Hi list, I just noticed that the two lexical weights have switched position in the translation model sometime in the last few months. Originally, it was p(s|t) lex(s|t) p(t|s) lex(t|s), now it is p(s|t) lex(t|s) p(t|s) lex(s|t).
This is most noticeable when looking at phrase pairs where t (or s) only occurs once: I ||| Ursprünglicher ||| 1 0.0002686 0.00025641 1 2.718 ||| 0-0 ||| 1 3900 1 the order of features is not too important when you optimize after training anyway, but for the interpretation of the models, and other scripts interacting with it, I'd prefer to have the old order back. Did the regression tests catch this flip, or did it happen before? Rico _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
