Abstract
People systematically associate colors and concepts, a phenomenon that can either help or hinder the interpretation of color in information visualizations. For instance, by applying optimization algorithms on color-concept association ratings, one can create palettes that are easily interpretable without legends (Schloss et al, 2018). Yet such optimization requires the designer to quantify associations between each concept and a large range of colors, to avoid the conflicts that arise when multiple concepts evoke the same strongest associates. Collecting association ratings for all possible colors and concepts is prohibitively costly and time-consuming. We therefore considered whether the space of color-concept associations can be expressed using low-dimensional representations. If so, that would mitigate the need for exhaustive human ratings and enable extrapolation of a limited set of ratings to new concepts and colors. To test this possibility, we collected color-concept association ratings for 30 concepts (spanning diverse concrete and abstract conceptual domains) and 58 colors (sampled uniformly over CIELAB space). Using principal components analysis (PCA) on the mean color-concept association ratings, we determined 8 ‘color profiles’ that strongly captured the structure of color-concept associations (90% variance explained). From these profiles, we fit regression models to estimate association ratings between new, unobserved concepts and colors. These models predicted both how sensitive each profile was to hue, lightness, and chroma and what blend of the 8 profiles best captured the color associations for a given concept. Using a leave-one-out approach and querying a subset of colors, we strongly predicted ratings for held-out concepts (mean correlation of 0.82 between true and predicted ratings). Our method can be used to automatically generate easily-interpretable color palettes for visual communication. Moreover, our results indicate that the mental representations underlying color-concept associations are highly structured, opening the way for a more principled understanding of color semantics.