Mapping Skill Demand and Occupational Interdependencies in Online Job Postings: A Sentence-BERT and Vector Space Approach

Abstract The accelerating pace of technological change, automation and artificial intelligence is profoundly restructuring global labour markets, shifting the focus of demand-side analyses from static occupational titles toward dynamic and multifaceted skill portfolios. Traditional macroeconomic monitoring instruments, such as employer surveys and administrative registries, frequently suffer from significant reporting latency, aggregate oversimplification and rigid taxonomic classifications that fail to capture emerging competencies in real time. This study investigates the structural interdependencies and statistical dispersion of demanded skills across nine major occupational groups by analyzing a massive repository of web-scraped job postings from leading commercial recruitment platforms in Poland (SOJO dataset, 2019–2023). Leveraging Natural Language Processing, raw vacancy descriptions were embedded using a deep contextual Sentence-BERT architecture and mapped into the European Skills, Competences, Qualifications and Occupations (ESCO) taxonomy via Facebook AI Similarity Search (FAISS) dense vector indexing at a strict similarity threshold. In addition, the distribution, concentration and dispersion of 74 medium-level skill concepts were assessed through box-plot parameters, interquartile ranges and outlier analyses. The empirical findings confirm that skill demand varies significantly across occupational categories in both type and frequency, while revealing strong cross-occupational penetration of digital literacy, operational coordination and transversal capabilities into traditionally manual occupations. These results provide actionable insights for technological forecasting, curriculum modernization and data-driven active labour market policies.