Model Selection for Homophone Spelling Correction in Low-Resource Languages: A Systematic Review with Khmer Case Study

This study conducts a comprehensive review of existing literature on spelling correction models designed for low-resource languages (LRLs). The research examines how these models handle homophone errors, using Khmer as a case study to highlight challenges in LRLs. Out of 174 academic publications, 54 papers were chosen for detailed study. The review covers various spelling correction models, including traditional, deep learning, and large language models. Traditional methods are affordable to implement but struggle to understand word context properly. Deep learning models (DLMs) provide the best balance between cost and effectiveness for correcting Khmer homophones. Large language models (LLMs) offer the highest accuracy but require significant computational power and may favor languages with more available data. The study also provides a three-stage framework for selecting appropriate models. When data is limited, traditional methods work best for homophone correction. When moderate amounts of data are available, DLMs should be the preferred choice. When both sufficient data and computational resources are accessible, LLMs deliver optimal results. These recommendations offer valuable guidance for researchers who develop spelling correction systems for LRLs, particularly when addressing homophone-related challenges.