Deep Latent-Variable Models For Neural Text Generation