Mert procedure of Moses provides an option called "--activate-features"
which performs something similar to the opposite of what you need.


--activate-features=STRING  ... comma-separated list of features to work on (if 
undef work on all features), others are fixed to the starting values (or better 
to their ratios)



You can specify one or more features to estimate, keeping the others fixed, in 
the sense that their their ratios are fixed.
In other word, this option groups the "not-active" features into one virtual 
new feature which is weighted as usual together with the "active" ones.

But, considering that mert provides normalization of the weights at each 
iteration, this option can not be used to fix just one weight.

Your request is also not well-defined because any set of weights can be 
proportionally augmented or reduced to give  one of them always the same value, 
without changing the output of the decoder: multiplying all feature weights by 
a constant value does not modify the translations, but only their scores by the 
same constant factor.

For instance, if you want to fix the LM (lm) and the word penalty (w), you can 
activate the translation model (tm) and the distortion model (d) you have to 
run mert with  this
--activate-feaures tm,d
where features are specified with their short names.


best,
Nicola



On 7/16/09 6:02 PM, "Marta Ruiz" <[email protected]> wrote:

Dear all,

I was optimizing with mert-moses-new.pl and I wonder how I can fix a lambda to 
one specific value
and keep that value fix during all the optimization.
Thanks in advance,

Marta
_______________________________________________
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