Hi,

We once empirically compared two different recombination schemes in a
hierarchical phrase-based system (without any kind of neural network
language model):

Recombination T.  The T recombination scheme recombines derivations that
produce identical translations. (I.e., hypotheses with the same
translation but different phrase segmentation are recombined.)

Recombination LM.  The LM recombination scheme recombines derivations
with identical language model context. (This is what we'd usually do.)


Cf. the following publication:

M. Huck, D. Vilar, M. Freitag, and H. Ney. A Performance Study of Cube
Pruning for Large-Scale Hierarchical Machine Translation. In Proceedings
of the NAACL 7th Workshop on Syntax, Semantics and Structure in
Statistical Translation (SSST-7), pages 29-38, Atlanta, Georgia, USA,
June 2013.
http://aclweb.org/anthology//W/W13/W13-0804.pdf


You'll find a couple of statistics and performance plots in the paper
for Chinese-to-English and Arabic-to-English translation tasks. This was
all done with the Jane SMT toolkit. 
http://www.hltpr.rwth-aachen.de/jane/
Note that the cube pruning k-best generation limit was applied after
recombination for the experiments in the paper.


I tend to think that you could do the "Recombination T" scheme if the
benefit of an RNNLM in terms of translation quality justifies it. There
might be pitfalls, e.g. regarding pruning settings and tuning on n-best
lists. Default Moses settings won't necessarily be a good choice.

Cheers,
Matthias


On Mon, 2016-03-07 at 15:41 -0600, Lane Schwartz wrote:
> Philipp,
> 
> Are you aware of any published work examining the importance of hypothesis
> recombination in terms of time/space/quality tradeoffs?
> 
> Lane
> 
> 
> On Mon, Mar 7, 2016 at 3:19 PM, Philipp Koehn <[email protected]> wrote:
> 
> > Hi,
> >
> > integrating this into the decoder will break all hypothesis recombination,
> > so it may be better (and definitely easier) to use the RNNLM to rerank
> > n-best lists.
> >
> > -phi
> >
> > On Mon, Mar 7, 2016 at 3:46 PM, Jake Ballinger <[email protected]>
> > wrote:
> >
> >> Hello everyone,
> >>
> >> Has anyone used an RNNLM language model instead of one of the recommended
> >> language models? I was specifically looking at the RNNLM toolkit provided
> >> by Tomas Mikolov at http://rnnlm.org/.
> >>
> >> Thank you!
> >>
> >> --
> >> Jake Ballinger
> >> Major: Computer Science
> >> Minors: Chinese, French, Spanish, & Math
> >> 443-974-6184
> >> [email protected]
> >> Box 582
> >>
> >> _______________________________________________
> >> 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



-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to