DOI: https://doi.org/10.61957/joerai-20250205

Title: Curriculum-Aligned Semantic Assessment: Improving Fairness and Accuracy in AI-Assisted Student Marking

Author: Mlalazi Sijabulisiwe

Abstract

As artificial intelligence becomes increasingly integrated into education, AI-assisted marking systems offer benefits such as reduced grading workloads and improved consistency. However, many existing models penalize students for deviating from a broader knowledge base rather than adhering to the prescribed curriculum, raising concerns of fairness. This paper presents a curriculum-aligned natural language processing (NLP) model designed to assess short and long-form student responses within the scope of a national syllabus. The system dynamically ingests curriculum documents, encodes key concepts into a hierarchical structure and semantically matches student answers using a fine-tuned BERT model. A simulated dataset of 1,400 primary science responses (including in-scope, out-of-scope and distractor answers) was used to evaluate the model. Results show a strong correlation with human marking (Pearsons r = 0.88) and improved precision and recall in identifying out-of-scope content (F1 = 0.90), outperforming a non-aligned baseline. These findings suggest that curriculum-aware assessment models can enhance fairness, uphold instructional integrity and support scalable, transparent evaluation in educational settings.

Keywords:Curriculum-aware assessment, Natural language processing in education, AI-assisted marking, Educational fairness, Syllabus-aligned evaluation, Explainable educational AI, Semantic scoring models.