Closed LHRYANG closed 1 year ago
Dear Ximing,
It seems that the training process only uses the likelihood loss. After training the model, during testing, you modify the traditional beam search decoding by introducing additional constraints. Is it right?
Hi, thanks for your question! Yes you're right, in training time we only use likelihood loss, and enforcing constraints in decoding time only.
Dear Ximing,
It seems that the training process only uses the likelihood loss. After training the model, during testing, you modify the traditional beam search decoding by introducing additional constraints. Is it right?