Hi, thank you. Interesting report. As a bonus, the fig. 5 explains how the phrases are extracted: "consistent word alignments". Finally I understand :-)
It might be a good idea to include that illustration on the page: http://www.statmt.org/moses/?n=Moses.Background It would make the concept easy to understand. Yours, Per Tunedal On Wed, Apr 3, 2013, at 14:32, Philipp Koehn wrote: > Hi, > > you will have to experiments yourself to see what works best. > We use by default grow-diag-final-and or sometimes grow-diag-final > > A clear effect of the different methods is the number of alignment points > and hence the number of phrase pairs extracted. There may not be a > clear picture with regard of quality --- at least when I ran experiments > with small amounts of training data. > http://homepages.inf.ed.ac.uk/pkoehn/publications/iwslt05-report.pdf > > -phi > > On Wed, Apr 3, 2013 at 7:54 AM, Per Tunedal <[email protected]> > wrote: > > Hi, > > Obviously, alignment is important for translation quality. > > Unfortunately, I don't fully understand the options. > > > > I've read the pages: > > http://www.statmt.org/moses/?n=FactoredTraining.TrainingParameters > > and > > http://www.statmt.org/moses/?n=Moses.Background > > > > Pros and cons of different settings? > > > > What's "final"? > > Intersect "creates a lot of extracted phrases": isn't that a good thing? > > > > Any hints on cases when a specific option might be useful? What's useful > > to try when translating between languages that differ a lot in word > > order and ways to express yourself (like Swedish and French)? And, on > > the contrary, languages that are very similar (like Swedish and Danish)? > > > > Yours, > > Per Tunedal > > _______________________________________________ > > 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
