This study addresses the real-time performance limits of Korean-language search systems caused by morphological complexity and the cost of semantic processing. We propose a hybrid query transformation method that couples rule-based preprocessing with a Transformer-based postprocessor. The rule-based stage simplifies agglutinative input, and the Transformer refines user intent and semantic context. On a curated Korean query set, our approach attains 89.0% Precision@5 (95% CI: 87.2–90.7) with 95 ms average latency (95% CI: 92–98), about 21% faster than an NLP-only baseline. User surveys and expert interviews further confirm practical applicability. To strengthen reliability and scope transparency, we report five-fold cross-validation, noise-robustness tests (spacing errors, minor typos), and comparisons against open proxy baselines (e.g., BM25+ KoNLPy). These additions clarify the study’s focus on Korean while providing reproducible evidence of robustness, positioning the framework as deployment-ready for Korean and a solid basis for future multilingual extensions.
Kim et al. (Tue,) studied this question.