Comment by pona-a

I wonder if these N-gram reduced models, augmented with confidence measures, can act as a very fast speculative decoder. Or maybe the sheer number of explicit rules unfolded from the compressed latent representation will make it impractical.