Semantic Clustering of Research Papers Using SBERT and K-Means to Support the Discovery of Potential Academic Collaborations
Abstract
Identifying potential research collaborations is critical for fostering interdisciplinary innovation and academic productivity. Traditional approaches, such as co-authorship network analysis, are limited in scope because they rely solely on past collaborations and fail to capture latent thematic connections among researchers. This paper proposes an integrated analytical framework that applies semantic content analysis to research publications to support the discovery and exploration of potential academic collaborations. We compare SBERT-based K-Means clustering with a traditional TF-IDF K-Means baseline and observe superior performance in capturing semantic relationships, as evidenced by higher Silhouette Scores and lower Davies-Bouldin Index values. The framework was evaluated using a text corpus comprising the titles, abstracts, and keywords of 1,335 articles authored or co-authored by researchers affiliated with a university over the past five years. Using the SBERT-based configuration, the approach identified twelve thematic clusters that correspond to interpretable research domains within the institution. Representative terms extracted for each cluster provide clear thematic boundaries and indicate the primary research focus of each domain. Furthermore, multi-dimensional analysis reveals that departmental affiliation patterns exhibit both expected disciplinary concentrations and cross-departmental research overlaps, while network visualization illustrates the interconnected research landscape and highlights authors who act as bridges between clusters.