BANG: Bridging Autoregressive and Non-autoregressive Generation with Large Scale Pretraining

TL;DR

BANG is a large-scale pretraining model that unifies autoregressive and non-autoregressive generation, improving performance across various NLP tasks by supporting multiple generation modes.

Contribution

The paper introduces BANG, a novel pretraining framework that bridges AR and NAR generation, enabling flexible support for different generation paradigms in a single model.

Findings

BANG significantly improves NAR and semi-NAR task performance.

BANG achieves comparable results to strong AR models.

BANG outperforms strong NAR baselines on multiple datasets.

Abstract

In this paper, we propose BANG, a new pretraining model to Bridge the gap between Autoregressive (AR) and Non-autoregressive (NAR) Generation. AR and NAR generation can be uniformly regarded as to what extent previous tokens can be attended, and BANG bridges AR and NAR generation by designing a novel model structure for large-scale pretraining. The pretrained BANG model can simultaneously support AR, NAR and semi-NAR generation to meet different requirements. Experiments on question generation (SQuAD 1.1), summarization (XSum) and dialogue generation (PersonaChat) show that BANG improves NAR and semi-NAR performance significantly as well as attaining comparable performance with strong AR pretrained models. Compared with the semi-NAR strong baselines, BANG achieves absolute improvements of 14.01 and 5.24 in the overall scores of SQuAD 1.1 and XSum, respectively. In addition, BANG achieves absolute improvements of 10.73, 6.39 and 5.90 in the overall scores of SQuAD, XSUM and PersonaChat respectively compared with the strong NAR baselines.