Fitting the CRFTagger on NER task with a lot of training data fails on too small memory

[FelixKirsch]

Fitting the CRFTagger on NER tasks with large training data needs a lot of memory. The active learning fails if the memory assigned to the docker container is too small. To fix this behavior training data has to be used in batches or maybe something like numpy memmap can be used. Or check how other ml models tackle this issue.