Anecdotally, this feature also isn't especially important, see e.g.: http://www.mt-archive.info/AMTA-2006-Lopez.pdf
On Mon, Feb 22, 2010 at 12:57 PM, Carlos Henriquez <[email protected]> wrote: > > The last weight from the phrase-table corresponds to the phrase penalty as > explained in > > http://www.statmt.org/moses/?n=FactoredTraining.ScorePhrases > > This features benefits translations that use less units to decode. > > It is the Euler's number and it is determined as such because the decoder > uses natural logarithms to add the weights, instead of multiplying > probabilities. Hence, 2.718 will become 1 during decoding as a way of saying > "I'm using yet another phrase unit to perform a translation". > > So you were quite right after all :o) > > -- > Carlos A. Henríquez Q. > [email protected] > > > ________________________________ > De: Calia <[email protected]> > Para: [email protected] > Enviado: lun,22 febrero, 2010 05:26 > Asunto: [Moses-support] 2.718 in the phrase-table > > > The exact purpose of the value 2.718 is what I am not sure of. > > As far as I know, the value is used for preferring the hypothesis with less > phrases to > > the one that use more phrases(words) for the same coverage of the source > sentence during > > the prefix cost comparison inside priority queue( for the hyps that covers > the same source range ). > > for example ) > > hyp1(I love you -> phrase_1 + phrase_2 ) vs hyp2(I love you -> phrase_1 + > phrase_2 + phrase_3 ) > > to prefer hyp1 to hyp2 by multiplying hyp1 by 2.718. > > This is how I understand the use of the value, since longer phrase has > empirically better translation > > than the one made up of word-based translation. > > Is there any one who can confirm my belief, or to correct my conclusion? > > p.s. Is 2.718 the Euler's number? If it is, why is the weight determined as > such? > > > > > > > > > > > > > > > _______________________________________________ > 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
