Effective Pivot Language Strategies for Low-Resource Translation via Prompting Large Language Models

Traditional fine-tuning methods for low-resource machine translation, while effective in enhancing translation quality, are often constrained by substantial computational and data requirements, limiting their practical utility. Prompt tuning large language models (LLMs) has recently emerged as a lightweight alternative, activating latent knowledge through carefully crafted prompts without altering model parameters. However, existing prompting strategies often struggle with semantic misalignment, particularly when translating between distant low-resource language pairs. In this paper, we investigate the efficacy of incorporating a pivot language strategy to address this “understanding misalignment”. Building upon the DUAT-I prompting framework, we propose two distinct pivot-based approaches using English as the interlingua: (1) a sentence-level two-stage translation pathway and (2) a wordlevel pathway focusing on granular lexical mapping. Extensive experiments on six low-resource language pairs from the FLORES-200 benchmark, utilizing prominent LLMs and evaluated with COMET, demonstrate that our pivot strategies, especially the sentence-level approach, significantly outperform direct translation and few-shot prompting baselines.