named parameters

[John Morgan]

Hello Chris,
I'm using cdec as a decoder for models built with moses. I'm building
the systems with a separate language model for corpora that come from
different domains. My understanding is that during mert tuning
weights for the separate LMs are estimated separately. The different
LMs are considered separate features. Can I use separate LMs as
separate features with cdec?
Can I put something like this in the cdec.ini file:
feature_function=LanguageModel_1 = lm_1 -o 3
feature_function=LanguageModel_2 = lm_2 -o 5
.
.
.

TThen put the following in a weights file:
LanguageModel_1 0.12345
LanguageModel_2 0.234567
LanguageModel_3 1.2345
.
.
.

John
--
Regards,
John J Morgan

[Chris Dyer]

Hi John,
This is something I've been meaning to fix. Right now cdec stupidly
doesn't let you specify the name of the language model feature, so if
you have more than one, they will conflict (with unpredictable
results). I'll add this in the next day or so (I'm currently back in
MD for my graduation, so I've got a few things going on at the
moment). It will be something like:
feature_function=LanguageModel -fid LM1 -o 3 foo.lm.gz
Which will tell cdec to use a feature name of "LM1" for that lm.
-Chris

[John Morgan]

Hi Chris,
Congratulations on the graduation.
I've played a little with vest. It will be important to be able to
specify multiple LMs as features whose weights are to be optimized by
vest. Right now vest seems to only run 1 iteration. I don't see any
errors. I can wait for vest to work with multiple LMs.
Thanks,
John

[Chris Dyer]

Hi John,
I looked a bit more closely, and someone has already added the ability
to name the LM parameter. See this page:

http://cdec-decoder.org/index.php?title=Feature_functions_for_rescoring_an_SCFG_translation_forest#LanguageModel

It's a bit unusual for vest to only run a single iteration, though.
Less than 6 is normal, but it's usually more than 2 or 3.

-Chris