Spelling Corrector for Turkish Product Search

Spell correction for e-commerce platforms presents unique challenges that are not adequately addressed by existing methods, which are primarily tailored for general purpose. These challenges include handling specialized terminology, brand names, and foreign terms frequently used in search queries. Traditional spell correction algorithms often fail to account for these complexities, leading to irrelevant search results and decreased user satisfaction. This study aims to develop and evaluate a novel spell correction algorithm specifically designed for the Turkish e-commerce context, with a focus on online food and grocery search queries. To enhance the spell correction process, we developed a two-module system consisting of a suggester and a ranker. The suggester module, leveraging TURNA, a pretrained transformer-based Turkish model, is designed to generate possible corrections by capturing the complex relationships within Turkish language data using spelling mistake weights derived from custom Turkish datasets. The ranker module then ranks these suggestions using various domain-specific features. Experimental results show that the proposed model successfully improves spell correction performance in the e-commerce domain, outperforming existing tools. However, the model's performance in general Turkish text correction is less effective, indicating areas for further refinement.