Models of similarity have traditionally assumed that the similarity relation is symmetrical. However, when reversed, similarity statements frequently have different properties from those of the original. Previous attempts to account for the asymmetry of similarity, have focused only on literal comparisons, resulting in a tendency to underestimate the degree of asymmetry in non-literal comparisons (i.e., similes). A model of similarity is proposed to account for the varying degrees of asymmetry found in all kinds of natural language comparisons. In this model, asymmetry is attributed to an imbalance in the salience of the shared attributes. Studies are reported which test key aspects of the model. The results appear to provide converging evi...