Learning Distributions of Meant Color

When a speaker says the name of a color, the color they picture is not necessarily the same as the listener imagines. Color is a grounded semantic task, but that grounding is not a single value as the most recent works on color-generation do. Rather proper understanding of color language requires the capacity to map a sequence of words is to a distribution in color space. To achieve this, we propose a novel GRU based model which when given a color description such as “light greenish blue” produces an estimated probability distribution across HSV color space. This model learns the compositional functioning of each term: “light”, “green”, “ish” and “blue” in the color name, and show together they determine the shape of the distribution. This allows the prediction of distributions for color description phrases never seen in training. Such capacity is crucial in allowing human computer interaction systems to interpret vague color statements, with particular use in recognising objects described by color.