The furrfect solution

Cattitude has always been at the forefront of technology. While other CAT tool suppliers struggle to keep up with the times, in Cattitude, important developments in translation technology are often implemented within days or even hours. Today is no exception. Cattitude just added a new feature based on the findings of Yasmin Moslem, Rejwanul Haque and Andy Way, who published a paper on Adaptive Machine Translation with Large Language Models.

One of the findings in said research is that showing ChatGPT 1-10 domain-specific translation pairs similar to the sentence to be translated (also known as fuzzy matches) can improve the translation of new sentences immediately. This is especially useful for high-resource languages.

Now, of course it’s very well possible to painstakingly copy your fuzzy matches and the sentence you’re currently working on to ChatGPT manually, but by the time you’re done, you could have probably come up with a translation yourself already. And as you know, machines are not here to threaten or delete us: they’re here to help us so that we can move faster, translate faster and therefore earn more money. Speed is key and that’s what we should use machines for.

Enter Cattitude. See that little button on the top? It copies whatever information ChatGPT needs to come up with a good suggestion for your current segment (as good as it technically possible at this moment, at least). Here is the result of one of the first tests with a mere 1-shot, but meanwhile a 5-shot approach has already been implemented:

The best of all worlds

It would be ideal to feed this information to ChatGPT automatically, and code-wise this is no issue or whatsoever, but as ChatGPT’s API is still very slow when it comes to generating responses, it’s currently quicker to just copy the info straight to ChatGPT on your second monitor and see what it comes up with. This is especially handy for longer segments that involve lots of manual labor. When it comes to shorter translations, the old route is still quicker.

In the future, when ChatGPT’s API has become a bit quicker, it’s very easy to automate the pipeline further and present ChatGPT’s suggestions straight in Cattitude’s interface.

Often I find myself making split-second decisions between using my own brains, ModernMT, DeepL and Amazon Web Services, and often the end result is combination of multiple models. ChatGPT and DeepL are great for longer and wordier segments, ModernMT is great for sentences with lots of jargon and sometimes Amazon Web Services can come up with that magical piece of the puzzle you just couldn’t remember. And yes, sometimes all of them produce nothing but garbage. In the end it’s us, the translators with their human brains, who need to make split-second decisions about which approach is best in which situation, and that is still something fortunately only humans can do.

I would like to thank Yasmin Moslem in particular for her indispensable help with developing this great new feature.