Hi Nicola, thank you for your answer. I conclude that "back-off weight" is the same as what's called "back-off cost" in Koehn's textbook on SMT. And that Moses use the back-off procedure described in the book. Anything on this in the Moses wiki? Yours, Per Tunedal
On Sun, Aug 25, 2013, at 9:53, Nicola Bertoldi wrote: > Hi Per, > > for all n-gram (but those of the highest order), the third field is the > logarithmic back-off weight (logBO) > if not reported, the weight is assumed equalt to 0 (in log scale) > > suppose you want to compute > logP(maison | cadre seulement) > > and the 3-gram > "cadre seulement maison" > is absent , i.e. logP(maison | cadre seulement) = 0.0 > > and the 2-gram is present as follows: > -0.7 cadre maison -0.1 > > hence, the LM is computed as: > > logP(maison | cadre seulement) + logBO(cadre seulement) * logP(maison | > seulement) = > = 0.0 + -4.57217 * -0.7 > > (sorry for the example, but I do not speak French) > > best > Nicola > > On Aug 23, 2013, at 9:20 PM, Per Tunedal wrote: > > > > > Hi, > > how do Moses calculate the probability of a sentence with an unknown > > word? How is the LM used? > > > > I've estimated a 3-gram LM with IRSTLM for a base line system, according > > to the instructions in the Wiki. The arpa-file contains entries like: > > > > -7.2625 redescendue -0.1681 > > -7.2625 serviabilité -0.1681 > > -2.51072 <unk> > > > > -3.26915 cadre très -0.096544 > > -4.52727 cadre lors > > -4.57217 cadre seulement > > > > I suppose the first number is the probability and the second number is > > the "back-off weight". Is it used somehow? In that case, what happens > > when it's absent (4.52727 cadre lors) ? > > > > 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
