From Lexical to Semantic: A Systematic Review of Off-Topic Detection Evolution in Automated Assessment

Off-topic detection is a vital challenge in natural language processing, especially as online learning, automated assessments, and AI-powered content platforms continue to expand – it ensures that learner responses or user-generated content align with intended prompts or topics Its applications span far beyond essay scoring to include spoken response assessment, educational dialogue systems, business writing analysis, and even harmful content moderation on social platforms This research aims to develop and enhance an off-topic detection model using AraBERT embeddings for Arabic student responses, focusing on boosting automatic assessment accuracy and elevating the overall robustness of educational AI systems. The findings demonstrate that our model reliably identifies relevance in short-answer tasks, achieving precision and recall comparable to or better than existing methods, suggesting that future work integrating deeper semantic, structural, and discourse-level features may further enhance assessment capabilities. [1], [2], [3], [4] This review aims to survey a range of studies in this domain, identifying the methods employed, assessing their significance, exploring the limitations they face, and highlighting emerging directions in the evolution of off-topic detection techniques within automated assessment systems.