Thanks for tips! Yes, I do want to use the same pruned tables for tuning and testing. Currently, I do have only one translation table.
KB On Tue, Aug 21, 2012 at 10:13 PM, Philipp Koehn <[email protected]> wrote: > Hi, > > there are several places where you could do this, depending > on what your goal is. > > If you want use the same pruned phrase table for tuning > and testing, it makes sense to add such a step into the > model training part. > > Currently, there is a phrase translation table step: > > build-ttable > in: extracted-phrases lexical-translation-table domains > out: phrase-translation-table > > which feeds into the configuration file creation step: > > create-config > in: reordering-table phrase-translation-table generation-table > sparse-lexical domains INTERPOLATED-LM:binlm LM:binlm > out: config > > So, you could put something inbetween, but changing the > input of the create-config step into: > > create-config > in: reordering-table pruned-phrase-translation-table > generation-table sparse-lexical domains INTERPOLATED-LM:binlm LM:binlm > out: config > > and creating a step > > prune-phrase-table > in: phrase-translation-table > out: pruned-phrase-translation-table > > However, this may get a bit messy when you have multiple phrase translation > tables (due to factored models, for instance), and lexicalized reordering > models > which have to have the same phrase entries as the phrase translation table. > In that case, it would better to place the pruning step after the > "create-config" > > prune-tables > in: config > out: pruned-config > > and then have the subsequent steps take pruned-config as input. > > -phi > > > On Tue, Aug 21, 2012 at 12:55 PM, Karel Bílek <[email protected]> wrote: >> Hello everyone, >> >> I have a question about EMS. >> >> I want to prune the phrase table right before TUNING step. I already >> do have a script for that; but I don't know where to "plug" it into >> config EMS file. >> >> I can do the pruning "by hand" and then write the address directly >> into "phrase-translation-table" but that seems to be against the >> spirit of EMS. >> >> Thanks, >> Karel Bílek >> >> _______________________________________________ >> 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
