i would say it would be difficult to do properly. the 2 directions are quiet separate, they don't share any models so there won't be any efficiency savings. the phrase table can be shared as the data for both direction are the same, just inverted, but the phrase table implementation has to be rewritten.
the global variables, chiefly StaticData, are a problem. Barry haddow has been making the decoder multithreaded, which is a good start, but a lot more needs to be done to make it work with multiple language directions. ie. probably about 6 months if you knew the code already Catalin Braescu wrote: > How hard would it be to change it? Did anyone made an analysis and > come up with an estimation of man-months of developing and of which > modules may need the rewriting? > > Catalin > Omlulu > > On Tue, Sep 15, 2009 at 3:22 PM, Hieu Hoang <[email protected]> wrote: > >> hi catalin >> >> it would be a nice feature to have, but Moses can't do this and it would be >> currently difficult to make it do this. >> >> you will need 2 separate instances of the decoder. >> >> Catalin Braescu wrote: >> >>> I wonder what would be required for Moses to handle more than 1 >>> language pair, 1-way? As in, the same instance of Moses to translate >>> both from Lang1 to Lang2 and from Lang2 to Lang1. Ideally to translate >>> to/from more than only 1 language pair. >>> >>> >>> >>> Catalin Braescu >>> Omlulu >>> _______________________________________________ >>> 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
