Probability-Deviation Guided Spelling Error Correction for Low-Resource Languages

Spelling Error Correction (SEC) is a fundamental task in natural language processing, essential for ensuring textual quality across applications from writing assistants to professional publishing. While deep learning approaches have advanced significantly for high-resource languages, low-resource languages face substantial challenges due to a scarcity of annotated data. The recent emergence of large language models (LLMs) pre-trained on vast multilingual corpora offers promising potential for few-shot and zero-shot learning in low-resource SEC. However, their direct application is hindered by issues such as hallucination and overcorrection. To bridge this gap, we propose a unified framework designed to effectively harness LLMs for robust SEC in low-resource settings. Our approach is built on two pillars: reasoning optimization and data enhancement. We introduce a novel pipeline that integrates error detection—leveraging token-level probability deviations within the LLM—with generative correction, using a selective adoption strategy that applies corrections only where both modules agree. Furthermore, we construct reproducible data augmentation pipelines to generate large-scale, task-specific pretraining data for Indonesian and Vietnamese, mitigating data scarcity. Evaluated on the International Conference on Multilingual Intelligent Information Processing (IMLIP 2025) shared task benchmarks, our method achieves state-of-the-art detection performance (85.0% F1 for Vietnamese, 77.8% for Indonesian) and significantly improves correction precision (from 50% to over 86%), demonstrating the framework’s effectiveness and generalizability for low-resource language SEC.