I'm not sure what the end goal is here, but decide whether you want to
call the feature API generically moses/LM/* or directly e.g.
http://kheafield.com/code/kenlm/developers/, SRILM docs, IRSTLM, docs, ...

Kenneth

On 06/10/14 09:53, Hieu Hoang wrote:
> you can't selectively prevent a feature function from being called.
> 
> You can zero out it's weight, and write another feature function which
> selectively copies the score from the original feature function into
> it's own score.
> 
> This is what the phrase table
>    PhraseDictionaryMultiModel
> does.
> 
> However, with language models, it's a bit more tricky. LMs are statefull
> features, ie. they return state information which affect recombination.
> You can't prevent the state from affecting recombination even if you the
> weights are zero
> 
> 
> 
> 
> 
> On 10 June 2014 16:10, Lars Bungum <[email protected]
> <mailto:[email protected]>> wrote:
> 
>     On 10. juni 2014 14:46, Hieu Hoang wrote:
> 
>         a language model ultimately inherits from FeatureFunction. It
>         loads data by overriding
>            virtual void Load();
> 
>     I see that this is called in void
>     StaticData::__LoadFeatureFunctions(), which in turn is populated by
>     FeatureFunction::__GetFeatureFunctions().  Is that correct?
> 
>     What I want is to load all the LMs initially, but not using them all
>     "side by side" while decoding, but in order to produce a special
>     feature.  Can I use the current infrastructure to do this?  By
>     giving their paths in moses.ini and somehow accessing them in my
>     custom feature inside FF/ ?   Loading the LMs each time inside the
>     feature will be to slow, so I need some way to access the
>     "centrally" loaded LMs.
> 
>     //LB
> 
> 
> 
> 
> -- 
> Hieu Hoang
> Research Associate
> University of Edinburgh
> http://www.hoang.co.uk/hieu
> 
> 
> 
> _______________________________________________
> 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