hengyicai / Adaptive_Multi-curricula_Learning_for_Dialog

The codebase for "Learning from Easy to Complex: Adaptive Multi-curricula Learning for Neural Dialogue Generation" (Cai et al., AAAI 2020)
MIT License
19 stars 3 forks source link

Model Confidence score #2

Open LuckyVicky001 opened 3 years ago

LuckyVicky001 commented 3 years ago

Hi, For the Model Confidence score, is the underlying dialogue model SEQ2SEQ? But I also found loss of dialogwae/hred/transformer/cvae. What's the meaning of these values and how did you use them? Thx, again.

hengyicai commented 2 years ago

Sorry for the late reply. When applying the "model confidence" curriculum on the specific model architecture (e.g., transformer or hred), the loss computed by the corresponding model architecture can be used to measure the "model confidence" score. You can also regard the seq2seq loss as "model confidence" for all the model architectures. In experiments, we did not find significant differences between these two options.