Abstract:

The clarity of teaching materials is an essential component of any learning process: it determines how well concepts are understood, how effectively reasoning is developed, and how valid the assessment outcomes are. In technical areas such as algorithms and data structures of Computer Science, even minimal ambiguity can completely change the interpretation of a problem or the choice of the appropriate solution method. Linguistic precision is therefore critical, as an unclear statement may lead to conceptual errors, obscure the task’s objectives, or invalidate comparisons between students.

We present an application designed to detect and correct grammatical problems, logical ambiguities, and specification gaps in multiple-choice questions.
The study was conducted in a core Computer Science course focused on algorithms and advanced data structures— a subject that forms one of the fundamental pillars of any computing degree. The course is part of the second-year curriculum of a Computer Science program offered by the National Distance-Learning University of Spain (UNED), where the clarity of instructional materials and assessment activities plays an even more critical role due to the online learning environment.

Our goal is threefold: to improve the quality of assessment items, to support teachers in designing exams and exercises, and to offer students unambiguous statements that promote deep and fair learning.

The use of artificial intelligence technologies enables generalizable detection and domain-specific improvements. The platform combines linguistic rules with AI-based language models to identify statements that admit more than one valid answer, contain syntactic errors, or rely on hidden assumptions. For instructors, it saves time during the review process and provides active assistance in writing pedagogically sound items; for students, it reduces frustration caused by poorly phrased questions and allows them to focus on conceptual reasoning. In addition, the application suggests clearer reformulations and provides concise justifications for each detected issue.

The system has been empirically evaluated on a curated set of multiple-choice questions containing different types of defects, including grammatical errors, logical ambiguities, and semantic inconsistencies. Several large language models were tested under different prompting strategies, including zero-shot and few-shot configurations. Corrections suggested by the models were manually assessed by instructors. The results show an overall error correction rate above 68%, with grammatical issues being successfully addressed in more than 80% of the cases. Semantic errors proved more challenging, although a relevant subset was also correctly identified and improved.

The contributions include:
(1) a catalog of recurrent ambiguity patterns in algorithmic questions;
(2) a methodological framework for the automatic evaluation of formative assessment items;
(3) an empirical evaluation and comparison of large language models and prompting strategies based on manually validated corrections; and
(4) a practical, open-access tool that enhances the quality, fairness, and effectiveness of the learning process through clearer questions and AI-assisted diagnostics.

Keywords:

Artificial intelligence in education, Assessment quality, Computer Science education, Algorithms and data structures, Natural language processing, Language models.