Abstract
This paper introduces a novel variational autoencoder model termed DVAE to prevent posterior collapse in text modeling. DVAE employs a dual-path architecture within its decoder: path A and path B. Path A makes the direct input of text instances into the decoder, whereas path B replaces a subset of word tokens in the text instances with a generic unknown token before their input into the decoder. A stopping strategy is implemented, wherein both paths are concurrently active during the early phases of training. As the model progresses towards convergence, path B is removed. To further refine the performance, a KL weight dropout method is employed, which randomly sets certain dimensions of the KL weight to zero during the annealing process. DVAE compels the latent variables to encode more information about the input texts through path B and fully utilize the expressiveness of the decoder, as well as avoiding the local optimum when path B is active through path A and the stopping strategy. Furthermore, the KL weight dropout method augments the number of active units within the latent variables. Experimental results show the excellent performance of DVAE in density estimation, representation learning, and text generation.