Abstract:

The evaluation process is an important but challenging part of educational systems. In this context, the biggest challenge, beyond designing high-quality assessment questions, is the alignment with predefined learning outcomes and educational standards, especially in low-resource languages. Alignment is not limited to learning standards; it also involves a clear connection among questions, answer options, and intended learning outcomes.

As part of the study, more than 6,000 date-specific questions were developed, each with a corresponding set of answers. The dataset included historical events with clearly stated dates and textual descriptions. In many cases, these descriptions included references to additional events. Empirical analysis showed that in 98% of cases, the first sentence of the description contains key information related to the mentioned date. To reduce semantic noise and ensure consistency in question generation, only the first sentence was used as the main source of information.

Questions were formed by converting existing sentences into interrogative form, following the grammatical rules of the Bosnian language. For each generated question, the correct answer was automatically derived from the associated historical event date. Distractors were generated through structured temporal offsets applied to the original historical date. Three difficulty levels (easy, medium, and advanced) were defined based on predefined temporal distance intervals. More difficult questions included answer options that were closer to the correct date, requiring more precise knowledge, whereas easier questions used larger time differences. This made it easier to control the level of difficulty without losing alignment with the predefined learning outcomes and standards.

All generated questions were manually reviewed to verify structural correctness, alignment with predefined learning outcomes, and appropriate difficulty classification. This review confirmed consistency between the questions, the corresponding answer sets, and the predefined learning standards. The generated questions will be integrated into an educational quiz platform developed in collaboration with the Centre for Peace and Multiethnic Reconciliation Mostar (CZM), intended for students and a wider audience interested in historical knowledge.

The results indicate that combining rule-based automation with expert validation can support more consistent alignment between curriculum standards and assessment practice.

Keywords:

Offline GenAI, learning outcomes verification, automated assessment generation, Bosnian language.