Semantic-aware conditional variational autoencoder for one-to-many dialogue generation