good idea. it would also be interesting to constrain a weight based on another weight. eg, the weights of phrase-table A = weights of phrase-table B
On Tue, Jan 24, 2012 at 3:27 AM, Thomas Schoenemann < [email protected]> wrote: > Hi! > > That's also the impression I got from a more thorough code review today. I > will work on implementing a positivity constraint, without changing the > default behaviour (in my own branch). > > Best, > Thomas > > ------------------------------ > *Von:* Barry Haddow <[email protected]> > *An:* Thomas Schoenemann <[email protected]> > *Cc:* "[email protected]" <[email protected]> > *Gesendet:* 19:00 Montag, 23.Januar 2012 > *Betreff:* Re: [Moses-support] Parameter ranges in MERT > > Hi Thomas > > As far as I recall, the parameter ranges were used for the random > restarts. These are sampled uniformly from the given range. > > cheers - Barry > > Quoting Thomas Schoenemann <[email protected]> on Mon, 23 Jan > 2012 09:42:38 +0000 (GMT): > > > Hi everyone, > > > > can anyone tell me if the current MERT implementation should respect > parameter ranges? It does require to input them, but to my knowledge these > ranges are ignored. > > > > > > Except for the word penalty I think all weights should be positive. Yet, > I often get negative ones from MERT, and they are not always corrected in > later iterations. > > > > Cheers, > > Thomas (currently University of Düsseldorf, Germany) > > > > > > --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 > >
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
