Hi Tom Yes, you should be able to reuse GIZA++ runs for training a system to translate in the opposite directions. I've done this before, although I use the experiment management system (ems), not train-factored-phrase-model.perl
Just remember to swap the giza and giza-inverse alignments. The final alignments are contained in files with names like fr-en.A3.final.gz - these are the only ones you need. regards Barry On Sunday 30 May 2010 10:58, [email protected] wrote: > In training a FR->EN model, train-factored-phrase-model.perl creates > GIZA++ (./giza.fr-en) and inverse GIZA++ (./giza.en-fr) directories. > > I also trained an EN->FR model model to a different root folder resulting > in GIZA++ (./giza.en-fr) and inverse GIZA++ (./giza.fr-en). > > Rather than go throught second training, is it possible to reuse first two > directories instead of retraining? It would seem that the -giza-f2e and > -giza-e2f parameters, along with the -first-step parameter picking up at > step 3 would allow this to work. > > However, I used diff to compare the corresponding two ./giza.en-fr > folders. There are difference and I don't know they prohibit effective > re-use. > > Thanks, > Tom > > _______________________________________________ > 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
