search

hiroki13 / span-based-srl

46 stars 9 forks source link

Let's beat the state of the art once again! #6

Open LifeIsStrange opened 4 years ago

LifeIsStrange commented 4 years ago

You are still the number one state of the art at semantic role labeling. But in the meantime there was a lot of progress in the Nlp world. ELMO has been succeeded by BERT, bert has been succeeded by spanBert and all of them have been succeeded by XLnet!

There is a popular library for easily switching from one transformer to another: https://github.com/huggingface/transformers

If you trained again your model, with XLnet instead of ELMO you would probably significantly increase accuracy at SRL!

Humanity needs more accurate SRL, please :}

LifeIsStrange commented 4 years ago

@hiroki13

logicReasoner commented 3 years ago

@LifeIsStrange Yeah, we need even better SRL. Still too many dumb errors in the output of the current SOTA :( Which model do you currently use?