Hello, I do not know why is this happening, but while testing the lemmatizer the best performance (word and sentence accuracy) was always obtained using the Perceptron trainer (for a number of languages), so I would recommend to train perceptron models (perhaps you have already tried this, but just in case).
Best, R On Fri, 21 Feb 2020 at 16:07, Markus Jelsma <markus.jel...@openindex.io> wrote: > > Hello, > > When GaussianSmoothing is enabled, the LemmatizerTrainerME gets stuck. The > loglikelihood gets warped to NaN right after the first iteration, like so: > Computing model parameters in 16 threads... > Performing 10000 iterations. > 1: ... loglikelihood=-1238383.3965208859 0.8032247485201534 > 2: ... loglikelihood=NaN 0.8178054720724305 > 3: ... loglikelihood=NaN 0.8032247485201534 > 4: ... loglikelihood=NaN 0.8032247485201534 > > Is this known to happen? Did i stumble upon some bug? > > Many thanks, > Markus
