Hi, I would go about it this way.
1. Tag the words that are assumed to be translated correctly in the target. 2. Align them to the source words using the reported word alignment. 3. Specify translations for correctly translated source word spans with XML markup. For you example, the end result should be: <x translation="the">la</x> gran <x translation="house">casa</x> roja You may want to throw in <wall/> reordering constraints, when suitable. Check the following paper, where TM matches were converted into such markup: http://homepages.inf.ed.ac.uk/pkoehn/publications/tm-smt-amta2010.pdf -phi On Mon, Nov 30, 2015 at 10:13 AM, Miguel Domingo <[email protected]> wrote: > Hello, > > I know that you can bring external knowledge to the decoder with *XML > Markup*, but this seems to be true only for the source side of the > decoder. I was wondering if there is something similar that could be used > on the target side. > > My idea was using moses both for translating and post-editing a sentence, > so that once the sentence has been translated you could manually select > those words that have been translated correctly and get a new translation > of the sentence (using the new knowledge). > > As an example, imagine we want to translate the sentence in Spanish *la > gran casa roja* (*the big red house*), and running moses' decoder we > obtain the hypothesis *the small green house*. We want to mark those > words that are correct (*the, house*) and obtain a new hypothesis in > which those words are left untouched (e.g. *the big green house*). > > Is there any way to achieve this? > > Thanks in advance, > > Miguel > > _______________________________________________ > Moses-support mailing list > [email protected] > http://mailman.mit.edu/mailman/listinfo/moses-support > >
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
