The evaluation of text complexity is an important topic in education. While this objective has been addressed by approaches using lexical and syntactic analysis for decades, semantic complexity is less common, and the recent research works that tackle this question rely on machine learning algorithms that are hardly explainable and are not specifically designed to measure this variable. To address this issue, we explore in this paper the engineering of novel features to evaluate conceptual complexity. Through the construction of a knowledge graph that captures the concepts present in a text and their generalized forms, we measure different graph-based metrics to express such a complexity. Eventually , early-stage evaluations based on a well-known public corpus of students' productions show that the use of these metrics significantly improves performance compared to a state-of-the-art binary neural network classifier.