Questions about the paper

[BlackLM12138]

Q1:
"In RSGAN, the generator G aims to generate realistic comments to fool a discriminator D, while the discriminator D aims to discriminate whether the comment c is more realistic than randomly sampled fake data generated by G." How do you select the comment c ? randomly select? Or each comment has a generated one to pair?
Q2:
Using the BLEU metric needs a reference. However, the dataset has no such reference of the generated comments, then how do you select the ground-truth reference to compare with the generated comments by BLEU?

Thank you