This paper considers the problem of automatic answering complex scientometric questions formulated in natural language (NL) over knowledge bases. The study is topical due to the limitations of modern large language models (LLMs): despite high understanding capabilities, they tend to generate inaccurate responses to user questions and may have outdated information in specialized subject areas. At the same time, knowledge graphs provide accurate and relevant information but require knowledge of a formal query language. A hybrid architecture-based solution is proposed: an LLM acts as an intelligent interface to an ontology-driven knowledge base, converting NL questions into correct SPARQL queries, and the results are returned to the user. To solve the problem, a specialized data corpus is compiled to train and test NL-to-SPARQL models in the field of control theory. The approach is implemented based on the ontology of scientific activity in the field of control theory and validated on the generated corpus of questions. Integration of the LLM with the ontology-driven knowledge base ensures a high accuracy of answers (about 99%), which confirms the prospects of this approach.