Visualize: Figure1;
Description: "We then use these features to
augment the input representation. Just as a semantic parser grounds an explanation by converting it into a logical form and then executing it, the features produced by BERT can be seen as a soft “execution” of the explanation on the input."