#faithfulness

Question 1: Single choice

Jain and Wallace tested whether the attention weights of recurrent (BiLSTM) text classifiers explain their predictions. What did they find?

Permuted or adversarial attention weights often left the prediction almost unchanged1
Attention weights agreed closely with gradient-based and leave-one-out importance scores2
Removing the attention layer always destroyed accuracy, which proves attention is the explanation3
Attention became a reliable explanation once its weights were normalised by a softmax4