Hi, I have the same experience, that the lexical reordering model does not always improve over only using the distance based model. I really think it depends to a large extent on the language pair and the corpus. Of course there is also the uncertainty of Mert, so it might be worthwhile to run several Mert runs if you haven't done that already.
I was in a team at the Dublin MT Marathon who worked on extending the reordering models in Moses. Galley and Manning (http://www.aclweb.org/anthology/D/D08/D08-1089.pdf) pointed out that the standard reordering model in Moses is inconsistent, since it is based on word alignments at training time, but on phrase alignments at decoding time. Our group implemented both a consistent phrase-based model, and the hierarchical reordering model suggested by Galley and Manning. This code is in the Moses trunk, but the Moses webpage has not been updated. There is some documentation on the MT Marathon wiki, however: http://statmt.org/mtm4/?n=Main.HierarchicalReordering. I think it might be worth it to investigate these models as well. There is some smoothing going on when training the reordering models. There is a flag to train-model.perl: --reordering-smooth to set the behavior of the smoothing. There is no proper treatment of unseen phrases at decoding time though, which isn't that much of an issue if the same data is used for training the phrase table and the reordering model, since there won't be any unknown phrase pairs, except for pass-through OOVs, then. But I think it would be a good idea to look into that anyway. It was one thing we discussed at MT Marathon, but never had time to do something about. /Sara 2011-08-31 14:00, Neda NoorMohammadi skrev: > Hello all, > > We are testing different reordering models in our system ( source > language structure is sov and target is svo). > The results are amazing! The distance based model improves the results > in compare to lexical reordering models. > > we want to know is it true for other pair languages and also: > Isn't it true that lexical reordering is offered to improve the > reordering models? Then why we have get just the reverse?! How we can > argument why it has happend to our system? > > Also I have checked the LexicalReorderingState.cpp class. There is a > commented part in the code which shows it has uses no smoothing model > for estimating none-seen events of this model. Am I right? > > Thanks > _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
