A question about “Automatically Identifying Words That Can Serve as Labels for Few-Shot Text Classification”

In the paper “Automatically Identifying Words That Can Serve as Labels for Few-Shot Text Classification”, I do not quite understand the following. Could you please give me an explanation?