Hi, in the experiments that I did, I found that kbmira is a little bit more stable (i.e., less variance between tuning runs) than MERT. It runs pretty much always to the maximum of 25 iterations.
-phi On Mon, Feb 24, 2014 at 10:01 AM, Marcin Junczys-Dowmunt <[email protected]> wrote: > Thanks, > what's your general experience with pro/kbmira and BLEU. For my metric, > it seems both methods do not really converge too well, although they > find improvements from time to time. Do you have a similar experience > with BLEU? If yes, do you let the optimization hit the 25 iterations limit? > Best, > Marcin > > W dniu 23.02.2014 17:55, Barry Haddow pisze: >> Hi Marcin >> >> It's mentioned here: >> http://www.aclweb.org/anthology/W/W12/W12-3139.pdf >> (section 4.2) but it's not something we spent much time experimenting >> with, >> >> cheers - Barry >> >> On 23/02/14 15:12, Marcin Junczys-Dowmunt wrote: >>> Hi, >>> Is there a publication that describes the idea behind >>> "--pro-starting-point" option in mert-moses.pl? This seems to be working >>> really well for me. >>> Best, >>> Marcin >>> _______________________________________________ >>> 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 _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
