Question about the paper

[TianyuZhuuu]

Dear authors,

Thank you for your contribution and congratulates!

After reading the paper, I'm still have some confusion.

1. For the abstract generator, does it reuse the hidden state produced by Input Article Encoder as encoder hidden states? or it concatenate all the selected sentences and run an encoder on top of that?

2. The loss function of the coherence model:
   
   The range of the coherence score is [-1, 1]. Ideally,
   , even if the pos/neg examples are correctly classified, the loss is 3. How to interpret the loss function?

Looking forward to your reply!