This repo contains the code to attack BiDAF.
The BiDAF is trained by ourselves.
We put related code in root directory.
You can reproduce the results by using the model parameters uploaded here: model.bin
and ema
.
You may read our run.py
for more information. You may also try try different attack scenarios (position targeted attack/answer targeted attack) and different attack methods (word-level/sentence-level) to see the effects.