DescriptionRecent research has used crowd sourced corpora of language to learn grounded meanings that associate color descriptions with uncertain regions in hue-saturation-value color space. In this paper, we explore the degree to which the interpretation of syntactically- complex color terms can be predicted compositionally from their constituents. Using both Elastic Net Regressors and Random Forest Regressors, we build models to predict the composed colors present in both Lux and in the tail data that was unused during the learning of Lux. We evaluate the performance of the models by assessing the learned parameters against the Lux parameters. We additionally look at novel human-generated descriptions and build a system that names colors productively.