latent 最近

（1）Jointly Optimizing Diversity and Relevance in Neural Response Generation
背景:
bland and generic responses.->diversify, cost of decreased relevance
以前使用CVAE实现diverse，VAE里面用高斯分布，高斯分布决定了diverse和relevant不可兼得
并且VAE存在一些问题
1）it suffers from the vanishing latent vari- able problem when applied to text generation tasks.->KL annealing method
2)VAE/CVAE conversation models can be limited to a simple latent representations such as standard Gaussian distribution->enrich the latent space by leveraging a Gaussian mixture prior
前人的方法(similiar to Luan2017)：
（误）

现在的方法：

两个要点：
1.插值法，提出$Z_{interp}$ = $u{Z}_{S2S}$