摘要
为促进文本域与视觉域之间的语义关联,提高英文翻译的准确性,提出一种基于深度跨域中介模块(deep crossdomain intermediary module,DCIM)的英文翻译方法,区别于传统的联合空间学习方法,DCIM作为中介模块指导图像中感兴趣区域与源词之间的交互。通过强化文本编码器和视觉编码器提取的表征之间的语义关联,使视觉表征在语义上得到增强并演化为跨域表征。通过双重注意力解码器,利用两种独立的注意力机制分别处理文本和跨域表征的上下文,预测目标词汇。实验结果表明,集成DCIM的模型在两种英文翻译任务中基于Meteor和Ribes指标取得了最佳翻译准确性和鲁棒性。
To promote semantical association between the text and visual domains and improve the English translation accuracy,an English translation method based on the deep cross-domain intermediary module(DCIM)was proposed.Unlike traditional joint space learning methods,DCIM served as an intermediary module to guide interactions between regions of interest(ROI)in images and source words.The semantical association between representations extracted by text encoders and visual encoders was enhanced,resulting in the evolution of visual representations into cross-domain representations.A dual-attention decoder was employed to utilize two independent attention mechanisms to process the contexts of text and cross-domain representations,thereby predicting target words.Experimental results show that the model integrated with DCIM achieves the best translation accuracy and robustness based on Meteor and Ribes metrics in two English translation tasks.
作者
黄笑菡
程时伟
张纪林
HUANG Xiao-han;CHENG Shi-wei;ZHANG Ji-lin(School of Social Development and Public Education,Huzhou Vocational and Technical College,Huzhou 313000,China;College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310014,China;School of Computer Science and Technology,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《计算机工程与设计》
北大核心
2025年第9期2502-2508,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(62172368)
浙江省自然科学基金项目(LQ23F020015)
浙江省教学改革基金项目(jg20230200)
湖州职业技术学院“镜湖人才”基金项目。
关键词
深度跨域中介模块
英文翻译
双重注意力解码器
文本域
视觉域
神经网络
自然语言处理
deep cross-domain intermediary module
English translation
double attention decoder
text domain
visual domain
neural network
natural language processing
作者简介
黄笑菡(1985-),女,浙江湖州人,硕士,副教授,研究方向为计算机辅助翻译、英语知识图谱;程时伟(1981-),男,浙江杭州人,博士,教授,研究方向为人机交互、普适计算;张纪林(1980-),男,浙江杭州人,博士,教授,研究方向为云计算与云存储、分布式机器学习等。E-mail:ern94ee61@21cn.com。
