Abstract
This study investigates whether morphology-aware input processing can improve question-answering performance for teaching programming in Kazakh, a low-resource language with rich morphological variation. Rather than proposing a new neural architecture, the study presents a controlled empirical evaluation of an adapted pipeline that combines FEMSeg_kaz preprocessing with morphology-aware post-processing (MAPP) normalization and domain-specific fine-tuning. A dataset of 50,386 Kazakh-language Python programming question–answer pairs was constructed with GPT-4o assistance and expert validation. Each pair consists of a student-style programming question in Kazakh and a corresponding instructional answer, usually including a short explanation and, when appropriate, a Python code example. Experiments with multilingual encoders and a controlled MiniLM setup show that morphology-aware normalization produces measurable improvements under fixed evaluation conditions, while its combination with domain-specific fine-tuning improves several answer selection and contextual-overlap indicators. The findings suggest that carefully controlled adaptation of established NLP techniques can support the development of Kazakh educational QA systems in morphologically rich, low-resource settings. Because no separate validation split or early stopping was used, these findings should be interpreted as controlled fixed-epoch comparison results rather than as fully optimized model selection results.