ZeUS: An Unified Training Framework for Constrained Neural Machine Translation

• Published in: IEEE access Vol. 12; pp. 124695 - 124704
• Main Author: Yang, Murun
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; Constrained neural machine translation; Constraints; Data models; Decoding; large language model; Large language models; Machine translation; Neural machine translation; Neural networks; Synthetic data; Terminology; Training; Translation; unified training framework
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3454510

Unlike general translation, constrained translation necessitates the proper use of predefined restrictions, such as specific terminologies and entities, during the translation process. However, current neural machine translation (NMT) models exhibit proficient performance solely in the domains of general translation or constrained translation. In this work, the author introduces the zero-shot unified constrained translation training framework, which adopts a novel approach of transforming constraints into textual explanations, thereby harmonizing the tasks of constrained translation with general translation. Furthermore, the author discovers the pivotal role of constructing synthetic data for domain-specific constrained translation in enhancing the model's performance on constrained translation tasks. To this end, the author utilizes large language models (LLMs) to generate domain-specific synthetic data for constrained translation. Experiments across four datasets and four translation directions, incorporating both general and constrained translations, demonstrate that models trained with the proposed framework and synthetic data achieve superior translation quality and constraint satisfaction rates, surpassing several baseline models in both general and contrained translation. Notably, ZeUS also exhibits significant advantages over multitask learning in constrained translation, with an average improvement of 7.25 percentage points in translation satisfaction rate (TSR) and 8.50 percentage points in translation completeness (TC).