Thanks Chris, Hieu.
These should be identical. However, it is possible that the beam
> search used by the decoder might run slightly differently in the two
> cases, leading to different results. With no pruning and unlimited
> stack sizes, the results should not be different (unless you have
> features sensitive to epsilon transitions).
>
I was using cube pruning algorithm with cube-pruning-pop limit 2000 and
stack size of 2000
I tried using stack algorithm with lattice input and same stack size(2000).
However, when I try the same with normal input, it takes too long to decode.
So, I can't compare the results.
Can I actually specify unlimited stack size? Wouldn't it take too long to
decode in that case with the normal stack algorithm?
If I use cube pruning, what kind of cube pruning pop limit should I set?
Amit
On Tue, Nov 16, 2010 at 6:37 AM, Chris Dyer <cd...@cs.cmu.edu> wrote:
> > I had a query with regard to use of lattice input in moses.
> > There is a little difference in the translations generated when I run
> moses
> > using the 'normal' input format and when I run it with 'lattice input'
> > format.
> > The translations weren't radically different - only a few phrases were
> > different.
> There are some slight differences in how the distortion scores are
> computed with lattices and without, which could explain some slight
> differences.
>
> >
> > When running moses without lattice input, each line in my input file
> looks
> > like the following:-
> > a b c d e f g h
> >
> > When running it using word lattices each line in my input file looks like
> > the following:-
> >
> ((('*EPS*',1.0,1),),(('a',1.0,1),),(('b',1.0,1),),(('c',1.0,1),),(('d',1.0,1),),(('e',1.0,1),),(('f',1.0,1),),(('g',1.0,1),),(('h',1.0,1),),(('*EPS*',1.0,1),),)
> >
> > Should there be any differences in the translations produced in the two
> > cases?
> These should be identical. However, it is possible that the beam
> search used by the decoder might run slightly differently in the two
> cases, leading to different results. With no pruning and unlimited
> stack sizes, the results should not be different (unless you have
> features sensitive to epsilon transitions).
>
>
> > When calling moses I give the parameters -inputtype 2 -weight-i 0.2.
> >
> > Also I wished to know, how is the 'weight-i' used here?
> > My understanding is that (weight-i)*log(path weights) + lambda1*lm + ....
> > determines the final log probability of a hypothesis. (where by path
> weights
> > I mean the product of the arc weights we specify in the lattice input
> format
> > for the path in question). Is it correct? and in that case should one
> also
> > perform some sort of tuning for this weight?
> Yes this is correct. You may tune this weight using MERT or whatever
> optimizer you like.
>
> >
> > Regards,
> > Amit
> >
> >
> >
> >
> >
> > _______________________________________________
> > Moses-support mailing list
> > Moses-support@mit.edu
> > http://mailman.mit.edu/mailman/listinfo/moses-support
> >
> >
>
_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support
