Graphical models consistently outperform n-gram models by a larger margin on sparse words than not-sparse words, and by a larger margin on polysemous words than not-polysemous words. One exception is the NB-R, which performs worse relative to Web1T-n-gram-R on polysemous words than non-polysemous words. For each graphical model representation, we show the difference in performance between that representation and Web1T-n-gram-R in parentheses. For each representation, differences in accuracy on polysemous and non-polysemous subsets were statistically significant at p < 0.01 using a two-tailed Fisher's exact test. Likewise for performance on sparse vs. non-sparse categories.