(15-10-2025, 08:11 PM)magnesium Wrote: You are not allowed to view links. Register or Login to view.I find that k-means analysis recovers three clusters of bifolia:
I was a bit taken aback when, instead of the images you intended to show, there was a message which read, "Content now viewable in your region." The UK now has an Online Safety Act, ostensibly to protect children from unsuitable content. What has happened is that the UK's Information Commissioner's Office has been threatening foreign websites with colossal fines unless an age check is used to restrict access. Imgur has responded by blocking all access from the UK. Of course, anyone in the UK can use a VPN to see Imgur hosted images.
By the way, the original word boxes measured by "Job" on voynichese.com appear to be quire accurate. I have been able to compute similarities between pairs of pages (for horizontal text only), separating the minim heights, ascenders and descenders, taking into account the 'shape' of each word'.
Below is the resulting 'correlation'-type matrix. Top left is f1r. The numbers indicate the quires.
[
attachment=11734]
(15-10-2025, 11:53 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.I have a first set of boxes along the direction of writing, but the word boundaries are not yet very accurate. Would you be interested in using these?
Rene very graciously provided me with his data, so I reran my analysis over the weekend. Thanks to Rene's dataset containing the direction of writing,
k-means clustering prefers two clusters rather than three, as the "Scribe 4" cluster's aspect ratio has (correctly) risen, now that I can correct for non-horizontal word orientations.
You could think of this analysis as a fancy, quantitative confirmation of the common observation that the writing associated with Voynich B tends to be smaller than the writing associated with Voynich A. Here are four plots all of the same data, one coordinate per bifolio, with each plot varying in its labeling so you can see how each trait varies across the clusters:
[
attachment=11760]
As you can see, my Cluster 2 is mostly consistent with Voynich B, Zandbergen's B and C languages, and Davis's Scribes 2–5, while my Cluster 1 is mostly consistent with Voynich A, Zandbergen's A and Ae languages, and Davis's Scribe 1.
These plots are generated using all words with a minimum of 10 occurrences across the VMS. But I can also do plots for individual common words, too. Here's what happens when we look solely at
daiin:
[
attachment=11761]