search

uclnlp / jack

Jack the Reader
MIT License
257 stars 82 forks source link

Fix DAM and make ESIM as close to paper as possible #357

Closed dirkweissenborn closed 6 years ago

pminervini commented 6 years ago

I've been working quite a lot for making those faithful to their respective papers:

https://github.com/uclmr/inferbeddings/blob/master/inferbeddings/nli/dam.py

https://github.com/uclmr/inferbeddings/blob/master/inferbeddings/nli/esim.py

dirkweissenborn commented 6 years ago

I know. DAM is mostly the same, ESIM is now built using our modular nli model (an example of how to use the modular nli model).

The problem with DAM is that you set the dropout rate to 0.8! You probably meant to set it to 0.2. Anyway, I will fix the NLI models, train and tweak them a little.

dirkweissenborn commented 6 years ago

results with the current implementations and configs are good enough for now