I think that we should consider a tighter coupling of the XML markup  
and the training process.  It seems logical to markup the training  
items along with the test items.  We could add a preprocessing where  
we replace XML delineated phrases in training items with their tag.   
Since the tags would then be part of the Giza training / phrase  
extraction they would be included in larger phrases, and get  
reordering probabilities.  This might be useful for putting in  
placeholders for names and numbers.

--Chris



On Dec 12, 2007, at 11:12 AM, Hieu Hoang wrote:

> yo loic
>
> i'm not the xml preprocessing guy for moses, but i don't think  
> marking up the input does, or should, change the phrase table.
>
> yes - the markup is primarily used to tell the decoder to avoid  
> certain words, or give extra info to translate them.
>
> but it might be a good idea to learn from the markup
>
> Loic Dugast <[EMAIL PROTECTED]> wrote: Hello
>
> I have a short question concerning the XML markup in Moses.
> What happens in the phrase table if I mark up some entity as say,
> type=person>Britney Spears ?
>
> Is there a way to generate such phrase pair as:
> Britney Spears is a great singer
> type=person>Britney Spears est une grande chanteuse ?
> ... while actually learning: is a great singer
> est une grande chanteuse ?
>
> (allright , with low probabilities associated)
>
> Or should I consider this being two different issues: 1) marking some
> phrases as entities (with XML markup, in order for example not to  
> translate
> them)
> 2) generalise the relevant phrase pairs (by adding an ENTITY_PERSON  
> *factor*
> maybe) ?
>
> Thanks
>
> Loic
>
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
>
>
>
>
> Hieu Hoang
> http//www.hoang.co.uk/hieu
>
> Support the World Aids Awareness campaign this month with Yahoo!  
> for Good
> _______________________________________________
> 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