CycleCVAE+T5 Implementation Details

[alxthm]

Hi,

Thank you for your work!

In the Fork or Fail paper, you mention using the T5 pretrained seq2seq model for the graph-to-text part (h_theta(y,z)) with y the linearized graph as a sequence input. Since you cannot simply concatenate z to the embedding of the graph y (like the BiLSTM model), I was wondering how you add the latent variable z to the input of the T5 seq2seq model?

Thanks,
Alexandre