Improving Variational Encoder-Decoders in Dialogue Generation