Hi, the size of the search graph corresponds to the number of hypotheses that are connected to full translation paths (including hypotheses that recombined away to such connected hypotheses).
If you are using cube pruning, the same number of hypotheses are created, no matter what the distortion limit is. With regular decoding, there will be actually more hypotheses created with larger distortion limit. It may be that large distortion limits leads to more hypotheses that are dead ends, i.e., hypotheses that initially survive stack pruning, but have no viable extensions. -phi On Tue, Jul 12, 2011 at 2:27 PM, Artem Sokolov <[email protected]> wrote: > Hi, > > I noticed that, when increasing the distortion limit (-dl option) the > size of the extended output search graph (-osgx option) gets smaller > and, vice versa, it gets larger if I lower the distortion limit. > > Could someone explain, please, what could be the reason for this? > As the weight of the distance penalty in the system is positive, I > would expect it to have the opposite behavior -- the size of the > search space should grow as one permits more distant reorderings. > > Thank you. > > -- > Artem Sokolov > _______________________________________________ > 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
