A Transformer-Based Embedding Approach to Developing Short-Form Psychological Measures

Se-jin JungJang-Won Seo^*

• Department of Psychology, Jeonbuk National University, Jeonju, Republic of Korea

Developing short-form psychological measures is essential not only for reducing respondent burden but also for saving time and economic resources. Existing approaches to short-form development typically require full-scale administration and rely on factor analysis or machine learning techniques based on response data. In contrast, the present study proposes a novel, data-independent method for item reduction using transformer-based semantic embeddings. Items from the International Personality Item Pool 50-item Big-Five Factor Markers (IPIP-50) were embedded using the sentence-t5-xxl model to generate dense semantic representations. These embeddings were clustered via K-means, and representative items were selected based on their proximity to cluster centroids. The resulting short-form, consisting of 30 items, preserved the original five-factor structure and demonstrated strong psychometric properties. When compared with existing item reduction techniques-namely, Classical Test Theory and a Genetic Algorithm-the proposed method achieved comparable levels of reliability, convergent validity, and predictive performance. These findings highlight the potential of transformer-based embedding approaches not only for efficient item reduction but also for informing item development. The results support the feasibility of a resource-efficient, linguistically grounded alternative to data-dependent reduction methods.