Predictive translation server


(Etienne Monneret) #1

It would be a great evolution if the translation server would accept an extra “begining” parameter for a sentence, to return with one (or N) possible way(s) of ending its translation. As far as I understand the way it’s working, I think this could be done by forcing the first tokens in the beam search… right ?

PS : a bit harder would be to get a predictive post-editing server = knowing a sentence to translate, and an already existing translation with a pointed position in it, in the beam search force all beginning and ending tokens, and let free exploration to fill the pointed position at best. The hard part is certainly to force the ending tokens in the beam search. The first above proposition of predictive translation server would be equivalent to set the pointed position at the end of the translation in this predictive post-editing server.

Lexical constraints at translation request time?
Adaptive Learning in OpenNMT
(Etienne Monneret) #2

DeepL did it : its editable target zone (with the keyboard, not only the mouse) is absolutely incredible ! Word completion and sentence auto-rewriting !
Wonderful !

(Terence Lewis) #3

Congratulations!. I know Kilgray are pushing for their MT providers (of which I am one) to offer something along these lines in memoQ 8.2. Are your developments a State secret? :slight_smile:

(Etienne Monneret) #4

Not sure to understand your question. If you want some details about me, just have a look at my profile on this forum, with a click on my icon.

In short : for now, in spite of the fact that I really would enjoy it, I don’t have the time to code in tools like ONMT. I’m just making integrations with our own system… where of course we are doing much more than only MT… of course also including some of our own technologies.


(Terence Lewis) #5

Sorry, my question arose because I misread your original post and didn’t notice you were referring to the DeepL website. I’ve just tried it myself and agree it is impressive.