- From: Olaf-Michael Stefanov <olaf@stefanov.at>
- Date: Fri, 19 Jul 2013 10:30:36 +0200
- To: public-i18n-its-ig@w3.org
Hi Yves, Dave, Phil, Arle, Serge, all, I fully agree, Yves. Your small suggestion to drop the "self-generated" part of the definition for MT confidence score make sense, is very reasonable, is still - to my understanding doable within ITS 2.0, and should be done. Kind regards, olaf-michael On 2013-07-17 18:50, Yves Savourel wrote: > Hi Dave, Arle, Phil, Olaf-Michael, all, > >> [13:26] <daveL> arle: for quality estimation with lucia specia >> this is definitely an assessment of the translation >> [13:26] <daveL> dave: in this case would it be the same of >> LQI conformance type > (BTW: the LQI type is 'non-conformance'). I could see some link with what QuEst does. But, to me, LQI has to do with marking out > problems. > > In this use case I have an MT candidate, I run some process on it and get a value that tells me how close the candidate seems to be > from a translation that is deemed ok by the given system. In real life I would, for example, use that value to possibly weed out > some MT candidates, or sort them, before they go to a translator. > > Providing an assessment of the translation quality seems to be exactly what MT confidence is suppose to be doing, with the exception > (currently) that this is to be done by the same MT system that generated the translation. > > My question then: does the current MT confidence a duplicate of the LQI/non-conformance? I think not. > > And I think--regardless of the merit of QuEst for this--a small change saying that MT confidence is not necessarily generated by the > same tool that created the MT candidate would make sense. > > Cheers, > -ys > > >
Received on Friday, 19 July 2013 08:31:06 UTC