Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TULUN: Transparent and Adaptable Low-resource Machine Translation

Raphaël Merx, Hanna Suominen, Lois Hong, Nick Thieberger, Trevor Cohn, Ekaterina Vylomova

TL;DR

Tulun addresses domain-specific MT for low-resource languages by integrating neural MT with LLM-based post-editing guided by glossaries and translation memories, deployed through an open-source web platform. It avoids model fine-tuning, enabling non-technical users to tailor translations to domain terminology while maintaining transparency through glossary and memory matches. Empirical results show substantial improvements in Tetun medical and Bislama disaster relief tasks (ChrF++ gains of 16.90–22.41) and consistent gains on FLORES-200 across six languages (average +2.83 ChrF++ over NLLB-54B), accompanied by strong usability (SUS 81.25). These findings demonstrate a practical, adaptable framework for terminology-aware MT in low-resource settings, with future work focusing on prompt engineering and offline deployment to broaden accessibility.

Abstract

Machine translation (MT) systems that support low-resource languages often struggle on specialized domains. While researchers have proposed various techniques for domain adaptation, these approaches typically require model fine-tuning, making them impractical for non-technical users and small organizations. To address this gap, we propose Tulun, a versatile solution for terminology-aware translation, combining neural MT with large language model (LLM)-based post-editing guided by existing glossaries and translation memories. Our open-source web-based platform enables users to easily create, edit, and leverage terminology resources, fostering a collaborative human-machine translation process that respects and incorporates domain expertise while increasing MT accuracy. Evaluations show effectiveness in both real-world and benchmark scenarios: on medical and disaster relief translation tasks for Tetun and Bislama, our system achieves improvements of 16.90-22.41 ChrF++ points over baseline MT systems. Across six low-resource languages on the FLORES dataset, Tulun outperforms both standalone MT and LLM approaches, achieving an average improvement of 2.8 ChrF points over NLLB-54B.

TULUN: Transparent and Adaptable Low-resource Machine Translation

TL;DR

Tulun addresses domain-specific MT for low-resource languages by integrating neural MT with LLM-based post-editing guided by glossaries and translation memories, deployed through an open-source web platform. It avoids model fine-tuning, enabling non-technical users to tailor translations to domain terminology while maintaining transparency through glossary and memory matches. Empirical results show substantial improvements in Tetun medical and Bislama disaster relief tasks (ChrF++ gains of 16.90–22.41) and consistent gains on FLORES-200 across six languages (average +2.83 ChrF++ over NLLB-54B), accompanied by strong usability (SUS 81.25). These findings demonstrate a practical, adaptable framework for terminology-aware MT in low-resource settings, with future work focusing on prompt engineering and offline deployment to broaden accessibility.

Abstract

Machine translation (MT) systems that support low-resource languages often struggle on specialized domains. While researchers have proposed various techniques for domain adaptation, these approaches typically require model fine-tuning, making them impractical for non-technical users and small organizations. To address this gap, we propose Tulun, a versatile solution for terminology-aware translation, combining neural MT with large language model (LLM)-based post-editing guided by existing glossaries and translation memories. Our open-source web-based platform enables users to easily create, edit, and leverage terminology resources, fostering a collaborative human-machine translation process that respects and incorporates domain expertise while increasing MT accuracy. Evaluations show effectiveness in both real-world and benchmark scenarios: on medical and disaster relief translation tasks for Tetun and Bislama, our system achieves improvements of 16.90-22.41 ChrF++ points over baseline MT systems. Across six low-resource languages on the FLORES dataset, Tulun outperforms both standalone MT and LLM approaches, achieving an average improvement of 2.8 ChrF points over NLLB-54B.

Paper Structure

This paper contains 39 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Glossary and translation memory integration in MT
  4. LLMs and Automated Post-Editing (APE)
  5. Our contribution
  6. System Design & Implementation
  7. System Design
  8. Translation View
  9. Glossary and Translation Memory View
  10. System Configuration
  11. Evaluation Mode
  12. System Architecture
  13. Backend
  14. Glossary and Translation Memory Retrieval
  15. Prompt Design
  16. ...and 24 more sections

Figures (3)

  • Figure 1: System overview with example translation from English to Tetun (en-tdt). The depth .9system components and depth .9data are configurable by end-users.
  • Figure 2: Translation View with the MT text, post-edited text, and the glossary entries and past translations relevant to this translation.
  • Figure 3: Eval mode: users can browse evaluation results, and see the reference translation