Tom Hoar pointed this out a while ago because KenLM's build_binary rejects positive log probabilities. Eventually, I gave in and added a -i (i is short for IRST bug) option to build_binary that maps positive log probability to 0.0. Happens just about every time somebody builds a decently-sized model with IRSTLM. . .
On 09/27/11 07:54, Nicola Bertoldi wrote: > Let's move into a private thread for next mails > to not bother moses-support list > > Please send me as much logging information as possible > > > Furthermore, try to compute the likelihood of the target sentence > which gives the positive logprob with compile-lm program of IRSTLM > run the command as follows > > echo "_your_setence_" | compile-lm _yourLM_ --eval /dev/stdin -d 3 > > > and collect stdout and sterr > > Which are the parameters you used for the estimation of your LM? > > Which version of IRSTLM are you using? > > cheers, > Nicola Bertoldi > > > On Sep 26, 2011, at 9:04 AM, Abbasov Samir wrote: > >> Hi, i have trained 6 gram language model with IRSTLM on billion words >> tokens, but some log probabilities are positive, which they must be >> negative, why? > > > _______________________________________________ > Moses-support mailing list > Moses-support@mit.edu > http://mailman.mit.edu/mailman/listinfo/moses-support _______________________________________________ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support