Yulia Kovalskaya 

Yulia Kovalskaya said:

the working principle is the same as with the DeepL glossaries

Well, "jein", as we might say in German. Yes and no. At this point, the AI is just given the recognized terms, without any metadata. Insofar yes, your termbase is in a way reduced to a glossary. I hope this will change and there will be a way to choose other fields from the termbase to be handed to the AI.

If you prompt the AI to give you several suggestions, you will see that it varies which of the synonyms it uses. And you can instruct it to use a fitting synonym. Insofar, no, it's not like a DeepL glossary. That will always slot in the first fitting term it comes across, even if it makes little sense. With the AI, you can instruct it how to select a synonym. You might have to custom tailor your prompt for a particular translation to tell AI what the context is, but even a general instruction works. See screenshots below, I created a termbase with two translations for the English word "stain", and gave it two sentences. I also instructed AI to explain its choice.

In the first sentence, MT got the correct term, which is accepted by AI. In the second sentence, MT got the wrong term, and AI corrects it:

This is just a quick test. I am sure this could be refined substantially.

Daniel, hello!

In the first place, thank you for the answer! As far as I can understand, this is the interface of the AI Assistant in Trados 2024... And once, the MT result is corrected I have to click on it to insert it in the segment, right?

ANd then the prompts...they are written for each term?? Or a kind of a list of prompts can be created to "train" AI???

Yulia Kovalskaya 

a) Yes, that is the AI Assistant in Trados 2024.

b) The way I set it up, you have to click to insert it. But you could automate this and use it as a batch task. As Paul pointed out, the current AIs are fuzzy – that's their glory and their weakness. So you will want to look at what it's doing, at the very least check the results.

c) One prompt, hastily written in this case. It makes sense to refine prompts. I could imagine customizing a prompt for a project, like “The text you are working on is about cleaning textiles.” If you can give the LLM relevant information, by all means do.

The one serious weakness I see is how forbidden terms are handled. They, too, are just passed on to AI as relevant terms. If your termbase has a lot of those, you might have to create a separate termbase to get rid of those. In the above example, if I had "Klecks" as forbidden term, I doubt AI could reliably select "Fleck". All it get is a list like: "stain-Fleck, stain-Klecks, stain-Beize". Across the board, handling forbidden terms is a topic that urgently needs attention IMHO.

But as for your main question, AI's terminology-aware translations are seriously above a DeepL-style glossary if you prompt it well. As I said in other posts before, I believe this is a big game-changer technology for this industry, just like the ascent of MT was ten years ago.

Would be worth checking whether a plain terminology-redacting task requires GPT-4o, or whether it can run on a simpler, faster and cheaper model like 3.5turbo. There is so much that could be done!

Daniel

Dear Daniel, 

sorry for the late reply. Very grateful for the explanations, got it: we have to "train" the AI Assistant to make it apply out Multiterm terminology properly. Until I have the version and test it, I will not see how it actually works. 

However, what I see on your screenshots is that the MT modification is performed in a separate window below, not directly in the segment. That could be less "comfortable and more time consuming. However, we have to give it a try. 

Best regards, Daniel!

Dear Daniel, 

sorry for the late reply. Very grateful for the explanations, got it: we have to "train" the AI Assistant to make it apply out Multiterm terminology properly. Until I have the version and test it, I will not see how it actually works. 

However, what I see on your screenshots is that the MT modification is performed in a separate window below, not directly in the segment. That could be less "comfortable and more time consuming. However, we have to give it a try. 

Best regards, Daniel!