it could be any number of things. all that tuning does is improve BLEU scores with respect to the given tuning set of sentences; so, this could result in:
--better BLEU scores, but worse subjective translation (since doing better at BLEU does not always guarantee better subjective translation --see our EACL paper on the subject) --better BLEU scores on a tuning set that is not representative of the actual intended test set (this is a domain problem) --your tuning set is far too small and you are overfitting with respect to it --bugs etc etc Miles On 01/02/2008, Panos <[EMAIL PROTECTED]> wrote: > > Hello, > > I built a baseline system for testing English to Greek translations. I > used the > bilingual corpus for the training process and the Greek translations of > the same > corpus for the language model (about 145000 sentences). Eveything seems ok > and > the system is able to produce some nice translations, domain-specific of > course. > However, the tuning process seems to create an .ini that produces pretty > bad > results. I tried the tuning process twice, one time with input and > reference > files of 2000 sentences and a second with 1000 sentences. Results are much > worse > than the ones I get with the untuned moses.ini. What am I doing wrong? > > Thanks in advance. > > Panos > > _______________________________________________ > Moses-support mailing list > [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] http://mailman.mit.edu/mailman/listinfo/moses-support
