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

Due to the miscellaneous ambiguity of semantics in open-domain conversation, current deep dialogue models disregard to detect potential emotional and action response features in the latent space, which leads to the general tendency to produce inaccurate and irrelevant sentences. To address this problem, we propose a semantic-aware conditional variational autoencoder that discriminates the sentiment and action responses features in the latent space for one-to-many open-domain dialogue generation. Specifically, explicit controllable variables are leveraged from the proposed module to create diverse conversational texts. This controllable variable can constrain the distribution of the latent space, disentangling the latent space features during training. Furthermore, the feature disentanglement improves the dialogue generation in terms of deep learning interpretability and text quality, which also reveals the latent features of different emotions on the logic of text generation.