Hi, if you have the computing resources, then running multiple tuning runs will give you more reliable results, regardless of use of MERT or k-best MIRA.
-phi On Mon, Sep 16, 2013 at 10:08 AM, Thomas Meyer <[email protected]> wrote: > Thanks, Philipp, I'll try that. > > As a follow-up question, I'm using batch-mira for tuning with these > features. Should I here also run several tunings for my systems and average > the scores when evaluating as is nowadays done with MERT to account for > instability? > > Thanks again, > Thomas > > > > On 16/09/13 10:55, Philipp Koehn wrote: > > Hi, > > one way to do this is to produce an n-best list of size 1, which then > reports a detailed score breakdown. > > -phi > > > On Mon, Sep 16, 2013 at 9:36 AM, Thomas Meyer <[email protected]> wrote: >> >> Hi, >> >> I started experimenting with sparse lexical source word deletion >> features in Moses. All works fine, but can I actually see somewhere the >> deletions that are made to the source texts? >> >> I am using the EMS with decoder verbosity set to 0. Maybe it's enough to >> just increase the verbosity level? >> >> Many thanks, >> Thomas >> _______________________________________________ >> 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
