Step-unrolled Denoising Autoencoders for Text Generation

TL;DR

SUNDAE is a novel non-autoregressive text generation model that iteratively refines random inputs into high-quality text, achieving state-of-the-art results in translation and promising results in language modeling and code generation.

Contribution

The paper introduces SUNDAE, a step-unrolled denoising autoencoder that improves upon diffusion methods with fewer iterations and better sample quality for non-autoregressive text generation.

Findings

Achieves state-of-the-art non-autoregressive translation results on WMT'14 English-German.

Produces high-quality unconditional language modeling results.

Effectively generates Python code and fills in arbitrary text templates.

Abstract

In this paper we propose a new generative model of text, Step-unrolled Denoising Autoencoder (SUNDAE), that does not rely on autoregressive models. Similarly to denoising diffusion techniques, SUNDAE is repeatedly applied on a sequence of tokens, starting from random inputs and improving them each time until convergence. We present a simple new improvement operator that converges in fewer iterations than diffusion methods, while qualitatively producing better samples on natural language datasets. SUNDAE achieves state-of-the-art results (among non-autoregressive methods) on the WMT'14 English-to-German translation task and good qualitative results on unconditional language modeling on the Colossal Cleaned Common Crawl dataset and a dataset of Python code from GitHub. The non-autoregressive nature of SUNDAE opens up possibilities beyond left-to-right prompted generation, by filling in arbitrary blank patterns in a template.