Anchor with just one feature about anchor HOT 2 CLOSED

Take a look at the examples (click the little boxes below). I'm guessing the model being used has a bias for predicting positive, such that even replacing most tokens in your long review with UNK does not cause a prediction of negative. If that is the case, an empty anchor or any arbitrary anchor is 'correct', i.e. under the assumption that the perturbation is replacing things with UNK this is a sufficient condition for a prediction of 'positive'.

from anchor.

Ah okay, I understand it now, thanks!

It just surprised me that out of this long text as input, only one word was selected as the anchor. But I understand it with those examples better.

from anchor.