Transformers: if crf needed when do ner?
If crf needed when do ner? In BertForTokenClassification, just Linear is used to predict tag. If not, why?
alphanlp
All 3 comments
A CRF gives better NER F1 scores in some cases, but not necessarily in all cases. In the BERT paper, no CRF is used and hence also no CRF in this repository. I'd presume the BERT authors tested both with and without CRF and found that a CRF layer gives no improvement, since using a CRF is kind of the default setting nowadays.
bheinzerling
on 3 Apr 2019
👍3
❤1
Issue #64 is a good reference for discussion on NER.
thomwolf
on 3 Apr 2019
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
![stale[bot] picture](https://avatars3.githubusercontent.com/in/1724?v=4&s=40)
stale[bot]
on 2 Jun 2019
Was this page helpful?
0 / 5 - 0 ratings
Related issues
rsanjaykamath
路
3Comments
guanlongtianzi
路
3Comments
0x01h
路
3Comments
iedmrc
路
3Comments
hsajjad
路
3Comments
Most helpful comment
A CRF gives better NER F1 scores in some cases, but not necessarily in all cases. In the BERT paper, no CRF is used and hence also no CRF in this repository. I'd presume the BERT authors tested both with and without CRF and found that a CRF layer gives no improvement, since using a CRF is kind of the default setting nowadays.