Distributional Learning of Appearance
In our experiments, the appearance of a word is modelled using a set of images of examples of that word, which in the figure we illustrate as a pile of images. The identification task is to determine which of two unlabelled novel appearances should be paired with a word (in this case cherry). The correct answer is on the right. The naming task is to determine which of two labels (in this case cherry and bridge) should be paired with an appearance. The correct answer is cherry. In both types of task the system is provided with knowledge of the appearance of a set of words (in the figure six) disjoint from those involved in the task.