI Assistant replaces the entire segment content with its output (changed or not), in which case the track changes mode would not be very helpful.

    In my other AI test project where I try to get AI to insert the correct synonym, I get this with Post-Edit Compare:

    Screenshot of a translation software interface showing a comparison between original and updated German translations of English source text, with match percentages and edit distances.

    I found that the AI Assistant reliably keeps its fingers off translated text if I tell it explicitly that it is not a translator but an editor, but I still like to be sure. (See segment 9!)

    The aim here is to be able to ask AI to apply a certain style to an entire text, such as making sure that British English is used throughout, and no US idioms. I'd definitely need to know what got changed.

    Daniel

    (Sorry, tried to update the screenshot but wasn't allowed to, so deleted the post and re-wrote it. Screenshot now has proper files as “original” and “updated”. Was very confusing before.)

    emoji


    Generated Image Alt-Text
    [edited by: RWS Community AI at 10:45 AM (GMT 0) on 14 Nov 2024]
  •  

    Why are all segments now marked NMT? Well, I'd say OpenAI touches each one of them and this causes the different origin. I think it reads each segment, applies AI, then writes the result. Even if the result is “don't change anything”, it still writes the unchanged segment.

    Seems unlikely and if I could make this work I could check it easily enough in two ways:

    1. add your initials to the start of each target segment before you run it and see if the result removed the initials, or
    2. use fiddler to see what gets sent and what comes back

    I already know this was never designed to work in this way, the developers would confirm this, so what we're testing here is almost a case of good bugs/bad bugs... and this may be a good bug ;-)

    Paul Filkin | RWS Group

    ________________________
    Design your own training!

    You've done the courses and still need to go a little further, or still not clear? 
    Tell us what you need in our Community Solutions Hub

    emoji