It's now easier than ever to optimize Smartcat for increased quality and throughput using a variety of generative AI tools built right into the platform.
In this interactive, hands-on session the Smartcat product team will:
➡️ Walk through the core Gen AI capabilities now available in the Smartcat AI platform
➡️ Dive deep into the nuts-and-bolts of optimizing Smartcat for different scenarios
➡️ Address questions and provide detailed guidance to help you utilize these AI tools with your own content
Webinar Q&A
Q: Is this only available with the Unite subscription plan?
A: I would suggest that you contact your customer success contact and ask about your subscription.
Q: If you would like to translate an XLIFF file, is using the AI prompt an added value? Or should we stick to the current model?
A: It could be an added value as shown by Andy if you want to generate some very specific terminology or style. We'll show you what else can be done in a moment.
Q: Is there a list of domains available?
A: LLMs cover most domains. They are trained using billions of documents.
Q: Will it be possible at some point to link the glossary prompt to specific glossaries instead of all glossaries available in a given project? My goal is to reduce false positives by asking the model to use a specific glossary but still keep reference glossaries in the editor for posteditors to check other terminology, such as polysemic terms, general terms, synonyms, etc, that I don’t want the model to use for translation in all cases.
A: Good question. At this time, we use glossaries associated with the project. One option in your case could be to fine-tune an engine using your data. But we can explore your suggestion.
Q: Is there any limitation on how long the prompt can be in case I want to be very specific with style, terminology, standard phrases, etc.?
A: Technically no. But this could slow down the translation process. If you want to impact the general output, fine-tuning an engine might be better suited.
Q: Can we add parallel texts as references, so the AI picks style from them?
A: You can fine-tune an engine where you can provide sentence pairs as a reference and the engine will try to match your style.
Q: When we run a translation using a prompt/the LLM, does it use the number of words in our subscription? Or how does that work?
A: Yes, words generated using a prompt of the LLM will consume Smartwords.
Q: Can you explain the significance of using curly brackets? Are these system terms that will only work if you type the word exactly as shown?
A: As shared by Jean-Luc, we have a small library of pre-configured prompts, which you can find on the right-hand side of the Prompt-generation UI.
Q: At the moment, in our projects, the generative AI preset with prompts does not work for strings that have tags (yellow tags for mark-ups, whenever there is formatting in the source). In these cases, the preset automatically falls back to another translation machine. Do you know of a workaround and how the generative AI preset works on strings with tags, too?
A: As shared by Jean-Luc, let’s look at some specific examples of the tags, and see if we can refine the prompt to ignore tags in strings
Q: Does it work as well if we are working with languages with cases?
A: Yes. So there are so-called low-resource languages. So the models, usually the models when they're trying to take the data, and obviously the most. The vast majority of the data is English data. That's why the llam's the language models show the best results in English. But as you try different other different languages. You would see the quality degrading and the more rare language you're you're trying to translate. Probably the worst results would be but the languages with case are not necessarily the low-resource languages. So yes, it works well. But as you go with one of the low resource languages to mention would be Armenian language, and one of the complex languages out there. You would. You might probably see some, some unexpected results. But you usually you would. You would see a good results for the for the cases.
