I had made an external script for this some time ago. It replaced each number with a tag before translation, and after translation it used the word alignment information to replace the tags with their stored value. But it only worked with phrase-based translation, not hierarchical. And it didn't work well enough, I dont remember exactly but i think it crashed when the translation didn't have the same number of tags as the source, maybe other problems. And anyway, if you plan to implement this "in the decoder", i suppose an external script is not what you want.
-- Raphaël Payen On Thu, Jun 6, 2013 at 9:59 PM, Hieu Hoang <[email protected]> wrote: > As you say, the number's in the input sentence would be unknown. However, > the reason to use placeholders is to make them known to the LM and > phrase-table so that they can have more accurate scores for them > > Therefore you need to replace numbers in your input sentence with > placeholders. > > You can then use the word alignment from the decoder to put the number > back. > > This is something I want to implement better in the decoder. So if you, or > anyone, is willing help me and contribute some time & code, I can help out. > > > > On 6 June 2013 15:25, Arezki Sadoune <[email protected]> wrote: > >> Dear Hieu Hoang, >> >> Thank you for the answer, >> >> Yes, I'm replacing the numbers with a placeholder on the training data and >> the LM as well, I thought this might address the issue of the number without >> losing too much translation quality, >> >> Regarding the input sentence I'm not interfering with the process assuming >> that the number will still the same as it is unknown..., An other option >> will be to make a script which, as you said puts back the original number >> for a 100% accuracy. >> >> Do you think I could use the both at the same time? >> >> Regards >> >> >> A.S >> >> >> _______________________________________________ >> Moses-support mailing list >> [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 > >
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
