ShannonAI / mrc-for-flat-nested-ner

Code for ACL 2020 paper `A Unified MRC Framework for Named Entity Recognition`
662 stars 118 forks source link

MSRA数据集start,end很快收敛,span的recall <50% #106

Open DSXiangLi opened 2 years ago

DSXiangLi commented 2 years ago

你好我在MSRA中文数据集上训练,发现start end loss很快就收敛了,但是span部分的recall上升非常缓慢。看了下metric因为span部分正负样本相差很大,请问有什么解决方案么?