Repeated words as decoded/generated output :(

[thefirebanks]

Hello! I was wondering if there was a reason why the final decoded sentence would be just the same word repeated min_dec_steps times? I am training this model on the Quora Question Pair dataset for the purpose of paraphrase generation, and I am getting results as such:

Original sentence is: "what books and study materials i will need for political science and ir for cse optionals ?" Hypothesis is: "and and and and and and and and and and and and and and and and and and and and" Reference is: "which of the study material is useful for the upsc cse optional political science and international relations ?"

I know this was originally intended for summarization but your RLSeq2Seq model worked really well for paraphrasing! And I believe the transfer learning component of this model would be essential for generalization, as the current paraphrase generation models also suffer from bad generalization capabilities. Thank you!