Since Giza is in effect running EM, you are at the mercy of however it gets initialised. Furthermore, if there are ties, results will depend upon how they are broken.
(Note also it is possible to have the same perplexity, but different parameterisations. This is because you may have a plateau on the likelihood surface) I would check to see how the models are initialised in the first instance Miles 2008/7/16 John D. Burger <[EMAIL PROTECTED]>: > Hi - > > I have recently run GIZA twice on the exact same input data, on the > same machine, with very different results. In the one case, it > finished normally, in the other, I got hillclimbing warnings: > > WARNING: already 41 iterations in hillclimb: 1.10041 2 33 26 > WARNING: already 42 iterations in hillclimb: 1.00001 0 33 26 > > and then the dreaded NaNs: > > THTo3: Iteration 1 > Reading more sentence pairs into memory ... > #centers(pre/hillclimbed/real): nan nan nan #al: nan > #alsophisticatedcountcollection: nan #hcsteps: nan > #peggingImprovements: nan > A/D table contains 0 parameters. > A/D table contains 0 parameters. > p0_count is 0 and p1 is 0; p0 is 0.999 p1: 0.001 > THTo3: TRAIN CROSS-ENTROPY nan PERPLEXITY nan > > As far as I can tell, these runs should be =exactly= the same - same > GIZA executable, same input data, same config file. And, indeed, the > cross-entropy and perplexity figures that GIZA prints out after each > model iteration match exactly ... until they don't, on the THTo3 line > above. > > Has anyone else experienced this? Is there a random component to the > model 3 code that might explain this? > > Sorry if this is something everybody already knows about. Thanks for > any info you can provide. > > - John D. Burger > MITRE > > > _______________________________________________ > 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
