If you work with factored models, typically it's expected that LMs are
trained for each factor. In your case, you have to train a LM on target
words and another LM on target POSs.

Class LMs are a feature provided by the IRSTLM toolkit: they model n-grams
of classes which translation (target) units are mapped to. I never used
them in combination with factored models, but in theory this should be
possible; nevertheless, they were introduced just for trying to efficiently
emulate the factored models in a single-factor framework.

Mauro

Joerg Tiedemann wrote:
> I have difficulties understanding the description about "class language
> models" at
> http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#ntoc8
>
> What I actually want to do is to use a language model on combined
> (concatenated) factors, let's say word+pos when decoding a factored
> model. I'm not sure if the "class language model" feature is the right
> thing to look at for doing this.
>
> for example, if my -translation-factors are 0-0,1 (0 being words and 1
> being target language POSs) then I would like to combine the translation
> prob's with a language model using factors 0,1 (word/pos) before the
> generation step from 0,1 to 0. and then maybe even adding a word
> language model on the generated surface string. what would be the proper
> way of doing this with moses?
>
>
> two other small things:
>
> I realised that moses has problems with xml-input when using lexicalised
> reordering. I get segmentation faults after decoding some sentences. It
> works fine using distance-based reordering.
>
> for the irstlm developers: it would be nice to change the hard-coded
> settings for gzip/gunzip in the bin/*.pl files to more general ones.
> otherwise I always have to do this by hand after downloading a new version.
>
> For example, replace
>         my $gzip="/usr/bin/gzip";
>         my $gunzip="/usr/bin/gunzip";
> with
>         my $gzip=`which gzip`;chomp $gzip;
>         my $gunzip=`which gzip`;chomp $gunzip;
>         $gunzip .= ' -d';
> or something like that.
>
> thanks.
> cheers,
> --
>
> Jörg
>
>
> ***********/\/\/\/\/\/\/\/\/\/\/\************************************
> **  Jörg Tiedemann                 [EMAIL PROTECTED]              **
> **  Alfa-Informatica               http://www.let.rug.nl/~tiedeman **
> **  Rijksuniversiteit Groningen    Harmoniegebouw, room 1311-429   **
> **  Postbus 716                    phone: +31 (0)50-363 5935       **
> **  9700 AS Groningen              fax:   +31 (0)50-363 6855       **
> *************************************/\/\/\/\/\/\/\/\/\/\/\**********
>
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
> .
>
>   


-- 
Mauro Cettolo
FBK - Ricerca Scientifica e Tecnologica
Via Sommarive 18
38100 Povo (Trento), Italy
Phone: (+39) 0461-314551
E-mail: [EMAIL PROTECTED]
URL: http://hlt.fbk.eu/people/cettolo

E cuale esie la me Patrie? cent, centmil, nissune
parcè che par picjâ lis bandieris spes a si picjin i omis

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to