[Paper Review] Deep Encoder, Shallow Decoder: Reevaluating Non-autoregressive Machine Translation

Jungo Kasai, Nikolaos Pappas+3 more|arXiv (Cornell University)|Jun 18, 2020

Natural Language Processing Techniques53 references64 citations

TL;DR

The paper argues that autoregressive models with a deep encoder and shallow decoder can outperform strong non-autoregressive models at similar speeds, and that traditional NAR evaluations underestimate AR speedups due to layer allocation, speed measurement, and distillation practices.

ABSTRACT

Much recent effort has been invested in non-autoregressive neural machine translation, which appears to be an efficient alternative to state-of-the-art autoregressive machine translation on modern GPUs. In contrast to the latter, where generation is sequential, the former allows generation to be parallelized across target token positions. Some of the latest non-autoregressive models have achieved impressive translation quality-speed tradeoffs compared to autoregressive baselines. In this work, we reexamine this tradeoff and argue that autoregressive baselines can be substantially sped up without loss in accuracy. Specifically, we study autoregressive models with encoders and decoders of varied depths. Our extensive experiments show that given a sufficiently deep encoder, a single-layer autoregressive decoder can substantially outperform strong non-autoregressive models with comparable inference speed. We show that the speed disadvantage for autoregressive baselines compared to non-autoregressive methods has been overestimated in three aspects: suboptimal layer allocation, insufficient speed measurement, and lack of knowledge distillation. Our results establish a new protocol for future research toward fast, accurate machine translation. Our code is available at https://github.com/jungokasai/deep-shallow.

Motivation & Objective

• Question conventional NAR speed-accuracy tradeoffs and evaluation practices.
• Investigate how encoder/decoder depth allocations affect AR and NAR performance.
• Assess the impact of knowledge distillation on both AR and NAR baselines under fair comparisons.
• Provide a revised protocol for evaluating fast, accurate MT models.

Proposed method

• Systematically compare autoregressive (AR) and non-autoregressive (NAR) models with varied encoder/decoder depths.
• Introduce and evaluate deep encoder–shallow decoder configurations for both AR and NAR.
• Measure inference speed using two metrics: S1 (one sentence at a time) and Smax (max batch size on hardware).
• Apply sequence-level knowledge distillation to both AR and NAR baselines for fair comparison.
• Analyze complexity and discuss how decoding iterations (T for NAR) impact total compute and speed.
• Conduct large-scale experiments on multiple WMT directions with standard preprocessing and evaluation (BLEU, SacreBLEU).

Experimental results

Research questions

• RQ1 Does a deep encoder with a shallow decoder provide a better speed-quality tradeoff for AR than for NAR?
• RQ2 How do speed measurements (S1 vs Smax) affect perceived advantages of AR vs NAR?
• RQ3 What is the impact of encoder/decoder layer allocation on translation quality and decoding speed?
• RQ4 Do knowledge distillation practices need to be applied equally to AR and NAR to ensure fair comparisons?
• RQ5 To what extent can AR models be sped up without sacrificing accuracy compared to strong NAR methods?

Key findings

• AR with a deep encoder and shallow decoder achieves comparable BLEU to a strong 6-6 AR baseline but with substantially faster S1 decoding.
• NAR models under a deep encoder–shallow decoder configuration generally lose BLEU compared with AR, and also show slower Smax performance than AR baselines.
• Under large-batch decoding, AR speedups remain robust while NAR speedups diminish as batch size grows.
• Knowledge distillation benefits both AR and NAR, but the accuracy gap between AR and NAR remains wide, widening when distillation is applied to both.
• Word order reordering and the number of decoder layers are key factors driving higher decoder depth requirements for NAR to perform well.