Aspect-Based Semantic Textual Similarity for Educational Test Items

In the educational domain, identifying the similarity among test items provides various advantages for exam quality management and personalized student learning. Existing studies mostly relied on student performance data, such as the number of correct or incorrect answers, to measure item similarity. However, nuanced semantic information within the test items has been overlooked, possibly due to the lack of similarity-labeled data. Human-annotated educational data demands high-cost expertise, and items comprising multiple aspects, such as questions and choices, require detailed criteria. In this paper, we introduce a task of aspect-based semantic textual similarity for educational test items (aSTS-EI), where we assess the similarity by specific aspects within test items and present an LLM-guided benchmark dataset. We report the baseline performance by extending the STS methods, setting the groundwork for future aSTS-EI tasks. In addition, to assist data-scarce settings, we propose a progressive augmentation (ProAug) method, which generates step-by-step item aspects via recursive prompting. Experimental results imply the efficacy of existing STS methods for a shorter aspect while underlining the necessity for specialized approaches in relatively longer aspects. Nonetheless, markedly improved results with ProAug highlight the assistance of our augmentation strategy to overcome data scarcity.