jarib / bert-norne

Setup

$ git submodule init && git submodule update
$ ./bin/fetch-norne.sh
$ pip install torch seqeval tensorboardX
$ cd transformers && pip install -r requirements.txt && pip install . && cd -
$ ./bin/train-ner.sh

Results

f1 = 0.8427046263345196
loss = 0.03901922195237588
precision = 0.8297126839523475
recall = 0.8561099060014461

The multilingual BERT model appears to perform worse on NER than jarib/spacy-nb.