Thank you all. Can you explain further what does it mean that MERT won't
know that the feature exists? Does that mean that the tuneable feature
weights are optimized assuming that all non-tuneable feature weights are
equal to zero?
In fact, in my understanding this should lead to a dramatic decrease of the
score Bleu on the tuning set, but this is not what I ended up with in my
tests, at least in some cases (decrease of just 0.19 Bleu when adding
tuneable=false to PhrasePenalty and Distortion on a tuning set constituted
by 574 segments).

Vito M.

2015-05-20 14:38 GMT+02:00 Rico Sennrich <[email protected]>:

> Matthias Huck <mhuck@...> writes:
>
> >
> > Hi Vito,
> >
> > tuneable=false should work.
>
> Just my usual caveat:
>
> if you use 'tuneable=false', the feature score(s) won't be
> reported to the n-best list, and MERT/MIRA/PRO won't even know that the
> feature exists. This is appropriate in some cases (keeping a feature weight
> at 0, or giving a high penalty to some glue rules to ensure that they are
> only used if no translation is possible without them), but in other cases,
> hiding important features causes the optimizer to search the wrong space.
>
> best wishes,
> Rico
>
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
>



-- 
*M**. Vito MANDORINO -- Chief Scientist*


[image: Description : Description : lingua_custodia_final full logo]

 *The Translation Trustee*

*1, Place Charles de Gaulle, **78180 Montigny-le-Bretonneux*

*Tel : +33 1 30 44 04 23   Mobile : +33 6 84 65 68 89*

*Email :*  *[email protected]
<[email protected]>*

*Website :*  *www.linguacustodia.com <http://www.linguacustodia.com/> -
www.thetranslationtrustee.com  <http://www.thetranslationtrustee.com/>*
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to