Hi,
When tuning using sparse features with PRO I am getting quite an 
impressive increase on the devset (compared to MERT without sparse 
features), but very inconsistent results on the test set with 
differences of several percent points between repeated experiments. I 
guess this is due to overfitting on the tuning set. The megam.opt binary 
seems to use regularization by default, but can be also be adjusted. Has 
anyone ever experimented with that?
Thanks,
Marcin
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to