Hi Barry, Thanks for the reference. I found why it can happen: those positive translation scores occur when some weighted feature components happen to be positive numbers and other negative weighted features are not small enough to sum up to a negative final score. (All the feature values but pharse penalty are log probabilities thus less than or equal to 0. If a feature weight tuned by MERT or whatever is a negative number as well, the result is a positive number.)
Best Regards, Rasul. Quoting Barry Haddow <[email protected]>: > Hi Rasul > > Positive scores are possible, but unusual, depending on the feature > weights. The scores can be interpreted as log probabilities, but they > have to be nornalised first. > > The standard Moses feature functions are explained here > http://www.statmt.org/moses/?n=Moses.FeatureFunctions > > cheers - Barry > > On 07/08/12 12:01, [email protected] wrote: >> Hi all, >> >> Is it possible to have a positive number for weighted overall decoder >> score? I previously thought that those are log probabilities thus >> negative. However, in a few n-best list, I found positive overall >> scores (the last score in each line) for some translations. This >> happened with both phrase-based and chart decoder. >> >> It would also be great if anybody direct me to the documentation about >> these score and how they are computed (not in general but specifically >> by the Moses decoders). I couldn't find by search. >> >> Thanks and regards, >> Rasul. >> _______________________________________________ >> Moses-support mailing list >> [email protected] >> http://mailman.mit.edu/mailman/listinfo/moses-support >> > > > -- > The University of Edinburgh is a charitable body, registered in > Scotland, with registration number SC005336. _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
