The Non-autoregressive Translation(NAT)model is a translation model that has made a sig-nificant breakthrough in inference speed by eliminating order dependencies in target sentences.However,the model has some limitations in translation quality.To investigate the fundamental reasons for this,this paper conducts a detailed and comprehensive preliminary study of attentional mechanisms,and the results reveal that the NAT model is inadequate in capturing localization features.Therefore,in this paper,a method for impro-ving the localization capability of the NAT model by explicitly introducing the surrounding lexical information is proposed.Specifically,Hybrid Grouped Linear Transformations are introduced in both the encoder and decod-er directions to attain representations that are more locally sensitive.Through experiments on the WMT14 Eng-lish-German and WMT16 English-Romanian datasets,the results demonstrate that the method improves the BLEU scores of 0.7 and 1.03 at a minor speed cost,respectively,which suggests that the proposed method has a significant effect and potential to enhance the localizability feature extraction of NAT models.
关键词
非自回归/局部性特征/混合分组线性变换/自回归
Key words
non-autoregressive/local characteristics/hybrid grouped linear transformations/autore-gressive
维普
万方数据