TULUN: Transparent and Adaptable Low-resource Machine Translation

2025-05-24

AI Summary

Low-resource languages exhibit suboptimal performance in domain-specific translation, and existing domain adaptation methods rely on model fine-tuning—rendering them inaccessible to non-technical users and small organizations. This paper proposes a zero-fine-tuning, terminology-editable, human-in-the-loop transparent adaptation framework that synergistically integrates neural machine translation (NMT) with large language model (LLM)-driven, terminology-aware post-editing. An open-source web platform enables collaborative construction and reuse of terminology databases and translation memory repositories. The approach eliminates the need for model retraining, significantly enhancing domain controllability and usability. On medical and disaster-response translation tasks for Tetum and Bislama, it achieves ChrF++ gains of 16.90–22.41 points; on the FLORES benchmark across six low-resource languages, it outperforms NLLB-54B by an average of 2.8 ChrF points.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Application Domains: Humanities & Computational Social Science

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

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.

Problem

• Enhancing low-resource machine translation for specialized domains
• Eliminating fine-tuning needs for non-technical users and small organizations
• Improving translation accuracy using terminology-aware human-machine collaboration

Innovation

• Combines neural MT with LLM-based post-editing
• Uses glossaries and translation memories for guidance
• Open-source web platform for collaborative terminology management