基于多粒度知识对齐的低资源机器翻译方法

王晓聪; 李英; 李衍铎; 高盛祥; 余正涛

doi:10.16112/j.cnki.53-1223/n.202511040002

昆明理工大学学报（自然科学版） ›› 2026, Vol. 51 ›› Issue (3) : 96 -108. DOI: 10.16112/j.cnki.53-1223/n.202511040002

计算机科学与技术

基于多粒度知识对齐的低资源机器翻译方法

王晓聪 ¹^,² ,
李英 ¹^,² ,
李衍铎 ¹^,² ,
高盛祥 ¹^,² ,
余正涛 ¹^,²^,^*

作者信息 +

Multi-granularity Knowledge Alignment for Low-Resource Machine Translation

Xiaocong WANG ¹^,² ,
Ying LI ¹^,² ,
Yanduo LI ¹^,² ,
Shengxiang GAO ¹^,² ,
Zhengtao YU ¹^,²^,^*

Author information +

文章历史 +

PDF (3048K)

摘要

低资源语言中,复合实体的深层语义结构难以学习,导致其翻译准确率低,进而影响整体翻译质量.为此,本文提出基于多粒度知识对齐的低资源机器翻译方法.首先,利用词对齐模型提取词级、实体级的对应翻译对,并基于此对齐关系设计词对齐、实体对齐和思考过程词对齐三种规则式对齐奖励;其次,结合格式和整体语义的一致性形成奖励信号;最后,通过强化学习框架,将这些奖励信号整合到翻译模型的训练过程中,使模型能够专注于源文本中复合实体语义关键信息的准确翻译.在中英、中越、中老三个语言对的实验表明本文方法在BLEU、COMET和实体准确率指标上均优于基线模型.详细的对比实验证明多粒度知识对齐方法能更好地理解复合实体的语义信息,从而有效缓解复合实体翻译中的歧义问题,提升翻译质量.

Abstract

In low-resource languages,the deep semantic structures of compound entities are difficult to learn,leading to low translation accuracy and consequently affecting overall translation quality.To address this issue,this paper proposes a low-resource machine translation method based on multi-granularity knowledge alignment.First,a word alignment model is used to extract corresponding translation pairs at the word level and entity level,based on which three rule-based alignment rewards—word alignment,entity alignment,and thought process word alignment—are designed.Second,consistency in format and overall semantics is incorporated to form reward signals.Finally,these reward signals are integrated into the training process of the translation model through a reinforcement learning framework,enabling the model to focus on accurately translating semantically critical information of compound entities in the source text.Experiments on three language pairs,Chinese-English,Chinese-Vietnamese,and Chinese-Lao,demonstrate that the proposed method outperforms baseline models in terms of BLEU,COMET,and entity accuracy metrics.Detailed comparative experiments show that the multi-granularity knowledge alignment method can better understand the semantic information of compound entities,thereby effectively alleviating ambiguity issues in compound entity translation and improving translation quality.

关键词

低资源机器翻译 / 多粒度知识对齐 / 大语言模型 / 强化学习

Key words

low-resource machine translation / multi-granularity knowledge alignment / large language models / reinforcement learning

引用本文

引用格式 ▾

王晓聪,李英,李衍铎,高盛祥,余正涛. 基于多粒度知识对齐的低资源机器翻译方法[J]. 昆明理工大学学报（自然科学版）, 2026, 51(3): 96-108 DOI:10.16112/j.cnki.53-1223/n.202511040002

登录浏览全文

4963

注册一个新账户忘记密码

参考文献

原文顺序 | 出版日期 | 本文引用

[1]	ZHAO W X, ZHOU K, LI J Y, et al. A survey of large language models[J]. Frontiers of Computer Science, 2026, 20(12):1-40.

[2]	WANG J A, MENG F D, ZHANG Y X, et al. Retrieval-augmented machine translation with unstructured knowledge[C]// Findings of the Association for Computational Linguistics: EMNLP 2025,Suzhou,China.Stroudsburg,PA, USA:ACL,2025:5858-5871.

[3]	MOSLEM Y, ROMANI G, MOLAEI M, et al. Domain terminology integration into machine translation:Leveraging large language models[C]// Proceedings of the Eighth Conference on Machine Translation,Singapore.Stroudsburg,PA, USA:ACL,2023:902-911.

[4]	BOGOYCHEV N, CHEN P Z. Terminology-aware translation with constrained decoding and large language model prompting[C]// Proceedings of the Eighth Conference on Machine Translation,Singapore.Stroudsburg,PA, USA:ACL,2023:890-896.

[5]	KIM S, SUNG M, LEE J, et al. Efficient terminology integration for LLM-based translation in specialized domains[C]// Proceedings of the Ninth Conference on Machine Translation,Miami,Florida,USA.Stroudsburg,PA, USA:ACL,2024:636-642.

[6]	HUANG Y C, LI B H, FENG X C, et al. Aligning translation-specific understanding to general understanding in large language models[C]// Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami,Florida,USA.Stroudsburg,PA, USA:ACL,2024:5028-5041.

[7]	ZHENG J W, HONG H H, LIU F Y, et al. Fine-tuning large language models for domain-specific machine translation[EB/OL]. Arxiv, 2024: 2402.15061.https://arxiv.org/abs/2402.15061.

[8]	ZHANG M S, LI Z H, FU G H, et al. Syntax-enhanced neural machine translation with syntax-aware word representations[C]// Proceedings of the 2019 Conference of the North,Minneapolis,Minnesota.Stroudsburg,PA, USA:ACL,2019:1151-1161.

