Autoregressive (AR) models represented by Transformer has attained the highest performance benchmarks in neural machine translation, which benefits from the powerful learning ability of the model with the attention mechanism. It differs from the way human translators translate a sentence, where prior knowledge plays a important role. Inspired by this, a knowledge distillation based translation prompt information fusion method is proposed to improve the AR model. It introduces two modules to improve the AR model: translation prompt information fusion and knowledge distillation. The main steps can be summarized as follows: Firstly, Training a non-autoregressive (NAR) model based on the bilingual corpus. Then, the translation prompt information generated by NAR are integrated into the AR model from two aspects. On the on hand, incorporating the translation produced by the NAR model into the decoder of the AR model. On the other hand, utilizing the output distribution of the NAR model to guide the output distribution of the AR model. Experimental results across several translation tasks with low-resource and rich-resource indicate the effectiveness of the proposed method.

Transformer-based Neural Machine Translation (NMT) models have attained state-of-the-art performance within the machine translation community, automatically acquiring translation knowledge from bilingual corpora through the attention mechanism. This contrasts with the approach of human translators, who heavily rely on prior knowledge during sentence translation. Inspired by this disparity, we propose a Word Translation Augmentation (WTA) method to enhance Transformer-based NMT models. Our methodology comprises three key steps: initially, we construct word alignment rules based on the training set; subsequently, we generate translation rules for source words in accordance with these alignment rules; finally, we integrate potential translation candidates for each source word into the NMT model during both training and testing phases. Additionally, the WTA method introduces the concept of Mixup for augmenting translation candidates of source words and employs two augmentation strategies to enrich the encoder. Experimental results on the WMT14 (German$\leftrightarrow$English) and AIChallenger 2018 (Chinese$\leftrightarrow$English) translation tasks demonstrate the effectiveness of our proposed method compared to strong baseline models.

Recent advancements in Transformer-based architectures have established new benchmarks in neural machine translation (NMT), though their efficacy remains constrained by the availability of large-scale parallel corpora. This limitation becomes particularly pronounced for low-resource language pairs, where the scarcity of bilingual training data often leads to suboptimal translation performance. To address this challenge, we introduce Keyword Exchange (KE), a novel data augmentation framework designed to enhance translation quality in data-scarce scenarios. The proposed methodology operates through a three-stage pipeline: (1) A selective filtering approach is first applied to identify and retain high-quality sentence pairs from both source and target language corpora; (2) The KE algorithm is then deployed on the filtered target-side sentences, strategically substituting keywords to generate syntactically valid yet semantically divergent pseudo-monolingual data; (3) Finally, an Alignment Mixture strategy synthesizes these augmented monolingual fragments with their corresponding source-side counterparts, producing a robust pseudo-parallel corpus that significantly expands the effective training set. Empirical validation across multiple low-resource translation benchmarks demonstrates the framework's superiority over conventional baselines. Notably, our approach achieves a maximum BLEU score improvement of 0.74 points compared to a strong Transformer-based baseline, outperforming several state-of-the-art data augmentation techniques.