Меню
Vol. 1 No. 30 (2026) / 2026-03-30 / Number of views: 140
Authors
Keywords
Link to DOI:
How to quote
In the context of a shortage of digital resources for the Kazakh language, the formalization of subject knowledge from the school history course of Kazakhstan is of particular relevance to support intelligent programmable systems. This paper presents the results of a study of intelligent programmable systems on the history of Kazakhstan, built on the basis of a full corpus of textbooks for grades 5-10, and analyzes the impact of the obtained knowledge graph on the quality of context search for QA tasks.
The purpose of the study is to develop an algorithm for term typing (person, period, place, source, event, unknown). To obtain initial "grain" labels, high-precision templates are used, after which training is carried out on XLM-R transformer embeddings using logistic regression. On validation by seed-labeling, accuracy of the order of 0.89 and F1 up to 0.92 for the main classes (person, period, place) are achieved.
The scientific novelty of the study lies in the development of a method for assessing the practical usefulness of an ontology formed from a QA set of 661 examples (528 train /133 test) with guaranteed correct answer spans. The proposed method compares several semantic search configurations: basic search across textbook sections, search across ontological "documents," and variants that expand queries and contexts using a thesaurus and ontology. The results show that using ontology documents increases P@1 from approximately 0.56 to 0.86 and nDCG@10 from ≈0.77 to ≈0.92; the combined ontology + expansion scheme maintains a high level of nDCG and almost complete coverage of relevant contexts.
Overall, the proposed ontology-oriented thesaurus and knowledge graph for the History of Kazakhstan not only structure the educational content for grades 5–10 but also demonstrate a substantial improvement in context retrieval quality for Kazakh-language QA systems, forming a foundation for future research on training and evaluating domain-specialized QA models.