Open alaakh42 opened 6 years ago
In fact, the model solves NER and uses the postags which are given during training AND testing. The labels are B-MISC I-MISC O ... not NB V ADJ....
So yes the results are good but remain on a prelabelling or a POS tagger, which can't be always available. You may never reach such result without POS tags.
In the
examples/CoNLL 2002.ipynb
directory the model uses the labels in the training step exactly inword2features
function . and yet you got an accuracy less than 100%