vq-vae-2

While primarily used for generative tasks in computer vision, VQ-VAE-2 has also been adapted for NLP tasks. It uses a discrete latent space and hierarchical structure.