Hrm. Your setup + commands look normal enough, but the result is not, so it's probably starving for memory. Have you checked on the memory usage during your failed training attempts?
You could also check by building a 4gram model on the full corpus, and a 5gam model on half the corpus. If these run quickly with no trouble, then I'd guess it's that a full 5gram LM is too large for your machine. If that is the case, then there is not much you can do. However, for the Europarl corpus, a 4gram model is sufficient, as a 5gram model does not provide any significant improvement in the 1st-pass system output. There is a table in the UW system description paper in the WMT-08 workshop that shows this :) ~amittai _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
