Grouping research proposals with funding agency requirements: A contextualized language model and constrained K-means clustering approach

Abstract

As the volume of research proposals increases and the interdisciplinary nature of research fields exacerbates the complexity of manual proposal grouping, the grouping of research proposals becomes increasingly vital in funding agencies' evaluation procedures. Although previous ontology- and word embedding-based approaches have made valuable contributions to the advancement of research proposal grouping, their practical utility has been limited due to a lack of consideration of funding agencies' requirements. This study proposes a systematic approach for grouping research proposals that aligns with three common requirements: size, cannot-link, and must-link constraints. The proposed approach utilizes KLUE-RoBERTa for proposal vectorization and constrained K-means clustering for proposal grouping with size constraints. We introduce a proposal pre-partitioning and a proposal vector centralization to simultaneously consider the cannot- and must-link constraints in grouping proposals. An empirical analysis of 3,665 proposals submitted to the National Research Foundation of Korea demonstrates the effectiveness and practicality of the proposed approach. Additionally, we conduct a comparative analysis of various combinations of methodological components to optimize this approach. The proposed approach is considered a valuable complementary tool for grouping proposals, enhancing the overall efficiency and effectiveness of the proposal evaluation system.