【摘要】 现有的篇章级神经机器翻译方法难以有效挖掘目标端远距离的上下文信息，翻译的译文不连贯.为此，文中提出融合目标端历史信息的篇章级神经机器翻译方法.首先，通过多头自注意力机制，获得源语言的上下文表征和目标语言的上文表征.然后，使用线性偏置注意力机制，动态地将历史信息注入当前目标语言表征.最后，通过融合源语言表征和经过增强后的目标语言上下文表征获得较优的篇章译文.在多个数据集上的实验表明，文中方法性能较优，在解码过程中融合通过循环机制建模的长序列信息，可有效提升篇章译文的连贯性和完整性.更多还原

【Abstract】 Existing document-level neural machine translation methods struggle to effectively capture long-distance contextual information on the target side, resulting in incoherent translations. To address this issue, a method for document-level neural machine translation with target-side historical information fusion is proposed. First, the contextual representations of the source language are derived via a multi-head self-attention mechanism. Second, the preceding context representations of the target language are obtained using another multi-head self-attention mechanism. Next, an attention with linear biases is employed to dynamically inject the historical information into the current target language representation. Finally, a higher-quality translation is obtained by integrating the source language representation with the enhanced preceding context representation of the target language. Experimental results on multiple datasets demonstrate that the performance of the proposed method is superior. Moreover, the proposed method effectively improves the coherence and completeness of document-level translations through incorporating long-sequence information modeled by recurrent mechanisms during decoding.更多还原