Non-Autoregressive Neural Machine Translation: A Call for Clarity

TL;DR

This paper critically evaluates non-autoregressive neural machine translation, emphasizing the importance of standardized benchmarks and providing insights to establish strong baselines for future research.

Contribution

It offers a comprehensive comparison of techniques for non-autoregressive translation, introduces standardized evaluation protocols, and provides open-source code for reproducibility.

Findings

Standardized BLEU, chrF++, and TER scores for four translation tasks.

Insights into effective baseline configurations using length prediction and CTC architectures.

Identification of inconsistencies in previous BLEU score reporting.

Abstract

Non-autoregressive approaches aim to improve the inference speed of translation models by only requiring a single forward pass to generate the output sequence instead of iteratively producing each predicted token. Consequently, their translation quality still tends to be inferior to their autoregressive counterparts due to several issues involving output token interdependence. In this work, we take a step back and revisit several techniques that have been proposed for improving non-autoregressive translation models and compare their combined translation quality and speed implications under third-party testing environments. We provide novel insights for establishing strong baselines using length prediction or CTC-based architecture variants and contribute standardized BLEU, chrF++, and TER scores using sacreBLEU on four translation tasks, which crucially have been missing as inconsistencies in the use of tokenized BLEU lead to deviations of up to 1.7 BLEU points. Our open-sourced code is integrated into fairseq for reproducibility.