Parallel Variational Autoencoders for Multiple Responses Generation

Variational autoencoders (VAEs) have been widely adopted in the field of diverse responses generation. The existing VAE-based responses generation models can be roughly divided into two categories: the first is to map multiple responses to a joint variable distribution; the second is to construct an individual variable distribution for each response. Models based on the joint distribution can generate diverse responses to some extent, but the difference among responses is inapparent. On the contrary, models based on the individual distribution can ensure the distinction among responses, while they are not good at modeling the correlation among responses. To combine these two kinds of VAEs and make the best of their respective advantages, we propose a Parallel Variational Autoencoders (PVAE) model that innovatively contains two variable distributions: correlation and distinction. The joint correlation distribution captures the common feature of all responses. Therefore, it can better retain the relevance to the given context. The individual distinction distribution models the unique content of each response, which promotes responses' diversity. In the generation step, keywords are sampled from these two distributions via the improved sampling process, which also enhances the diversity of responses. The experimental results demonstrate that our model achieves state-of-the-art performance on a weibo dataset.