TULUN: Transparent and Adaptable Low-resource Machine Translation
This addresses the challenge of domain adaptation for low-resource MT, making it more accessible to non-technical users, though it is incremental as it builds on existing MT and LLM methods.
The paper tackles the problem of machine translation for low-resource languages in specialized domains by proposing Tulun, a system combining neural MT with LLM-based post-editing using glossaries, which achieved improvements of 16.90-22.41 ChrF++ points over baselines on medical and disaster relief tasks and 2.8 ChrF points over NLLB-54B on FLORES.
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.