search

lizekang / ITDD

The source code of our ACL2019 paper "Incremental Transformer with Deliberation Decoder for Document Grounded Conversations "
MIT License
86 stars 17 forks source link

PPL scores #14

Closed dishavarshney082 closed 4 years ago

dishavarshney082 commented 4 years ago

While training I am observing that the ppl score decreases upto 10000 steps but it starts increasing after that. I am not able to understand this behaviour. Do you have any idea ?

lizekang commented 4 years ago

Maybe it's overfitted. But perplexity can not represent the quality of the response. Experimentally, the response with a little bit of overfitting is better than the response with the lowest ppl score.