i think in theory, the parameters don't have to scale, but in practise
they have to be otherwise the beam setting will screw up decoding. MERT
normalise the L1 weights, eg. these are my weights for a tuned (syntax)
model:
weight L1 L2
0.11 0.11 0.01
0.12 0.12 0.02
-0.03 0.03 0.00
0.03 0.03 0.00
-0.71 0.71 0.50
1.00 0.53
On 02/05/2010 10:11, Miles Osborne wrote:
there is a large amount of randomness involved with parameter tuning.
each time you run it (using the same language resources) you might
get different parameters,
also, the parameters are not scaled. this means that one run might
give you these values:
10 20 30
and the next run might give you these ones:
0.1 0.2 0.3
Miles
On 2 May 2010 09:34, Somayeh Bakhshaei <[email protected]
<mailto:[email protected]>> wrote:
Hi All,
A problem:
Isn't it true that the parameter tuning must gain the structure of
the language so i must get the same set of tuned parameters sets
with different kind of tune sets?
So why with changing the tuning set i get different amounts for
parameters?
another awful result:
I changed my test set, the Bleu result changed from 19 to 3 !!!!!
How its may while there is no overlap between none of the test
sets and train set?!!
------------------
Best Regards,
S.Bakhshaei
_______________________________________________
Moses-support mailing list
[email protected] <mailto:[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
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
