Hi, On Tue, Feb 28, 2012 at 2:43 PM, Dennis Mehay <[email protected]> wrote: > > Hi all, > > Sorry if this question is answered somewhere, but I can't seem to find what I > need in the archives or in the moses documentation (or in my perusal of the > relevant files of the code). > > What I want to know is what the scores in the lexicalized reordering tables > correspond to (precisely, I mean -- I know they are smoothed MSD relative > frequencies). > > What I mean is, does: > > blah ||| blee ||| 0.1 0.2 0.3 0.4 0.5 0.6 > > mean that blah<->blee is 0.1 likely to remain monotone wrt the previous > phrase (backward monotone relative freq), or wrt the next phrase (forward > monotone relative freq)? Does 0.5 refer to a swap relative frequency wrt the > previous phrase or the next? Is 0.6 'other' (discontiguous) wrt the previous > or next phrase, etc., etc.? It seems to me (from actually computing relative > frequencies) that the scores are: > > blah ||| blee ||| <mono-backward> <swap-backward> <other-backward> > <other-forward> <swap-forward> <mono-forward>
Looking at the code in scripts/training/lexical-reordering/reordering_classes.cpp, specifically Model::score_fe, it seems to be first with respect to the previous phrase (is that what you mean by backward?), and then with respect to the next phrase. The order "mono, swap, other" is correct. > but that doesn't make much sense. I am sorry that you feel that way. > Oh, and while I'm on it, does anyone have an opinion on whether "word-based" > reordering extraction works better or worse or is generally indistinguishable > from "phrase-based" reordering extraction. I think there was some work on this by Richard Zens, who also built discriminative reordering models. I do not have a definite answer to this, I never tried it myself. -phi _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
