Deep Encoder, Shallow Decoder: Reevaluating Non-autoregressive Machine Translation

Much recent effort has been invested in non-autoregressive neural machine\ntranslation, which appears to be an efficient alternative to state-of-the-art\nautoregressive machine translation on modern GPUs. In contrast to the latter,\nwhere generation is sequential, the former allows generation to be parallelized\nacross target token positions. Some of the latest non-autoregressive models\nhave achieved impressive translation quality-speed tradeoffs compared to\nautoregressive baselines. In this work, we reexamine this tradeoff and argue\nthat autoregressive baselines can be substantially sped up without loss in\naccuracy. Specifically, we study autoregressive models with encoders and\ndecoders of varied depths. Our extensive experiments show that given a\nsufficiently deep encoder, a single-layer autoregressive decoder can\nsubstantially outperform strong non-autoregressive models with comparable\ninference speed. We show that the speed disadvantage for autoregressive\nbaselines compared to non-autoregressive methods has been overestimated in\nthree aspects: suboptimal layer allocation, insufficient speed measurement, and\nlack of knowledge distillation. Our results establish a new protocol for future\nresearch toward fast, accurate machine translation. Our code is available at\nhttps://github.com/jungokasai/deep-shallow.\n