Abstract: Despite years of development and significant progress in document-level neural machine translation, most efforts have focused on building effective network structures from a model perspective by utilizing contextual word information, neglecting the guidance of cross-sentence discourse structure and rhetorical information for the model. Addressing this issue, under the guidance of Rhetorical Structure Theory, a method for separately encoding discourse units and rhetorical structure tree features is proposed. Experimental results show that the proposed  method enhances the encoders ability to represent discourse structure and rhetorical aspects. The improved model surpasses several high-quality baseline models, achieving notable improvements in translation performance across multiple datasets. Additionally, significant improvements in translation quality are demonstrated through the proposed quantitative evaluation method and human analysis.

Key words: neural machine translation, discourse analysis, document-level translation, rhetorical structure theory