How does NMT learn from our translations, or does it?


I have used NMT for almost two months now and I can confirm that it is far superior to the old AdaptiveMT. Wow!

One thing that is puzzling me though, is how does NMT learn from my translations, or does it?

I have the "Update" box ticked next to "SDL Language Cloud" (in Project settings\All language pairs\Translation Memory and Automated Translation) but the NMT engine doesn’t seem to ’learn’ anything from my translation. For example I am working on a large translation and the term "OECM" is being used hundreds of times, and yet, the NMT keeps giving me the wrong acronym in French even though I correct it each time.

In a way I am not too surprised that the NMT isn’t learning properly as this type of AT no longer uses Adaptive engines. However, I am wondering what purpose the ticked "update" box serve. Is it just used to feed back to the SDL database to improve the NMT at a later stage? It would be interesting to know.

Now the big question: is there any scope for NMT to work with Adaptive engines in the future? (apologies if this question sounds silly but I have no idea how NMT is programmed...).

If NMT could learn from my own translations I can’t begin to imagine what tremendous potential this would have, especially combined with Termbases and TMs.

By the way, if you set up a beta group to test NMT from English into French, you can count me in!


Parents Reply Children