One strategy which Google used (at their NIST presentation this year) was to have phrase tables with multiple case variants. That would tightly couple case within MERT and you'd not suffer from making independence assumptions arising from reintroducing case info later.
(Using case / no-case as factors might emulate this, but I suspect this will be overkill since for most words you'd only have a single lower cased version ie a single factor) Hmm, I see a future MSc project here. Miles 2008/6/25 John D. Burger <[EMAIL PROTECTED]>: > Miles Osborne wrote: > > this is actually an argument for *preserving case* at all stages, rather >> than throwing information away and only re-introducing it later (with >> possible random effects like that one). >> > > Yeah, John Henderson and I have had that argument, several times. My > concern has been that you introduce data scarcity issues by splitting the > counts for rare words. I'm beginning to suspect, however, that this would > be outweighed by the better discrimination you'd get by keeping case > distinctions. Of course, with Moses, you could have two factors, with and > without case. I wonder if anyone has tried that ... > > out of curiosity, what happens when you use -1 instead (no limits on >> reordering) >> > > Well, it takes =way= longer, of course. As for scoring, it does poorly, > losing as much as a point of BLEU, whether scoring caselessly or not. > Interesting to try, though. > > - John Burger > MITRE > > -- 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
