Hi Taraka >From the mert.log, I can see that LENGTH is still set to 4, since the number of bleu statistics is 9, which is 2*LENGTH+1. You need to run 'make clean all' for the change to take effect.
You should also check that you have an up-to-date version of mert-moses-new.pl, since an earlier version did not handle feature weights with exponentials. Your feature weights and bleu scores look very strange though Best point: 1 2.23227e-40 1.87459e-40 5.10829e-41 7.07497e-40 -1.83438e-40 -8.1584e-40 6.61028e-40 => 0.96344 I don't know how you get such a high bleu score and such low weights. Is there anything odd about the data? Did you get similar outputs from earlier iterations of mert? Maybe you could post the decoder output and reference file, regards Barry On Monday 12 January 2009 21:17:23 K.Taraka Rama wrote: > Dear Barry, > > Yes, I changed the LENGTH variable. I am attaching the mert.log file. > > On Tue, Jan 13, 2009 at 2:34 AM, Barry Haddow <[email protected]> wrote: > > Dear Taraka > > > > Answers inline > > > > On Monday 12 January 2009 20:29:46 K.Taraka Rama wrote: > > > Dear Experts, > > > > > > I have the following doubts regarding the new mert implementation: > > > > > > 1. I know that the mert implementation is based on the Och's paper > > > of Minimum Error Rate Training and Discriminative models for SMT, which > > > tries to optimise the feature function's weights. In Moses website it > > > is written that there are 5 translation weights. What do the 5 > > > translation feature functions stand for? > > > > Forward and backward phrase translation (log) probabilities, forward and > > backward lexical translation weights, and a phrase penalty. > > http://www.statmt.org/moses/?n=FactoredTraining.ScorePhrases > > > > > 2. What are the default weights for the feature functions? > > > > These are usually set to 0.2 by train-factored-phrase-model.perl > > > > > 3. I have the following problem when I run the new mert script. > > > > > > "Failed to parse mert.log, missed Best point there. at > > > > /home/taraka/moses/bin/moses-scripts/scripts-20081123-1240/training/mert- > >mo > > > > >ses-new.pl line 638." > > > > > > I have tried to change the value of "n" in Bleu score function and > > > recompile it. It worked for the first time. It didnt give any error. > > > The next time I did the same thing and it started giving the above > > > error. I tried with "--continue" switch, still it didnt change any > > > thing. > > > > I take it you mean that you changed the LENGTH variable? > > Possibly mert crashed. Could you post the mert.log file? It should be in > > the > > working directory. I don't think the --continue option works (yet) with > > the new mert. Make sure you use a clean working directory for a new run, > > > > regards > > Barry > > > > > I believe there is no problem with the data, because I was able to run > > > > the > > > > > old implementation of MERT on it and also test the models. _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
