NJTSC: Advancing Low-Resource Spelling Correction through Neighboring-Language Joint Training and In-Context Learning

Spelling correction plays a crucial role in text preprocessing and language digitalization for lowresource languages, serving as a fundamental task for improving text quality and promoting digital inclusion. Addressing spelling errors in Southeast Asian languages such as Indonesian and Vietnamese is particularly challenging due to limited corpora, high annotation costs, and unique linguistic characteristics. This paper addresses the IMLIP 2025 Shared Task by proposing a Neighboring Languages Joint Training Framework for Spelling Correction (NJTSC). Our team, Jiujiuwo, introduces a neighboring languages joint training strategy that exploits typological similarities through large language model (LLM)-based translation and direct data incorporation from similar languages. Furthermore, we employ in-context learning (ICL) with structured prompt templates to enrich model inputs and utilize Low-Rank Adaptation (LoRA) for efficient fine-tuning. Experimental results show that NJTSC achieves competitive performance with fewer than 3 billion parameters, demonstrating the effectiveness of our approach for low-resource spelling correction.