基于约束解码与最小贝叶斯的多模态语言翻译模型解码方法研究  

Research on Decoding Method for Multimodal Language Translation Model Based on Constrained Decoding and Minimum Bayes

作　　者：于英俊 YU Yingjun(Xi'an Fanyi University,Xi’an 710105,China)

机构地区：[1]西安翻译学院,西安710105

出　　处：《自动化与仪器仪表》2024年第11期172-176,共5页Automation & Instrumentation

基　　金：陕西省教育厅科学研究计划项目资助《延安红色旅游翻译与传播路径研究》(23JP053);2022年度陕西高校青年创新团队“中华优秀文化翻译与国际传播创新团队”阶段性成果。

摘　　要：针对跨语言翻译模型准确性较低,易产生歧义等问题,研究搭建了基于Transformer的多模态语言翻译模型,并提出了一种新的解码方法。结果表明,在英语-越语数据集中,完整模型的双语评估替补值最高,为28.51%。在newstest2017、newstest2018和cwmt2018数据集中,研究所提多模态翻译模型的双语评估替补值均最高,分别为24.52%,24.47%和24.48%。采用双语评估替补效用函数的模型在双语评估替补指标方面的表现最好。研究结果显示,所提翻译模型具有较好的语言翻译性能和准确性,能够有效对传统的Transformer模型进行改进,且在长句子翻译上具有较好的表现,因此,具有重要的实际应用价值和前景。研究能够为英文翻译提供一定的技术支持,促进跨文化传播以及跨国贸易的深入开展。A multi-modal language translation model based on Transformer was studied and constructed to address the issues of low accuracy and ambiguity in cross language translation models,and a new decoding method was proposed.The results showed that in the English Vietnamese dataset,the complete model had the highest substitute value for bilingual evaluation,at 28.51%.In the datasets of Newstest2017,Newstest2018,and CWMT2018,the multimodal translation model proposed by the research institute has the highest bilingual evaluation substitute values,at 24.52% and 24.47%,respectively.And 24.48%.The model that uses bilingual evaluation substitute utility function performs the best in bilingual evaluation substitute indicators.The research results show that the proposed translation model has good language translation performance and accuracy,can effectively improve the traditional Transformer model,and has good performance in long sentence translation.Therefore,it has important practical application value and prospects.Research can provide certain technical support for English translation,promote cross-cultural communication,and deepen cross-border trade.

关 键 词：翻译模型 最小贝叶斯 约束解码 多模态 英语翻译 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]