Bourdieu Vectors: A Framework for Quantifying Taste in Social Space

This paper introduces Bourdieu Vectors, a computational framework for estimating social, cultural, economic, and symbolic capital based on cultural practices. Each practice is represented as a multidimensional vector generated via prompt-based LLM querying. Validation demonstrates robust quantitative performance across multiple dimensions: (1) statistically significant group separation in a 1,000-item dataset (PERMANOVA pseudo-F =151.05, p = 0.001), (2) meaningful correlations with empirical GSS 1993 data (n = 1, 583; r up to +0.58; 69.8% significant correlations at p < 0.05; mean classification accuracy = 65.8%, ROC-AUC = 0.629), which remained robust to demographic confounders in multivariate regression (70% significant; β up to 1.02), and (3) high intra-model consistency across five LLMs at deterministic settings (temperature=0: cosine similarity 0.91–1.00; mean inter-model agreement 0.86), with moderate stability at stochastic settings (temperature=1: cosine similarity 0.84–0.95). The framework enables systematic analysis of taste-based social positioning and is publicly accessible at https://bourdieuvectors.com.