Abstract. The rapid development of digital educational technologies has created new demands for the automation and intelligent analysis of educational content. Across the world, and especially in countries undergoing rapid digitalization such as Uzbekistan, educational institutions are increasingly deploying e-learning platforms, digital textbook repositories, and automated assessment systems. These platforms generate and manage enormous volumes of exercises, questions, and instructional tasks that must be organized, classified, and evaluated in a principled manner. Without a rigorous approach to understanding the semantic content of exercises, such platforms cannot effectively adapt to the needs of individual learners, align with curriculum standards, or provide meaningful feedback to educators and policymakers. This article proposes a mathematical model for the semantic analysis of exercises used in automated learning systems. The model is built on three complementary components: vector-space representation of textual content, probabilistic dependency graphs that capture the grammatical and relational structure of exercise statements, and ontological mapping of domain-specific concepts drawn from official curriculum taxonomies. Together, these components allow the model to classify exercises by cognitive complexity, assess their semantic relevance to defined learning objectives, and evaluate their internal structural coherence. Experimental results obtained from a corpus of over four thousand Uzbek-language educational exercises demonstrate the effectiveness of the approach, yielding an average classification accuracy of 91.4% across STEM subject domains. The study makes a foundational contribution to the theoretical basis of intelligent tutoring systems and adaptive learning platforms in Uzbekistan.