Problem

Students transferring between two-year and four-year colleges frequently encounter difficulties in having their previously earned credits fully recognized. This problem, known as course articulation, often leads to students losing valuable credits, extending their time to degree completion, and incurring unforeseen educational expenses. The current system for establishing course equivalencies between institutions is largely manual, making it an exceedingly time-consuming, resource-intensive, and inconsistent process that cannot realistically scale to the millions of potential articulations required across various educational systems. While some efforts have explored using artificial intelligence to assist this work, their adoption in practice has been limited, partly due to concerns from faculty and staff regarding algorithmic accuracy and trustworthiness.

Method

This study developed a stakeholder-informed AI system to enhance course articulation within the State University of New York (SUNY) system. The methodology began with conducting surveys among articulation staff and faculty to gauge their perceptions and adoption rates of recommendations generated by a baseline algorithm (SBERT). Feedback from these surveys highlighted that current models often relied on superficial lexical matching, rather than capturing true semantic content. Based on these insights, the researchers developed Shared Space Alignment (SSA), a supervised alignment technique. SSA, inspired by the Multilingual Pseudo-Supervised Refinement (MPSR) approach, learns orthogonal transformation matrices for each college to align course vector representations into a shared space, making course equivalency a nearest-neighbor search problem, while shifting optimization towards the destination institution's space. The approach utilizes course titles and catalog descriptions, which are vectorized using various generations of NLP models (Word2Vec/DescVec, SBERT, and OpenAI) and, in some configurations, augmented with student enrollment histories (Course2vec). Model performance was evaluated using recall@1 and recall@5 against existing SUNY articulation agreements, and by analyzing the dispersion of course embeddings within Classification of Instructional Programs (CIP) categories to ensure content-based grouping. To identify new articulation opportunities, a similarity threshold for course equivalency was determined using AUC-ROC analysis, based on existing agreements and pseudo-negative samples.

Results

Initial surveys revealed an average adoption rate of 61.23% for baseline algorithmic recommendations among faculty and staff, who expressed a strong need for more accurate, content-based matches. The advanced method, combining OpenAI and Course2vec models with SSA, achieved significantly higher performance, reaching a recall@1 of 0.764 and recall@5 of 0.928. This represents a 5.5-fold improvement in accuracy compared to previous state-of-the-art methods. SSA itself contributed substantially to this gain, increasing performance by an average of 121.017% across different NLP generations. Furthermore, SSA successfully reduced the dispersion of course embeddings within academic disciplines (CIP categories) across 45 out of 46 system-wide categories and for 91.76% of institution-CIP pairs. This qualitative analysis confirmed that SSA effectively mitigates the influence of superficial linguistic artifacts (e.g., generic terms or institutional boilerplate), leading to more accurate content-based clustering. Based on a defined similarity threshold, the improved model identified 2,787,526 additional potential articulation pairs, which is 17.759 times more than the currently existing agreements. Projecting these new opportunities with the observed 61.23% adoption rate, the study estimates an 11.87-fold increase, or 1,706,802, in valid credit mobility opportunities for students that would otherwise go unrecognized.

Implications

This study demonstrates that an AI system designed with significant stakeholder input can dramatically improve course articulation processes, thereby expanding student credit mobility and potentially transforming institutional decision-making. By surfacing 92% of articulation matches within the top five suggestions, the AI can streamline the workflow from manual review to an auditing process, where faculty committees verify algorithmically predicted equivalencies. To maximize these benefits, it is crucial to enhance student awareness through academic advising, which could also be augmented by AI systems to navigate complex transfer requirements. Beyond course articulation, this methodology holds promise for other higher education applications, such as Common Course Numbering and Credit for Prior Learning processes, helping institutions better recognize diverse educational pathways. However, the study notes limitations, including that adoption projections were based on a lower-stakes scenario with a less accurate baseline model, and SSA's dependency on existing articulation data, suggesting avenues for future research into adoption in real-world settings and reducing data dependency.