My basic plan for experimentation was to use a translation model only. Of course, with the current implementation (as I understand it) many permutations generated by the translation model alone will have the same phrases in a different order and thus have the same probability. I was thinking about perhaps modifying the translation model such that it generates sentences from phrases that overlap on both the source side and target side and scoring in a way that gives a higher rank to translations generated with greater degrees of overlap.
Do you think it would be easier to write the new translation model from scratch or adapt the existing one? James ________________________________ From: [email protected] [[email protected]] on behalf of Hieu Hoang [[email protected]] Sent: 05 November 2013 12:12 To: Read, James C Cc: Daniel Cer; Kenneth Heafield; [email protected] Subject: Re: [Moses-support] gappy phrases If you're gonna throw most of Moses away, you may consider using and extending the basic decoder I've just added to github https://github.com/moses-smt/mosesdecoder/tree/master/contrib/basic-decoder Or write your own. You're welcome to add it to moses too On 5 November 2013 07:29, Read, James C <[email protected]<mailto:[email protected]>> wrote: Interesting. This isn't going to be an issue for the kind of experiments I will be running, though. I won't be using a language model or a reordering model or a beam search during decoding (I know, not much of Moses left). My main usage of 'Moses' at the moment is training translation models so I can run experiments with 'units of translation' in isolation from other variables in the system (language model, reordering model, beam search). I would really like to be able to run some experiments also with discontinuous phrases both source side and target side. Any idea what kind of changes I would need to make to the training process to be able to learn these kind of transformations? I suppose I'm also going to need to modify the operation of the translation model to get these working as well. thanks, James ________________________________ From: [email protected]<mailto:[email protected]> [[email protected]<mailto:[email protected]>] on behalf of Daniel Cer [[email protected]<mailto:[email protected]>] Sent: 04 November 2013 19:43 To: Kenneth Heafield Cc: [email protected]<mailto:[email protected]> Subject: Re: [Moses-support] gappy phrases Hi everyone, Ken and I just spoke about this. Here's a quick synopsis of our semi-recent experience at Stanford with discontinuous/gappy phrases: * Source side gaps are effectively free and don't really degrade decoding time. * Target side gaps are fine for smaller beam sizes (e.g., < 200). * When using large beams, our current implementation slows down dramatically. For example, with a stack size of 500, I think it was sometimes taking over an hour to translate some sentences. While discontinuous phrases can moderately increase the BLEU score, but we get a bigger increase by just using very large beam sizes. Dan On Sun, Nov 3, 2013 at 8:34 PM, Kenneth Heafield <[email protected]<mailto:[email protected]><mailto:[email protected]<mailto:[email protected]>>> wrote: Hi, I'll throw in the anecdote that gappy phrases are currently not in use at Stanford. My predecessor told me that it took a lot longer and only improved BLEU slightly on Chinese-English. But it's also possible that something didn't get passed down correctly from Michel to my predecessor to me. . . Kenneth On 11/03/13 14:18, Read, James C wrote: > My understanding is that they used a similar approach as the grammar > extraction to extract the gappy phrases. Would it be a massive undertaking to > get Moses to support this? > > James > ________________________________________ > From: Barry Haddow > [[email protected]<mailto:[email protected]><mailto:[email protected]<mailto:[email protected]>>] > Sent: 30 October 2013 09:26 > To: Read, James C > Cc: > [email protected]<mailto:[email protected]><mailto:[email protected]<mailto:[email protected]>> > Subject: Re: [Moses-support] gappy phrases > > No, but it does support hiero and syntax models. > > On 29/10/13 22:23, Read, James C wrote: >> Hi, >> >> does anybody know if Moses supports gappy phrases >> http://www-nlp.stanford.edu/pubs/naacl10-discontinuous_phrases.pdf >> >> James >> >> _______________________________________________ >> Moses-support mailing list >> [email protected]<mailto:[email protected]><mailto:[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. > > > _______________________________________________ > Moses-support mailing list > [email protected]<mailto:[email protected]><mailto:[email protected]<mailto:[email protected]>> > http://mailman.mit.edu/mailman/listinfo/moses-support > _______________________________________________ Moses-support mailing list [email protected]<mailto:[email protected]><mailto:[email protected]<mailto:[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 -- 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
