Core Concepts
VOLTA introduces a novel framework that enhances generative diversity in natural language generation by combining Transformer models with VAE and InfoGAN, improving quality while maintaining diversity.
Abstract
The paper introduces VOLTA, a framework that bridges Transformer models with VAE and InfoGAN to enhance generative diversity.
VOLTA improves generative diversity by introducing a cross-attention-based connection between the latent space and the decoder.
The framework accommodates various Transformer architectures and supports both continuous and discrete inputs.
Comprehensive experiments across different NLG tasks demonstrate VOLTA's effectiveness in enhancing generative diversity while upholding quality.
Stats
"Our model utilizes a default configuration comprising 32 Gaussian latent variables, along with 4 uniform latent codes."
"The VAE components in VOLTA add a mere 0.46M parameters."