The relevance of this research is driven by the growing need to improve the efficiency of semantic information retrieval amid the rapid expansion of text data, particularly in low-resource languages such as Kazakh. The purpose of the research is to develop a justified approach for selecting and comparing text vectorization models used in intelligent search systems, considering the morphological and syntactic features of the Kazakh language, and to construct a mathematical model for computing semantic similarity in a multidimensional vector space. The methodology is based on the empirical testing of six models (TF-IDF, Word2Vec, FastText, GloVe, BERT, and KazBERT) on a corpus of 24,000 Kazakh texts. Vectorization was performed using CLS-tokens; morphological preprocessing employed the Kaznlp tool. Model effectiveness was assessed using precision, recall, and F1-score metrics. The results demonstrated that KazBERT, combined with morphological analysis, achieved the highest accuracy in handling variable word forms, outperforming multilingual BERT by 11–15% and TF-IDF by over 30%. FastText showed strong resilience to morphological variation but was less effective with syntactically complex queries. The scientific novelty lies in the development of a hybrid model for intelligent search adapted to the agglutinative nature of the Kazakh language, and in the introduction of a custom morpho-syntactic metric that increases sensitivity to grammatical features. The conclusions confirm that adapting vector models to account for grammar significantly enhances retrieval relevance.
Sadykova et al. (Sat,) studied this question.