On the low-density latent regions of VAE-based language models