AligNART: Non-autoregressive Neural Machine Translation by Jointly Learning to Estimate Alignment and Translate
This work addresses translation inconsistency for machine translation researchers and practitioners, offering an incremental improvement by explicitly modeling alignment to simplify decoding.
The paper tackles the multi-modality problem in non-autoregressive neural machine translation, which causes issues like token repetition, by introducing AligNART, a model that uses full alignment information to explicitly reduce target distribution modality, achieving BLEU scores comparable to state-of-the-art models on WMT14 En↔De and outperforming previous non-iterative models on WMT14 En↔De and WMT16 Ro→En.
Non-autoregressive neural machine translation (NART) models suffer from the multi-modality problem which causes translation inconsistency such as token repetition. Most recent approaches have attempted to solve this problem by implicitly modeling dependencies between outputs. In this paper, we introduce AligNART, which leverages full alignment information to explicitly reduce the modality of the target distribution. AligNART divides the machine translation task into $(i)$ alignment estimation and $(ii)$ translation with aligned decoder inputs, guiding the decoder to focus on simplified one-to-one translation. To alleviate the alignment estimation problem, we further propose a novel alignment decomposition method. Our experiments show that AligNART outperforms previous non-iterative NART models that focus on explicit modality reduction on WMT14 En$\leftrightarrow$De and WMT16 Ro$\rightarrow$En. Furthermore, AligNART achieves BLEU scores comparable to those of the state-of-the-art connectionist temporal classification based models on WMT14 En$\leftrightarrow$De. We also observe that AligNART effectively addresses the token repetition problem even without sequence-level knowledge distillation.