Individuals face an increasingly large number of social and technological risks. How these risks are perceived is of critical interest to researchers and policymakers alike. In recent years, researchers have demonstrated that high-dimensional word embeddings derived from text have the potential to improve the prediction of risk perception (Bhatia, 2019). A further alternative and potentially more powerful approach is free association: past research has found embeddings derived from free associations to be more predictive of human judgments and behavior than those derived from text. Analyses by Bhatia (2019) suggest that embeddings trained on free associations are better at predicting risk perception than embeddings trained on text and on par with the so-called psychometric approach to risk perception, which uses human jugdments on nine central dimensions of risk. We build on this work by collecting risk ratings and psychometric ratings for a larger set of risk sources (n=1004), and evaluate a collection of psychometric, text-based, and free association-based representations in terms of their ability to predict risk perception (all prediction is done at the aggregate, rather than the individual participant level). We find that despite being trained on orders of magnitudes less data, the free association-based embeddings are top performers in both the individual and ensemble representation comparisons, with an ensemble of the psychometric and free assations explaining over 90% of the risk perception variance. Finally, we investigate how our model can be applied to improve the communication of risk information between researchers, policymakers, and the general public.