Vocabulary size by Illustration Type

[ReneZ]

Large differences in sample text size are certainly a problem but it can be addressed by subsampling.

You can split each section into chunks of ~1000 vords each (maybe rounded to full pages) to match text length to the smallest and 2nd smallest ones (Astro and Zodiac).

Yes, this is a valid approach statistically, but of course it weakens the statistics of the larger sections.
It also removes the meaning of the idea of a 'corpus'.