Who is even still working on a solution for the VMS?

[quimqu]

I am working on a solution. And it's quite promising, actually. Here is the MD5 of the main proposition, since such things seem to be the common practice around here:

    ec1581a8934c179303b8bd94c58ccc9d

The theory is that the VMS text is meaningless. I have a system that can generate Voynichese (but without meaning), and accounts for the various characteristics of the text:

• Repeating words,
  
• Consecutive words that only differ by one or two "tokens",
  
• A system of common prefixes and suffixes,
  
• Positive autocorrelation of word lengths (words that are in short-short-long-long order rather than long-short-long-short as one would expect in a natural language),
  
• "Topic-words", i.e., larger structural features of the text, in which page-specific "topics" seem to emerge (which Gaskell & Bowern mentioned in their paper as being a major limitation of today's theories that the VMS text is meaningless).
  

Furthermore, the method I found is consistent with what was available with early 15th century techniques, so does not require a Cardan-Grille like Rugg & Taylor suggested (which came much later), or fancy disks and Fontana-like mnemonic tecnhiques.

This, of course, would not prove that the VMS is gibberish. Nor would it exclude that some Voynichese words are meaningful but steganographically hidden in meaningless text. This method would simply show a way to generate meaningless Voynichese without much trouble.

Of course, I don't make any assumptions on why one would generate 200 pages of meaningless text.

Anyway, once I'll write it up, in a way that is clear for readers, I will post it here

Hello,

I have worked on a very simple substitution model that generates Voynich-like text. It matches the entropy values, follows a very similar Zipf distribution, and is surprisingly easy to implement.

The basic idea is to find a natural text with a similar Zipf curve and then assume that Voynich words were generated beforehand through small modifications (changing a character, adding one, deleting one, etc., roughly Levenshtein distance = 1). In my case, I found that "De docta ignorantia" has the best fit in terms of Zipf. After that, I simply replace the most frequent natural-language word with the most frequent Voynich word, the second most frequent with the second most frequent, and so on.

The real problem is not generating Voynich-like words. The real problem is reproducing the text structure: the use of gallows at the beginning of paragraphs, the distribution of word endings at the ends of lines, the preferred prefixes at the beginnings of lines, and similar positional effects.

In other words, generating Voynich-like words and replacing one word with another can reproduce many of the global statistical properties almost automatically. But reproducing the structural patterns of the manuscript is much harder.

Does your method generate these structural features as well?