[9]	OCH F J, NEY H. A systematic comparison of various statistical alignment models[J]. Computational Linguistics, 2003, 29(1):19-51.

[10]	ARTETXE M, SCHWENK H. Margin-based parallel corpus mining with multilingual sentence embeddings[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics,Florence,Italy.Stroudsburg,PA, USA:ACL,2019:3197-3203.

[11]

HASLER E, GISPERT A, IGLESIAS G, et al. Neural machine translation decoding with terminology constraints[C]// Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human LanguageTechnologies,Volume 2 (short Papers), New Orleans,Louisiana.Stroudsburg,PA,USA:ACL,2018:506-512.

[12]	LI Z, ZHENG M, SONG M Y, et al. TAT-R1:Terminology-aware translation with reinforcement learning and word alignment[EB/OL]. Arxiv, 2025: 2505.21172.https://arxiv.org/abs/2505.21172.

[13]	GUO D Y, YANG D J, ZHANG H W, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning[J]. Nature, 2025, 645(8081):633-638.

[14]	HE M G, LIU Y L, TAO S M, et al. R1-T1:Fully incentivizing translation capability in LLMs via reasoning learning[EB/OL]. Arxiv, 2025: 2502.19735.https://arxiv.org/abs/2502.19735.

[15]	REI R, STEWART C, FARINHA A C, et al. COMET:A neural framework for MT evaluation[C]// Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP),Online.Stroudsburg,PA, USA:ACL,2020:2685-2702.

[16]	FENG Z P, CAO S S, REN J H, et al. MT-R1-zero:Advancing LLM-based machine translation via R1-zero-like reinforcement learning[C]// Findings of the Association for Computational Linguistics:EMNLP 2025, Suzhou,China.Stroudsburg,PA,USA:ACL,2025:18685-18702.

[17]	WANG J A, MENG F D, ZHOU J. DeepTrans: Deep reasoning translation via reinforcement learning[EB/OL]. Arxiv, 2025: 2504.10187.https://arxiv.org/abs/2504.10187.

[18]	CUI G F, WANG P C, LIU Y, et al. CRPO:Confidence-reward driven preference optimization for machine translation[C]//Findings of the Association for Computational Linguistics: ACL 2025,Vienna,Austria.Stroudsburg,PA, USA:ACL,2025:560-574.

[19]

DONG T Y, LI B, LIU J S, et al. MLAS-LoRA:Language-aware parameters detection and LoRA-based knowledge transfer for multilingual machine translation[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (volume 1: Long Papers),Vienna,Austria.Stroudsburg,PA, USA:ACL,2025:15645-15660.

[20]	NGUEFACK I N, FINKELSTEIN M, SAKAYO T S. Pretraining strategies using monolingual and parallel data for low-resource machine translation[C]// Proceedings of the Sixth Workshop on African Natural Language Processing (AfricaNLP 2025),Vienna,Austria.Stroudsburg,PA, USA:ACL,2025:31-38.

[21]	GHANIZADEH M A, DOUSTI M J. Dynamic jointly batch selection for data efficient machine translation fine-tuning[EB/OL]. Arxiv, 2025: 2511.04406.https://arxiv.org/abs/2511.04406.

[22]	LIU A X, FENG B, XUE B, et al. DeepSeek-V3 technical report[EB/OL]. arXiv, 2024: 2412.19437.https://arxiv.org/abs/2412.19437.

[23]	SHAO Z H, WANG P Y, ZHU Q H, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models[EB/OL]. Arxiv, 2024: 2402.03300.https://arxiv.org/abs/2402.03300.

[24]	WANG J A, MENG F D, ZHOU J. ExTrans: Multilingual deep reasoning translation via exemplar-enhanced reinforcement learning[EB/OL]. Arxiv, 2025: 2505.12996.https://arxiv.org/abs/2505.12996.

[25]	SCHULMAN J, WOLSKI F, DHARIWAL P, et al. Proximal policy optimization algorithms[EB/OL]. Arxiv, 2017: 1707.06347.https://arxiv.org/abs/1707.06347.

[26]	CONNEA U A, LAMPLE G, RANZATO M A, et al. Word translation without parallel data[J]. Arxiv,2017:1710.04087.

[27]	ZHUANG X H, GAO S X, YU Z T, et al. Low resource neural machine translation model optimization based on semantic confidence weighted alignment[J]. International Journal of Machine Learning and Cybernetics, 2024, 15(10):4325-4340.

[28]

SCHWENK H, WENZEK G, EDUNOV S, et al. CCMatrix:Mining billions of high-quality parallel sentences on the web[C]// Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (volume 1: Long Papers),Online.Stroudsburg,PA, USA:ACL,2021:6490-6500.

[29]	POST M. A call for clarity in reporting BLEU scores[C]// Proceedings of the Third Conference on Machine Translation: Research Papers,Belgium,Brussels.Stroudsburg,PA, USA:ACL,2018:186-191.

[30]	VERL Team. VERL: Volcano engine reinforcement learning framework[EB/OL].[2026-03-25].https://github.com/verl-project/verl. 2024.

[31]	JALILI SABET M, DUFTER P, YVON F, et al. SimAlign:high quality word alignments without parallel training data using static and contextualized embeddings[C]//Findings of the Association for Computational Linguistics: EMNLP 2020,Online.Stroudsburg,PA, USA:ACL,2020:1627-1643.