Hi, there is no specific existing feature function that allows for exactly that.
The closest approximation would be to use the WordTranslation feature for this factor. This would learn binary features for each mapping of one annotation to the other, hopefully learning to prefer some mappings (where they agree) to others. You can set this in EMS with: sparse-features = "word-translation all factor 1-1" where "1" is the factor number for your annotation. If you want to implement exactly what you describe, you should look at the word translation feature, and write a very similar feature function. -phi On Fri, Jul 24, 2015 at 6:03 AM, Marco Damonte <[email protected]> wrote: > Hi everyone, > > I'm using EMS to run experiments involving semantic annotations as > factors. I would like to try adding sparse features to use the fact that > when a word and its translation have the same annotation, there is good > chances that it is a good translation. > > I read the tutorials on Moses website but it's still not clear to me how > this works. Can anyone help me with this? > > Regards, > Marco > > _______________________________________________ > 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
