Hi, when using the --pro-starting-point option, you must not specify the --pairwise-ranked parameter, or undefined things happen.
When using --pro-starting-point you may also want to try out --historic-best (which uses all prior PRO starting points and best found points). -phi On Tue, Sep 27, 2011 at 11:42 AM, Fabienne Braune <[email protected]> wrote: > Hi all, > > I tried to use MERT with the --pairwise-ranked option but the length > parameter seems to be strangely tuned : If I take the weights obtained with > PRO and use them on the test set I get a bad BLEU score and bad length ratio > (0.96...). > > I would like to combine PRO with line MERT : > - I tried to use the option --pro-starting-point. When I enter this option > before --pairwise-ranked on the command line, then the weights and log files > I obtain look exactly like tuning with --pairwise-ranked only > -When I use the option --pro-starting-point alone then I get very different > feature weights and log files then when using --pro-starting-point with > --pairwise-ranked. > > Does someone know if I am using the right option? > > Thanks a lot for any answer! > > Cheers, > Fabienne > > _______________________________________________ > 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
