CRF model is not saved to correct path when training NER about deeppavlov HOT 3 CLOSED

We will fix the bug in the next release.
Indeed, in some cases CRF does not have a big impact on token classification model, so you can set the flag "use_crf": false

from deeppavlov.

BTW results with and without CRF are pretty much the same. The CRF is probably good in combination with the original Bi-LSTM model (here) but it has no impact on model with BERT.

These experiments came to the same conclusion:

• NER-BERT-CRF (see section Performance)
• NER using BERT-CRF (see table 2 on page 6)

from deeppavlov.

#1637 fixed this issue. Will be in DeepPavlov 1.2.0.

from deeppavlov.