search

hsgodhia / hred

Implements the paper " Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models" by Serban et al (currently on the MovieTriples dataset)
119 stars 30 forks source link

decode for generating only one response? #3

Open luomuqinghan opened 6 years ago

luomuqinghan commented 6 years ago

Thank you for your code! But in original HRED, for k context and one response, HRED generates k utterances. It seems that you only generate the final response in training process?

hsgodhia commented 6 years ago

hi @luomuqinghan yes currently one the final utterance is decode, after reading the paper it seemed like that may be the right approach? Do you think every utterance along the way must be decoded based on the previous history?

luomuqinghan commented 6 years ago

Yes, it seems that the paper of "HRED" did in this way when training. In test, it did in your way.