Hi João,

if you do not get different results for a beam of 1000 vs a beam of 10000
then, yes, that means that you do not get additional search errors by cutting
the stack sizes by a factor of 10. For Europarl trained models this is
consistent with my findings (see also my 2004 AMTA paper on Pharaoh).
I have seen different results for other models, especially if they are
larger. Obviously the reordering limit plays a role as well.

-phi

On Nov 19, 2007 4:50 AM, João Graça <[EMAIL PROTECTED]> wrote:
> Hello,
>
> I am trying to do some experiments with the beam size in moses. So i run the
> same experience with the default beam and with 1000 10000 using option -s on
> moses decoder. For my surprise the results for beam = 1000 10000 where
> exactly the same. What does this mean, that the decoder is not making
> mistakes by pruning, that the difference between 1000 and 10000 is not
> enough to fix the mistakes, or that i am doing something wrong.
>
> Thanks a lot,
>
> João Graça
>
> _______________________________________________
> 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

Reply via email to