Writing a fake Voynich by rules: copies, habits and the link between words

[quimqu]

Following Koen's point about the binding, nesting and the singulions, I have been searching deeplier in the LSA results of Davis and Layfield.

First: if we try all the possible ways to order the bifolios and we keep the best one, we will always find a "best one" just by chance (like throwing 24 dices and you keep the best one, just because it had better luck than the others). The proof: my scribe’s text has no correct order, and the same method still finds a better one, with an even bigger gain than in the real text.

Second: For each quire, I compared a nested structure with a singulion structure. So in the singulion, each time you open a bifolio, you see the pages of the bifolio. So the singulions win only with LSA, and only because they put the two inner pages of each bifolium face to face. Other methods (plain TF-IDF, shared vocabulary, glyph trigrams, cross-entropy, and LDA topic models) don’t see this difference.

So why with LSA? Well, LSA does not compare words. First, it learns from the manuscript wivh words tend to appear near from each other and makes topics. The compares the topics (not words) between the pages. Those topics are not about content, but about style (fammilies of words).

Anyway, what I found consitently is that the four pages of a bifolio are detected as simmilar in all the methods tested. Even my scriba's text has this behaviour. So each bifolio is a writing session. The singulions result is something that happens only with LSA: what is LSA measuring that the other methods don't... I don't know and I haven't found it yet.