GEA-NMT:图卷积增强的多任务低资源机器翻译模型  

GEA-NMT:Graph Convolutional Network Enhanced Multi-task Adapting Neural Machine Translation

作　　者：张宝兴 彭敦陆[1] 王雅峰 ZHANG Baoxing;PENG Dunlu;WANG Yafeng(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Shanxi Provincial Natural Gas CO.,LTD,Xi′an 710016,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093 [2]陕西省天然气股份有限公司,西安710016

出　　处：《小型微型计算机系统》2024年第9期2156-2164,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61772342)资助.

摘　　要：在低资源语种上训练高质量机器翻译系统的主要障碍是平行语料的稀缺,一些低资源语言在语言上与高资源语言相关或相似,相似的语言可能使用相同的字符集或类似的句法结构.本文利用低资源语种上的单语数据及与之相关联的高资源语种和目标语言的平行语料,尝试将翻译知识进行迁移,从而完成低资源语种和目标语言间的翻译模型训练.本文提出一种图卷积增强的多任务训练低资源神经机器翻译模型(GCN Enhanced multi-task Adapting Neural Machine Translation,GEA-NMT),结合降噪自编码器、生成对抗训练、回译和桥接方法,并使用图神经网络学习句法依赖关系以进一步提升低资源语种上的神经机器翻译模型效果.The main obstacle to training high-quality machine translation systems on low-resource languages is the scarcity of parallel corpora.Some low-resource languages are linguistically related or similar to high-resource languages.Similar languages may use the same character set or similar syntactic structures.This paper utilizes monolingual data in low-resource languages and the associated parallel corpora of high-resource languages and target languages to try to transfer translation knowledge,so as to complete the translation model training between low-resource languages and target languages.This paper proposes a graph-convolution-network-enhanced multi-task adapting low-resource neural machine translation model(GEA-NMT),which combines denoising autoencoders,generative confrontation networks,back-translation and bridging methods,and uses Graph Neural Network to learn syntactic dependencies to further improve the effect of neural machine translation models on low-resource languages.

关 键 词：低资源神经机器翻译 图神经网络 多任务训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]