I will get the blowup but since its generating phrases for a sentence only it 
can fit in memory.
Ofc the time would be more but im eager to see what happens.


Sent from Samsung Mobile

-------- Original message --------
From: Barry Haddow <[email protected]> 
Date: 24/10/2014  20:09  (GMT+09:00) 
To: Raj Dabre <[email protected]> 
Cc: [email protected] 
Subject: Re: [Moses-support] Phrase pair generation at run time using 
Source-Pivot and Pivot-Target phrase tables 
 
On 24/10/14 11:06, Raj Dabre wrote:
> That is a good starting  point suggestion.
> Many thanks.
>
> Perhaps an easier option would be to do the synthesis offline. In 
> other words, take your two tables and create a pivot table from them, 
> and then use it like a normal phrase table.
>
> This I have been doing for 6 months but the phrase tables generated 
> are super-huge where the size almost is a square of the original size. 
> I end up having to keep a threshold frequency and kill potentially 
> good phrase pairs. Thats why I want to generate it online and keep all 
> the pairs.

But if you synthesise the phrase pairs inside the decoder, then won't 
you get the same n^2 blow-up? I think you have to find a good way to 
prune however you implement the synthesis.


>
> Thanks again.
>
> On Fri, Oct 24, 2014 at 6:58 PM, Barry Haddow 
> <[email protected] <mailto:[email protected]>> wrote:
>
>     Hi Raj
>
>     You could create a custom phrase table implementation to produce
>     your synthesised phrase pairs. Have a look at the existing phrase
>     table implementations in moses/TranslationModel. In particular,
>     you need to subclass PhraseDictionary. The method
>     GetTargetPhraseCollectionLEGACY() returns a collection of phrase
>     pairs, given a source phrase.
>
>     Perhaps an easier option would be to do the synthesis offline. In
>     other words, take your two tables and create a pivot table from
>     them, and then use it like a normal phrase table.
>
>     cheers - Barry
>
>
>     On 21/10/14 11:14, Raj Dabre wrote:
>
>         Hello,
>         I am currently doing research on using pivot languages for
>         Phrase based SMT.
>
>         My current method involves the usage of alternate decoding
>         paths feature to combine multiple synthesized Source-Target
>         phrase tables. (I have noticed that not many people exploit
>         this method or even if they do.... they don't mention it clearly).
>
>         However pre-synthesized  phrase tables need to be pruned to
>         remove low probability phrase pairs and I would like to
>         generate phrase pairs via a pivot at run time. I am ok with
>         taking additional decoding time.
>
>         I am aware that Bertoldi (2008) had already mentioned that he
>         had used this method but this is not present in the moses
>         decoder release.
>         I would very much like to implement this but do not know where
>         to start.
>         If someone could tell me the section of code that reads in
>         phrase pairs given a source phrase I think I might be able to
>         do something.
>         Any help would be appreciated.
>
>         Thanks in advance.
>
>         -- 
>         Raj Dabre.
>         Research Student,
>         Graduate School of Informatics,
>         Kyoto University.
>         CSE MTech, IITB., 2011-2014
>
>
>
>         _______________________________________________
>         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.
>
>
>
>
> -- 
> Raj Dabre.
> Research Student,
> Graduate School of Informatics,
> Kyoto University.
> CSE MTech, IITB., 2011-2014
>


-- 
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

Reply via email to