The gift that keeps on giving

There’s a lot of buzz around ChatGPT. Most people, including myself, don’t even want to hear the word anymore. Gurus pop up everywhere, claiming they are experts and charging thousands of euros for teaching you how to “talk” to this language model. “Prompt engineering”, they call it. Many of these are false claims. But here’s Yasmin Moslem, an outstanding AI researcher and machine translation specialist, who once more comes up with a method to improve the quality of machine translations generated by ChatGPT in a simple, beautiful and very elegant way. The feature was implemented in Cattitude 1,5 hours later. No other CAT tool can match that implementation speed.

Here you could already read how we succesfully implemented machine translation by ChatGPT at a lightning-fast speed. Today we outdid ourselves by improving on this 1,5 hours after the newest research in this field was published. So what’s the deal?

All warm and fuzzy

We already could generate pretty good machine translations by feeding ChatGPT our best 10 fuzzy matches (or less but as many as possible if the prompt length does not allow it), but ChatGPT didn’t always adhere to our terminology database. Of course not, as it doesn’t know how our client wants us to translate certain terms. The research paper above shows that this problem can be solved very easily by feeding ChatGPT a short terminology list besides the fuzzy matches that were already being fed. And indeed, this works. Look at the following:

Here we see a very so-so machine translation by DeepL (ModernMT’s suggestion was even worse: you can see it crossed out in the top row of the yellow square on the bottom). The following is ChatGPT’s take after only feeding it fuzzy matches:

Better already, but this is as good as it gets in this case. If we merely rephrase the first part to Ga als 6e over de rondelijn en, we’re done. That’s a lot less work than typing the entire sentence from scratch. Now in this particular case ChatGPT got the terminology right, but suppose we want it to translate car as autootje (carlet) instead. This is of course pure nonsense, just to illustrate our purpose.

This is the result after telling ChatGPT explicitly that we want it to translate car as carlet:

As you can see, it neatly translated car as little car (autootje), exactly as we wanted.

I’m sure you can think of a thousand reasons why this may be very handy in your own use case scenario.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.