Bridging the Multilingual Gap in Educational Question Generation via Knowledge Distillation
Abstract
Question Generation (QG) is central to Intelligent Tutoring Systems, but routine Large Language Model (LLM) inference is difficult to deploy in resource-constrained educational settings, including much of the Global South. We study whether sequence-level knowledge distillation can transfer multilingual, topic-controlled QG from a teacher LLM into a deployable Small Language Model (SLM) without commissioning per-language datasets. Using XQuAD contexts in 11 languages, Google Gemini 2.5 Flash generates synthetic training pairs and an mT5-small student is fine-tuned on the result. Against zero-shot LLMs and a bespoke augmented mT5 baseline, the distilled student retains 86--87\% of the teacher's lexical and WikiSemRel topic-control performance while using a far smaller model. The results suggest that distillation is a practical pathway for multilingual educational NLP under limited data and compute.