[Article] Not Deciphering The Voynich Manuscript

[Torsten]

An article about the Voynich manuscript at lesswrong.com: You are not allowed to view links. Register or Login to view.

The author argues the VMS is a meaningless fabrication, analysed with the Zandbergen-Landini transcription. Much of it replicates known results: the very low glyph-level conditional entropy from Lindemann & Bowern 2021, and Timm’s finding that near-identical words cluster both within lines and between adjacent lines beyond what a shuffled-page baseline predicts. The author pushes back on Lindemann & Bowern’s view that a hoaxer couldn’t have produced such language-like statistics. 

The new parts:

• He reports the Zipf fit explicitly (parameter 1.854), noting that an order-1 Markov process lands in the same natural-language range, so a Zipf fit is weak evidence either way.
  
• He trained a GRU on a subset of pages and on control corpora, then masked earlier lookalike words. Prediction degraded much more for Voynichese than for Latin or a Markov process, and slightly less than for text from the self-citation generator.
  
• The degradation is strongest when lookalike pairs differ at both the start and end of the word, which he argues fits a copy-and-edit process better than any real inflection system.
  

The comments are worth reading too: Michael Roe says he spent years with the Voynich discussion group, including Landini, and thinks the text is a made-up pseudolanguage, though he calls the statistical evidence highly suggestive rather than conclusive. He suggests checking whether the two proposed “authors” differ more than two writers of one language could.

It’s interesting to set this next to Michael Roe’s early contribution to the Voynich mailing list in 1993 (“A word to unlock the Voynich MS”). In that email he proposed "okal" as a possible way into the Vonyich text and suggested reading the word starts as case markers. voynich.nu presents the resulting word pattern in EVA in section 3.2.2 (“Mike Roe’s generic word”) of his word-structure page (see You are not allowed to view links. Register or Login to view.).

[npcompl33t]

If they trained the GRU on held-out pages from the entire manuscript, its pretty easy to understand the results. We already know the gradient is the number 1 feature, it dominates PCA1 and MDS1, and would no doubt be the exact thing the GRU matches onto. Masking a lookalike word degrades performance because the 'lookalikeness' of the word is what allows the GRU to locate where on the curve it is and how it should adjust its probabilities. They even say this in their statement:

" In other words, the trained model relies on the earlier lookalike in downstream prediction."

This person should have read your most recent paper about the problems applying static-models to a dynamic-text!

[Mauro]

The two articles referenced in the 'Coda' section are also quite interesting, expecially the first one about the Papal cipher. I was unaware there existed deliberately ambiguos ciphers where the decrypter had to supply additional information (mostly by his knowledge of the language) for a successful decryption. My guess (but just a guess) is this reduces the entropy of the encrypted text, given information has been removed from it, so this might be relevant to the VMS too (not that I'm endorsing the cipher hypotheses, I'm pretty much agnostic at the moment).