It looks like they're re-fitting the same model with different subsets of characters though -- a next step might be to use all the data and fit a regression model or some kind of hierarchical model with partial pooling which shares statistical strength across the different classes of character.