Hi Maria

The old mert used two different stopping criteria:
(i) If the weights didn't change by more than a certain amount (defaulting to 
0.00001)
(ii) If no new translations were produced after the weight update.

Criterion (ii) was never actually implemented for new mert, and in some 
situations the weights can oscillate so that criterion (i) is never fulfilled 
and training never stops. I've used new mert fairly extensively and never 
observed this, but since we removed old mert, there's been a couple of 
reports of this behaviour with new met.

The best solution would be to revisit the stopping criteria, and either 
reinstate (ii), and/or add a stopping criterion based on the change in bleu 
score. This may well be done at the mt marathon, but probably not before. 
Until then a simpler solution would be to add a 'max iterations' argument, 
and default this to (say) 20, since most mert runs have finished at that 
stage. I will check something in shortly for this,

regards
Barry

On Wednesday 01 September 2010 14:46, Maria Holmqvist wrote:
> Dear moses developers,
>
> I recently installed a new version of Moses decoder and training
> scripts (revision 3407).
> The problem is that the tuning step never stops. It goes on for 50-80
> iterations (until I interrupt it) but Bleu on the devset doesn't
> really improve after the 15 first iterations.
> I've run a few experiments with different amounts of Swedish-English
> data and different symmetrization heuristics, but the problem remains.
>
> In comparison I trained and tuned a model with an older version of
> moses and cmert. This tuning stopped after about 10 iterations as
> normal.
>
> Have someone else experienced this problem? Do you have any ideas on
> how I can solve it or should I go back to the old Moses version?
>
> /Maria
> _______________________________________________
> 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