Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback

Petruta-Simina Coroian; Gabriele Kotsis; Ismail Khalil

doi:10.1007/978-981-96-4291-5_32

Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback

Petruta-Simina Coroian
, Gabriele Kotsis
, Ismail Khalil^*

^*Corresponding author for this work

Institute of Telecooperation

Research output: Chapter in Book/Report/Conference proceeding › Conference proceedings › peer-review

Abstract

This paper presents a study on improving the quality of neural machine translation (NMT) for the English-Romanian language pair using Reinforcement Learning from Human Feedback (RLHF) via Direct Preference Optimization (DPO). Despite advancements in NMT, challenges remain, particularly for low-resource languages and personalized translations. By incorporating human feedback, the proposed approach demonstrates improvements in translation accuracy and naturalness. Although traditional metrics, such as BLEU and chrF++, yielded slightly lower scores for the DPO-trained model, human assessments indicate that the DPO-trained model better aligns with human preferences, particularly in everyday conversational contexts.

Original language	English
Title of host publication	Information and Communication Technology
Subtitle of host publication	13th International Symposium, SOICT 2024, Danang, Vietnam, December 13–15, 2024, Proceedings, Part IV
Publisher	Springer Singapore
Pages	394-405
Number of pages	12
Edition	1
ISBN (Electronic)	978-981-96-4291-5
ISBN (Print)	978-981-96-4290-8
DOIs	https://doi.org/10.1007/978-981-96-4291-5_32
Publication status	Published - 26 Apr 2025

Publication series

Name	Communications in Computer and Information Science
Volume	2353

Fields of science

102013 Human-computer interaction
102002 Augmented reality
102006 Computer supported cooperative work (CSCW)
102027 Web engineering
202038 Telecommunications
102021 Pervasive computing
102015 Information systems
102025 Distributed systems
102 Computer Sciences

JKU Focus areas

Digital Transformation
Sustainable Development: Responsible Technologies and Management

Access to Document

10.1007/978-981-96-4291-5_32

Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback
Khalil, I. (Speaker)
14 Dec 2024
Activity: Talk or presentation › Poster presentation › science-to-science

Cite this

Coroian, P.-S., Kotsis, G., & Khalil, I. (2025). Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback. In Information and Communication Technology: 13th International Symposium, SOICT 2024, Danang, Vietnam, December 13–15, 2024, Proceedings, Part IV (1 ed., pp. 394-405). (Communications in Computer and Information Science; Vol. 2353). Springer Singapore. https://doi.org/10.1007/978-981-96-4291-5_32

Coroian, Petruta-Simina ; Kotsis, Gabriele ; Khalil, Ismail. / Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback. Information and Communication Technology: 13th International Symposium, SOICT 2024, Danang, Vietnam, December 13–15, 2024, Proceedings, Part IV. 1. ed. Springer Singapore, 2025. pp. 394-405 (Communications in Computer and Information Science).

@inproceedings{9502ff1b86d74dea8f5ac2147dfb0b9a,

title = "Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback",

abstract = "This paper presents a study on improving the quality of neural machine translation (NMT) for the English-Romanian language pair using Reinforcement Learning from Human Feedback (RLHF) via Direct Preference Optimization (DPO). Despite advancements in NMT, challenges remain, particularly for low-resource languages and personalized translations. By incorporating human feedback, the proposed approach demonstrates improvements in translation accuracy and naturalness. Although traditional metrics, such as BLEU and chrF++, yielded slightly lower scores for the DPO-trained model, human assessments indicate that the DPO-trained model better aligns with human preferences, particularly in everyday conversational contexts.",

author = "Petruta-Simina Coroian and Gabriele Kotsis and Ismail Khalil",

year = "2025",

month = apr,

day = "26",

doi = "10.1007/978-981-96-4291-5\_32",

language = "English",

isbn = "978-981-96-4290-8",

series = "Communications in Computer and Information Science",

publisher = "Springer Singapore",

pages = "394--405",

booktitle = "Information and Communication Technology",

edition = "1",

}

Coroian, P-S, Kotsis, G & Khalil, I 2025, Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback. in Information and Communication Technology: 13th International Symposium, SOICT 2024, Danang, Vietnam, December 13–15, 2024, Proceedings, Part IV. 1 edn, Communications in Computer and Information Science, vol. 2353, Springer Singapore, pp. 394-405. https://doi.org/10.1007/978-981-96-4291-5_32

Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback. / Coroian, Petruta-Simina; Kotsis, Gabriele ; Khalil, Ismail.
Information and Communication Technology: 13th International Symposium, SOICT 2024, Danang, Vietnam, December 13–15, 2024, Proceedings, Part IV. 1. ed. Springer Singapore, 2025. p. 394-405 (Communications in Computer and Information Science; Vol. 2353).

Research output: Chapter in Book/Report/Conference proceeding › Conference proceedings › peer-review

TY - GEN

T1 - Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback

AU - Coroian, Petruta-Simina

AU - Kotsis, Gabriele

AU - Khalil, Ismail

PY - 2025/4/26

Y1 - 2025/4/26

N2 - This paper presents a study on improving the quality of neural machine translation (NMT) for the English-Romanian language pair using Reinforcement Learning from Human Feedback (RLHF) via Direct Preference Optimization (DPO). Despite advancements in NMT, challenges remain, particularly for low-resource languages and personalized translations. By incorporating human feedback, the proposed approach demonstrates improvements in translation accuracy and naturalness. Although traditional metrics, such as BLEU and chrF++, yielded slightly lower scores for the DPO-trained model, human assessments indicate that the DPO-trained model better aligns with human preferences, particularly in everyday conversational contexts.

AB - This paper presents a study on improving the quality of neural machine translation (NMT) for the English-Romanian language pair using Reinforcement Learning from Human Feedback (RLHF) via Direct Preference Optimization (DPO). Despite advancements in NMT, challenges remain, particularly for low-resource languages and personalized translations. By incorporating human feedback, the proposed approach demonstrates improvements in translation accuracy and naturalness. Although traditional metrics, such as BLEU and chrF++, yielded slightly lower scores for the DPO-trained model, human assessments indicate that the DPO-trained model better aligns with human preferences, particularly in everyday conversational contexts.

U2 - 10.1007/978-981-96-4291-5_32

DO - 10.1007/978-981-96-4291-5_32

M3 - Conference proceedings

SN - 978-981-96-4290-8

T3 - Communications in Computer and Information Science

SP - 394

EP - 405

BT - Information and Communication Technology

PB - Springer Singapore

ER -

Coroian PS, Kotsis G , Khalil I. Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback. In Information and Communication Technology: 13th International Symposium, SOICT 2024, Danang, Vietnam, December 13–15, 2024, Proceedings, Part IV. 1 ed. Springer Singapore. 2025. p. 394-405. (Communications in Computer and Information Science). doi: 10.1007/978-981-96-4291-5_32

Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback

Abstract

Publication series

Fields of science

JKU Focus areas

Access to Document

Activities

Enhancing Neural Machine Translation with Direct Preference Optimization Using Human Feedback

Cite this