Text Adversarial Generation Based on Latent Variable Models

Abstract:
Variational auto-encoders (VAEs) have received much attention recently as an end-to-end architecture for text generation with latent variables. The variational auto-encoder model's encoder cannot fully extract the semantic information of text, which leads to the poor quality of text generation. A text adversarial generation model named VSGAN based on variational auto-encoder is proposed. VSGAN uses VAE to extract short-distance text features and long-distance text features, constructs fusion hidden variables after full connection, and generates text sequences by decoding fusion hidden variables. The intensive adversarial training method is introduced to improve the quality of text generation. Experimental results show that the accuracy and authenticity of the proposed model are further improved compared with the baseline model.