Towards a Computational Model of Human Word-Color Associations

Abstract

Computational models of multimodal associations can help us to better understand the ways different domains of human knowledge and experience interact with and supplement each other. Natural language and color are two domains of particular interest, as they are both integral to our experience of the world and are powerful communication devices in their own rights. However, current computational models of human word-color associations attempt to bring the color domain closer to the distributional semantic domain by treating color as a lexical entity like any other target word. My work aims to preserve the rich information contained in human beings’ experience of color by maintaining color as a perceptual experience tied to some underlying understanding of word meaning. I first establish a dataset of human color annotations for words that represent varying degrees of abstractness and emotional content. Then, I develop three computational models that are grounded in color data: a distributional semantic model, an image analysis model, and finally, a Bayesian representativeness model. I find that the Bayesian representativeness model is best able to discern meaningful structure in input color data, which allows it to most closely emulate humans' psychological color associations for words.