厦门大学学报（自然科学版）2025，Vol.64Issue(6)：970-982,13.DOI:10.6043/j.issn.0438-0479.202412008
基于多任务学习的词汇约束神经机器翻译方法
A lexically constrained neural machine translation method based on multi-task learning
摘要
Abstract
[Objective]Neural machine translation(NMT)refers to using neural network models to translate source language into accurate target language,whereas lexically constrained translation(LCT)requires the generated translation to include pre-specified segments in the target language.The end-to-end training approach employed by NMT models prioritizes overall sequence optimization over the forced alignment of specific lexical items,hence making it challenging for existing LCT methods to strike a balance between translation quality and constraint accuracy.[Method]To address this issue,we propose a multi-task learning-based LCT method that leverages inductive biases from multiple tasks to guide the model in developing translation capabilities for lexical constraint scenarios.This method designates LCT as the primary task and incorporates auxiliary tasks such as NMT,target-side monolingual text generation(TMTG),and source-side token type labeling(STTL)for the multi-task learning.[Results]Experimental results demonstrate that the proposed method achieves improvements over the current state-of-the-art baseline methods,with a BLEU score increase of 2.28,a BLEURT score improvement of 0.62,and a Window-Overlap score increase of 1.50.[Conclusion]By leveraging information exchange and collaborative interactions among tasks,the proposed multi-task learning-based lexical constrained neural machine translation method enhances cross-task knowledge transfer,thus ensuring the accurate generation of constrained terms while maintaining high translation quality and fluency.关键词
神经机器翻译/词汇约束/多任务学习/知识融合Key words
neural machine translation(NMT)/lexical constraint/multi-task learning/knowledge fusion分类
信息技术与安全科学引用本文复制引用
叶娜,夏宇轩,张桂平,杨晨,王雪妮..基于多任务学习的词汇约束神经机器翻译方法[J].厦门大学学报（自然科学版）,2025,64(6):970-982,13.基金项目
国家自然科学基金(U1908216) （U1908216）
