Thanks everyone for the replies. Alex's comment about the phrase-based systems is probably correct. In our experiment the cube pruning diversity was helpful with a smaller beam. But this was on a slightly different real-time setting, where the speed was critical.
I guess the diversity would help in hierarchical and I am planning to implement this in Hiero system that we have developed in-house (@ SFU) called Kriya. Unfortunately, I am not very familiar with the Moses code-base. - Baskaran On Tue, Oct 18, 2011 at 5:25 AM, Hieu Hoang <[email protected]> wrote: > oh yes, my mind is a sieve! > > i think it would be easy for someone to do. > > fyi, the cube popping loop is in > ChartCell.cpp line 112 > so i guess what you have to do is manipulate what hypotheses are > returned by Pop() in > RuleCubeQueue.cpp line 44 > > > > On 18/10/2011 17:08, Alexander Fraser wrote: > > Hi Hieu -- > > > > I think that was Hieu :-), Christian F, Christian H and I, when we > > implemented the phrase-based cube pruning at MTM2. > > > > The cube pruning diversity number sets the minimal number of > > expansions for each coverage in a cube that will be expanded. This is > > OR'ed with the normal cube pruning expansions. The implementation of > > this is at line 130 of SearchCubePruning.cpp and line 395 of > > BitmapContainer.cpp. > > > > In the phrase-based experiments we did, using this parameter did not > > help search performance. So it may be that, at least for phrase-based, > > the intuition that certain *coverages* are systematically incorrectly > > pruned is not good (and instead one must look at individual > > hypotheses, which is what cube pruning does). > > > > This idea of diversity in hierarchical could be easily (trivially?) > > implemented in the same way, maybe it would help there. > > > > Cheers, Alex > > > > > > On Mon, Oct 17, 2011 at 2:56 PM, Hieu Hoang<[email protected]> wrote: > >> cube pruning *diversity*. I think philipp koehn added some heuristics to > >> phrase-based, not sure how it works > >> > >> On 17 October 2011 19:53, Lane Schwartz<[email protected]> wrote: > >>> Wait - seriously? Cube pruning is not implemented in the hierarchical > >>> model? > >>> > >>> On Sat, Oct 15, 2011 at 1:42 AM, Hieu Hoang<[email protected]> > wrote: > >>>> hi baskaran > >>>> > >>>> no, it's not presently used it hierarchical pb models. It would be > >>>> interesting to try it, especially for syntax models, as arguably the > >>>> lack of > >>>> diversity is one of the problem with these models > >>>> > >>>> On 15/10/2011 04:40, Baskaran Sankaran wrote: > >>>> > >>>> Hi, > >>>> > >>>> I wonder, whether the param cube-pruning-diversity (cbd) is being used > >>>> when > >>>> using the moses_chart for hierarchical phrase-based models? If so, > does > >>>> it > >>>> have the same meaning as in the regular phrase-based decoders? > >>>> > >>>> Thanks > >>>> - Baskaran > >>>> > >>>> > >>>> _______________________________________________ > >>>> 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 > >>>> > >>>> > >>> > >>> -- > >>> When a place gets crowded enough to require ID's, social collapse is > not > >>> far away. It is time to go elsewhere. The best thing about space > travel > >>> is that it made it possible to go elsewhere. > >>> -- R.A. Heinlein, "Time Enough For Love" > >> _______________________________________________ > >> 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
