Semantic applications require expressive schemas, or ontologies, to exploit the full potential of knowledge graphs, which have become ubiquitous to express and organize knowledge in many domains. Existing knowledge graphs are rich in factual information, but the available schemas, handcrafted by knowledge engineers, are often minimal and not always complied with by the contributors of factual data. A key challenge to bridge this gap is to develop methods to induce schema axioms from factual data. This task can be broken down to two steps: (i) generating candidate axioms and (ii) scoring them against the available evidence for acceptability. Here, we focus on the latter, crucial step. Current methods such as OWL2Vec*, Onto2Vec, and OPA2Vec have shown promise, but accurately predicting the acceptability of axioms not logically deducible from an ontology remains a challenge. This paper introduces OCASP, OWL Class Axiom Score Predictor, a novel active learning approach to address this issue. The approach exploits a semantic Web reasoner, a data-driven axiom scoring heuristic, and an existing semantic similarity, which we extend to include complex class axioms, this defining an axiom-based embedding space. Our approach aims at providing a more efficient and robust solution for expressive schema induction, overcoming the limitations of existing methods.