Abstract
Enriching educational ontologies automatically from low-resource-language text remains an unsolved integration problem: conventional large language model (LLM)-to-knowledge-graph pipelines require a vocabulary-alignment step and lack hallucination control at the commit boundary. This study removes both bottlenecks and tests whether the resulting ontology can drive measurable learning gains. The core idea is schema co-design: the JSON schema constraining LLM output is the OWL T-box of the target ontology, so extracted records are directly populatable and no alignment step is needed; a literal-presence filter rejects entities absent from the source text before commit, the HermiT reasoner verifies consistency, and Kazakh, Russian and English labels are generated in a single follow-up call. The pipeline (Python 3.9, OWLready2) was benchmarked on 90 Kazakh paragraphs across GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro, and the resulting adaptive textbook was evaluated quasi-experimentally with 65 Grade 6 students (control n = 32, experimental n = 33) in two Astana schools over 16 weeks. Of 312 source chunks, 271 (86.9%) survived the full validation chain, yielding 1,847 OWL individuals, 2,931 object-property assertions and 6,512 data-property annotations across 47 classes; the literal-presence filter intercepted 5.2% of validated responses and the reasoner a further 1.1%. Extraction F1 reached 0.89 (GPT-4o), 0.87 (Claude) and 0.86 (Gemini), with hallucination rates of 4.9–8.3%; the one-time corpus build cost USD 3.47–9.64, with zero marginal LLM cost per learner. The experimental group outperformed controls on task accuracy (+19.4 points, d = 2.30), repeated errors (−53.6%, d = 1.85) and sessions to mastery (−34.3%, d = 1.50), all p < 0.001 under Bonferroni correction. The novelty lies in making the ontology T-box itself the extraction schema, combined with pre-commit literal grounding, validated in a real low-resource classroom deployment.