Abstract We present semantic atlases for journals in the life sciences and social sciences by encoding over 30 million abstracts with domain‐adapted transformer models, aggregating paper embeddings into journal centroids, and applying UMAP for visualization. We validate the resulting 2D manifolds through strong correlations between positional proximity and external measures of topical and authorship overlap. Semantic axes reveal interpretable gradients along biologically meaningful dimensions and a latent quality axis that recovers prestige hierarchies. The approach also positions journals with elevated retraction rates peripherally, highlighting latent signals of publication integrity. We introduce effective radius as a measure of geometric properties; larger radii are associated with higher impact and citations. These findings demonstrate that transformer embeddings provide a continuous, citation‐independent framework for decoding journal roles, thematic specialization, prestige, and knowledge dispersion, offering new quantitative tools for the science of science and dynamic monitoring of scholarly ecosystems.
Yang et al. (Thu,) studied this question.