Promises and Pitfalls of Generative Masked Language Modeling: Theoretical Framework and Practical Guidelines

TL;DR

This paper develops a mathematical framework for Generative Masked Language Models (GMLMs), analyzes their capabilities and limitations, and demonstrates practical improvements in machine translation speed with minimal quality loss.

Contribution

It provides the first comprehensive theoretical analysis of GMLMs, proposes an iterative decoding method, and offers practical guidelines for model design and optimization.

Findings

Achieved 2-3x speedup in machine translation with minimal quality loss.

Developed a mathematical framework analyzing sample complexity and inference trade-offs.

Provided empirical recommendations and insights into error modes of GMLMs.

Abstract

Autoregressive language models are the currently dominant paradigm for text generation, but they have some fundamental limitations that cannot be remedied by scale-for example inherently sequential and unidirectional generation. While alternate classes of models have been explored, we have limited mathematical understanding of their fundamental power and limitations. In this paper we focus on Generative Masked Language Models (GMLMs), a non-autoregressive paradigm in which we train a model to fit conditional probabilities of the data distribution via masking, which are subsequently used as inputs to a Markov Chain to draw samples from the model, These models empirically strike a promising speed-quality trade-off as each step can be typically parallelized by decoding the entire sequence in parallel. We develop a mathematical framework for analyzing and improving such models which sheds light on questions of sample complexity and inference speed and quality. Empirically, we adapt the T5 model for iteratively-refined parallel decoding, achieving 2-3x speedup in machine translation with minimal sacrifice in quality compared with autoregressive models. We run careful ablation experiments to give recommendations on key design choices, and make fine-grained observations on the common error modes in connection with our theory. Our mathematical analyses and empirical observations characterize both potentials and limitations of this approach, and can be applied to future works on improving understanding and performance of GMLMs. Our codes are released at https://github.com/google-research/google-research/tree/master/padir