Figure 3.
Results of Simulations 2 and 3. Ellipses delimit the area corresponding to 90% of vowel tokens for Gaussian categories (a) computed from men’s, women’s, and children’s production data in Hillenbrand et al. (1995), recovered in Simulation 2 by (b) the lexical-distributional model, (c) the infinite mixture model, and (d) the gradient descent algorithm, and recovered in Simulation 3 by (e) the lexical-distributional model with αL = 10, 000, (f) the lexical-distributional model with αL = 10, (g) the infinite mixture model, and (h) the gradient descent algorithm.