On the (in?)feasibility ontology-based educational question generation with an LLM

The idea was simple: replace the PhD student’s cumbersome NLP pipeline for the surface realisation step to generate the ontology-based educational questions, answers, and feedback with an LLM. We’d start with English and then expand to other languages relevant for both the Horizon Europe project “STARWARS” and the South African project KeBoNa that we’re part of and leading, respectively, being French, Italian, Arabic, and isiXhosa. For instance, given Disjointwith(Valve ConnectionPoint), an education question may be, e.g., “Is a valve a connection point?”, to which the answer is “no” and the explanation that it’s not so because they’re disjoint classes, or, when also availing of an automated reasoner, e.g., “True or false: valves and connection points are distinct network elements.” to which the answer is “true”, and the feedback may be “True, because they’re disjoint and valve and connection point both are an apparatus that is a node that is a network element”. There are different…