Testing the extrapolation quality of word frequency models