Non-autoregressive Machine Translation with Target Language Components in English Corpus

Non-autoregressive Transformer (NAT) neural machine translation models generate target words in parallel, but their word-by-word independence often results in repetitive or missing translations, yielding lower quality compared to autoregressive models. To address this limitation, we propose a NAT model that integrates target language syntactic components through a dual-decoder framework. The first decoder produces an initial translation, while the second iteratively refines it using sentence constituent attention, embedding richer contextual and structural knowledge of the target language. Experiments on WMT14 and WMT16 benchmarks show improvements of 0.615 BLEU and 1.05 BLEU, respectively, over the strongest baseline models, demonstrating the effectiveness of the proposed approach.