As another member of the sesame street, BigBird beats BERT on GLUE tasks. The most signifcant improvement comes from the RTE task.
Paper: https://arxiv.org/pdf/1901.11504.pdf
A community version: https://github.com/namisan/mt-dnn
If you're interested in contributing this, we're more than happy to help. Let's coordinate here
MTL(MultiTask Learning) is one of the most important topic in NLP research with LM.
It's heavily related on developing MTL frame on gluon-nlp.
pytext(https://arxiv.org/pdf/1812.08729.pdf) already implement on MTL for production use.
Most helpful comment
MTL(MultiTask Learning) is one of the most important topic in NLP research with LM.
It's heavily related on developing MTL frame on gluon-nlp.
pytext(https://arxiv.org/pdf/1812.08729.pdf) already implement on MTL for production use.