dialog context c:前k-1個utterances，conversational floor（1或0）,meta features(topic)
latent variable z: capture distribution of valid responses
x: response utterance
y:linguistic features(knowledge-guided CVAE)

p(z|c):prior network
p(x|z,c): response decoder，用q(x|z,c) recognition network來模擬
生成過程：sample z, generate x如圖c所示

訓(xùn)練過程，如圖b所示，通過max L目標函數(shù)得到q p的兩個參數(shù)，從而得到z的分布。再由z生成response

The neural network architectures for the baseline and the proposed CVAE/kgCVAE models. The dashed blue connections only appear in kgCVAE