prunes non-initial rules (i.e. rules with non-terminals) from the rule table
based on their frequency counts. In Zollmann, Venugopal, Och, and Ponte (2008),
pruning hierarchical rules that occur only once was found to significantly
decrease rule table size without harming translation quality.
Also, add TUNING:filter-settings and EVALUATION[:<set>]:filter-settings
variables so that this can be enabled in the EMS.
git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@4033 1f5c12ca-751b-0410-a591-d2e778427230
an unknown word label file has been generated. Also, disable this option
by default since it can greatly increase number of hypotheses generated
and hasn't been shown to help translation yet.
git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@3277 1f5c12ca-751b-0410-a591-d2e778427230
