ondrejklejch / acoustic_punctuation Goto Github PK

Hello，I wrote a seq2seq model with Pytorch to do punctuation restoration. And there is a question

My model:

• architechture
  • encoder: GRU
  • decoder: GRU with attention
  • loss: Cross Entropy after a softmax
• output label
  • 5 types
  • (1) ',' (2) '.' (3) '!' (4) '?' (5) 'blank'

But after experiments, I found my model only learning to pridict 'blank'. The model tends to predict all input words to be a 'blank'. What can I do to fix this? A special loss or other thing?

Hi @ondrejklejch, can you please add license?