Semantic demand analysis for academic library acquisition decisions

Academic library acquisitions based on historical circulation systematically fail to anticipate demand in rapidly evolving domains. This study presents a semantic demand analysis framework that integrates large-language-model-derived text embeddings with borrowing-behavior data to reduce supply–demand misalignment, where large language models serve as instruments for embedding, metadata completion, and quality assessment, while the core contribution is the density-based quantification of latent demand validated against independent evidence. Using 195,823 circulation records and metadata for 52,187 titles from a Chinese university library (2020–2025), the authors construct semantic vectors, dynamic reader profiles, a KDE-based imbalance index in high-dimensional embedding space, and a composite scoring model. Benchmarked against collaborative filtering, matrix factorization, and linear discriminant analysis baselines, the top 500 recommendations reduce the zero-circulation rate from 28.4% to 4.2% and cut the Gini coefficient from .65 to .42. Eighteen supply-deficient domains are identified, triangulated with zero-result online public access catalog queries. Open-source embedding replication confirms reproducibility without proprietary dependence.