It will be better if there is a Coverage Attention Module


I Think this paper(Modeling Coverage for Neural Machine Translation) have done a great job, and I really found the over-translation and under-translation in the translate results given by OpenNMT.

Alternative methods for <UNK> substitution
(srush) #2

Thanks we will look into adding this. If you were interested in sending a PR we would be very thankful as well :slight_smile:


It’s hard for me to change OpenNMT with this complex module, I have some code in a issue( works for coverage but it is just a simplest version with the samiliar idea, I’m testing whether it could work now.