Background: Risk attitudes are widely recognized as domain-specific, as reflected in frameworks such as DOSPERT. Yet most domain taxonomies remain top-down and may lag behind lived experience. Recent advances in natural language processing (NLP) enable a bottom-up approach by extracting domain structure from how people describe risks in their own words.


Objectives: We aim to (1) derive contemporary risk domains from free-text descriptions of everyday risks and risk decisions, and (2) develop and validate a typology of risk preference by deriving a parsimonious set of multidimensional risk profiles (Frey et al., 2023) that complements variable-centered accounts.


Methods: We collected open-ended responses from 860 chinese adults describing personally salient risks and risk decisions. Texts were embedded using a state-of-the-art sentence model (e.g., bge-large-zh-v1.5). We applied UMAP for dimensionality reduction and HDBSCAN for density-based clustering to construct an interpretable semantic map of risk themes. Guided by the resulting domain structure, we adapted and administered a domain-specific risk preference measure. We then conducted model-based clustering (latent profile analysis and other clustering solutions) to identify a parsimonious set of risk profiles, and characterized them via cognitive patterns, decision tendencies, and sociodemographic correlates.


Results: We obtained an interpretable and updateable taxonomy of risk domains that includes both classic domains (e.g., health/safety, finance) and emergent themes (e.g., online fraud, novel technology risks). We also identified a small number of robust multidimensional risk profiles that capture systematic heterogeneity in risk preference across domains for most individuals, providing a scalable typological framework for risk research and applied risk governance.