Context-Filter for Document-Level Neural Machine Translation
In the field of document neural machine translation,existing models often fail to effectively filter the incorporated contextual information,leading to increased noise and a decline in model performance.To address this issue,the paper proposes a machine translation method based on Context-Filtered Document Translation(CFDT).This method achieves fine-grained management and effective utilization of document contextual infor-mation by introducing Intra-Sentence Constraint Attention(ISCA)and Filter Attention(FA)mechanisms.The ISCA module focuses on intra-sentential information,enhancing the model's ability to represent long sequences;the FA module,on the other hand,filters out the most relevant contextual information for the current word transla-tion by calculating attention scores and applying masking operations,thereby excluding redundant and irrelevant information.Experimental results show that this method has achieved significant improvements in BLEU scores for Chinese-English and English-German translation tasks compared to existing models,proving the positive role of context information filtering in document-level neural machine translation.
万方数据
维普
