A Generalized Framework of Sequence Generation with Application to Undirected Sequence Models

TL;DR

This paper introduces a unified framework for sequence generation that applies to both directed and undirected neural models, enabling new decoding strategies and achieving competitive translation results.

Contribution

It proposes a generalized model of sequence generation that unifies decoding in directed and undirected models, facilitating the adaptation of decoding algorithms across these paradigms.

Findings

Achieves constant-time translation comparable to linear-time methods.

Demonstrates competitive performance on WMT'14 English-German translation.

Unifies various neural sequence models under a single framework.

Abstract

Undirected neural sequence models such as BERT (Devlin et al., 2019) have received renewed interest due to their success on discriminative natural language understanding tasks such as question-answering and natural language inference. The problem of generating sequences directly from these models has received relatively little attention, in part because generating from undirected models departs significantly from conventional monotonic generation in directed sequence models. We investigate this problem by proposing a generalized model of sequence generation that unifies decoding in directed and undirected models. The proposed framework models the process of generation rather than the resulting sequence, and under this framework, we derive various neural sequence models as special cases, such as autoregressive, semi-autoregressive, and refinement-based non-autoregressive models. This unification enables us to adapt decoding algorithms originally developed for directed sequence models to undirected sequence models. We demonstrate this by evaluating various handcrafted and learned decoding strategies on a BERT-like machine translation model (Lample & Conneau, 2019). The proposed approach achieves constant-time translation results on par with linear-time translation results from the same undirected sequence model, while both are competitive with the state-of-the-art on WMT'14 English-German translation.