hi prashant
My suggestions:
1. create a phrase table that loads only the rules for each sentence
from a text file just before it decode. This is done for the
hierarchical decoder when it uses suffix arrays. You need to write a
per-sentence phrase table for the phrase-based model.
2. you can filter the text phrase-table for each sentence and restart
moses for each sentence. If the language model is binarized, it will be
quick to restart.
3. you can use the existing text phrase-table, write some code to
change the scores, and retranslate the sentences.
On 24/08/2012 17:13, Prashant Mathur wrote:
Some things that I was already planning to do in adaptive decoding
1. One can query the phrase pair and get the scores
2. One can change the scores given the phrase pair
3. Update scores of selected phrase entries after decoding a sentence
This implementation would help in a scenario where I have a feedback
in the form of posteditions by a translator.
--
Prashant
On 08/24/2012 05:43 PM, Nicola Bertoldi wrote:
Hi Barry, Hi Prashant
could the implementation of "parallel data structure"
be a nice project for the forthcoming MT Marathon?
Which characteristics should it have?
I have some ideas about it
Nicola
On Aug 24, 2012, at 5:25 PM, Barry Haddow wrote:
Hi Prashant
If you want to update the phrase table after it's loaded, then you
want to use the in-memory one (PhraseDictionaryMemory) rather than
the on-demand one (PhraseDictionaryTree) since the entries in the
latter are loaded from disk as required and changes may be overwritten.
If you update the in-memory table though, you should remember that
Moses assumes it is thread-safe, and also that Moses prunes it at
load time to (a default of) 20 entries per source phrase.
I actually think that creating a "parallel data structure", or in
fact a wrapper that implements the same interface as the
PhraseDictionaries and delegates to one of the real ones, could be a
cleaner solution,
cheers - Barry
On 24/08/12 15:03, Prashant Mathur wrote:
Hi All,
I am trying to change the phrase scores after the phrase table and
reordering model is loaded in the memory. Is it possible? If so,
which class should I look into?
If not.. is it possible to maintain a parallel data structure and
prioritize it over the already loaded scores?
--
Prashant
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
Moses-support mailing list
[email protected] <mailto:[email protected]>
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
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
