Ontology-Guided Evaluation of LLM-Generated Content: A Neurosymbolic Framework for Knowledge Engineering Education

Evaluating LLM-generated content in expert-knowledge domains is challenging. Reference-based metrics miss semantic correctness, and LLM-as-a-judge approaches inherit the opacity and hallucination risks they were meant to detect. We propose a neurosymbolic framework in which an ontology serves as the primary control layer over neural evaluation. It formalizes expert assessment criteria as ontological concepts, grounds neural predictor initialization from this formalization, and reincorporates predictions into the ontology for expert-controlled score interpretation. We instantiate this framework in a Knowledge Engineering education case study: the automated evaluation of LLM-generated explanations of ontology modeling mistakes. Experiments on real-world student data show consistent improvements over pre-trained baselines, particularly on under-represented scoring classes. Beyond evaluation, the ontology-formalized scoring levels provide a natural reward signal for generation control via reinforcement learning, opening a path toward ontology-guided LLM generation.