Hi, Ok, I understand this approach now. I generally like the idea of backoffs and it give more opportunity to further tune the engines. And it's even more powerful when multiple LMs goes in place. Just I think the MERT might have difficulties to handle such huge numbers of parameters to be tuned (especially for Czech where we hit the sparse issues ).
Tomas -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Philipp Koehn Sent: Monday, November 3, 2014 11:47 PM To: Tomas Fulajtar Cc: moses-support ([email protected]) Subject: Re: [Moses-support] Combine models with backoff Hi, you should just add multiple language models - tuning will find proper weights for them. There is no way to do backoff with language models, just interpolation. -phi On Mon, Oct 27, 2014 at 12:49 PM, Tomas Fulajtar <[email protected]> wrote: > Hi all, > > > > I would like to combine two phrase based engines. One, smaller is > trained on desired domain data, but with limited corpus size. The > second is the legacy one with huge phrase table and LM, but with kind > of older/more obsolete terminology. Thus the idea is to combine both > to preserve domain/language style from the first engine, but also > reduce OOV with application of the second engine. > > > > I think what I am looking for is the Back-off model - use the small > one as a preferred one , and then the second in case of phrases not > found. I have setup such a config in accordance with > http://www.statmt.org/moses/?n=Moses.AdvancedFeatures#ntoc25,. > > > > [feature] > > PhraseDictionaryCompact name=A > > PhraseDictionaryCompact name=BackOff > > > > [mapping] > > 0 T 0 > > 1 T 1 > > > > [decoding-graph-backoff] > > 0 > > 1 > > > > [weight] > > A = 0 0 0 0 > > BackOff = 0 0 0 0 > > > > And it seems to work (weights were tuned afterwards with mert). > > > > I have also read the > http://comments.gmane.org/gmane.comp.nlp.moses.user/10099. However > there is not mentioned how the LMs combination could be managed. I > can add both to ini file and perform the weights tuning, or is it > better to set the weights manually? I believe that phrase table > backoff would ensure the preference of model A terminology, while > combination of both LMs would make the translation smoother as it can benefit > from the second, bigger LM. > > > > Could you please correct my assumptions? I hope the explanation does > make some sense… > > > > Thank you very much, > > > > Tomas > > > _______________________________________________ > 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
