ai-forever/ner-bert

Explanation for the tables

qiuwei opened this issue · 1 comments

qiuwei commented

Hi, what does "Total spans in test set" mean in the tables?

BTW, i got similar results on conll2003 data.
It seems bert is not so effective for NER.

king-menin commented

  1. Span is a concatination of the sequence: B_ORG, I_ORG, I_ORG => ORG. There are counting of this spans in the table. In all articles we can see metrics for IOB (as conll-2003, with B_ORG and so on).

  2. For example Google achived with BERTLARGE 96.6 (dev set) 92.8 (test set) on conll-2003 (as we can see in their paper).
    I think, that BERT is very effective and universal for all tasks 👍 .