Achieving Professional Translation in the Military Field Through Fine-Tuning LLM

Despite the remarkable advancements in large language models (LLMs) that have significantly addressed natural language translation challenges in general-purpose domains, their application to specialized machine translation (MT) remains constrained by substantial technical barriers, particularly in high-stakes sectors demanding exceptional translational accuracy and contextual fidelity, such as military intelligence and biomedical sciences. In the military domain, rule-based machine translation (RBMT) demonstrates efficacy in achieving precise translation of specialized terminology at granular levels. However, sentence- and paragraph-level translations necessitate contextual comprehension for accurate semantic representation, rendering RBMT approaches inadequate. To address these limitations, we constructed a military translation dataset (Military-MT) incorporating translation samples across three granularity levels: terminology, sentence, and paragraph. Building upon this dataset, we propose an instruction-oriented efficient fine-tuning methodology. Evaluations conducted across three foundational model baselines reveal that our approach significantly enhances zero-shot machine translation (MT) capabilities of base models in military-specific translation tasks.