Candidate Soups: Fusing Candidate Results Improves Translation Quality for Non-Autoregressive Translation

TL;DR

The paper introduces 'Candidate Soups', a method that combines multiple candidate translations in non-autoregressive models to significantly improve translation quality without sacrificing inference speed, outperforming autoregressive models in some cases.

Contribution

It proposes a novel ensemble-like approach called 'Candidate Soups' that leverages model uncertainty to enhance NAT translation quality, a significant step beyond existing methods.

Findings

Significant translation quality improvements on WMT benchmarks.

Best variant surpasses autoregressive models in three translation tasks.

Achieves 7.6 times faster inference than AT models.

Abstract

Non-autoregressive translation (NAT) model achieves a much faster inference speed than the autoregressive translation (AT) model because it can simultaneously predict all tokens during inference. However, its translation quality suffers from degradation compared to AT. And existing NAT methods only focus on improving the NAT model's performance but do not fully utilize it. In this paper, we propose a simple but effective method called "Candidate Soups," which can obtain high-quality translations while maintaining the inference speed of NAT models. Unlike previous approaches that pick the individual result and discard the remainders, Candidate Soups (CDS) can fully use the valuable information in the different candidate translations through model uncertainty. Extensive experiments on two benchmarks (WMT'14 EN-DE and WMT'16 EN-RO) demonstrate the effectiveness and generality of our proposed method, which can significantly improve the translation quality of various base models. More notably, our best variant outperforms the AT model on three translation tasks with 7.6 times speedup.