Hi, the motivation for backward probabilities is the noisy channel model that reformulates p(e|f) into p(e)*p(f|e). The phrase penalty feature value is exp(1), because this way it simply counts the number of phrases used. You are free to change anything you like, and you may even get improvements that way.
-phi On Thu, Jan 15, 2009 at 12:50 AM, K.Taraka Rama <[email protected]> wrote: > Hi, > > What are the 5 weights in translation model in "mert" meant for ? Barry > suggested that they can be the probabilities found in > http://www.statmt.org/moses/?n=FactoredTraining.ScorePhrases > > Forward and backward phrase translation probability, lexical weights and > phrase penalty. > > But in the problem formulation we only take the forward probabilities into > consideration. What is the reason for having 5 feature functions? > > The phrase penalty always seems to be exp(1). What is the reason for it? Can > I change it? > > -- > With Regards, > Taraka Rama. > > > _______________________________________________ > 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
