"The Currier languages revisited" revisited

[kckluge]

3. Not using the text on circles and labels: I strongly suspect that this is the main reason for the visible differences. I especially suspect the circular texts. It may be worth to try this out.

To clarify, the experiment in the thread starter used line types

* P -- lines in running paragraphs
* C -- text on the circumference of circles in the circular diagram type pages,

but not

* R -- radial text in circles in the circular diagram type pages
* L -- labels

Running the bigram frequency stats on all lines of text results in this list of 40 most frequent bigrams (while there is some reordering of relative frequency, it looks like it mostly overlaps the list for just P & C lines):

Currier: 89 OF OE CC 4O C8 SC 8A C9 FC OP AM AR CO AE FA OR SO ZC O8 PC AN PA EF FS S9 PS ZO RA ES 9F S8 AJ FO PO F9 BS 2A 9P EO

Translated into EVA: dy ok ol ee qo ed che da ey ke ot aiin ar eo al ka or cho she od te ain ta lk kch chy tch sho ra lch yk chd am ko to ky pch sa yt lo

Looking at the effects of adding R and L lines of text into the data makes clear that it's the inclusion of labels (especially the Zodiac folio labels) that explains the difference. Here's the plot with "R" lines added to "P" & "C" lines -- there is minimal difference from the result without the "R" lines:

   

When the labels ("L" lines in the transcription file) are added in, it shifts the relative frequencies (as well as changing the PCA axes used for projection), producing a plot that looks very much like yours (f116v winds up being a big outlier on the plot's Y value, so I've clipped the range to keep it from reducing detail):

   

The labels in the zodiac pages are clearly the cause of the "bridge" in your original results. Even with the way they influence the PCA, if you leave them out of the plot the A & B language pages still split (with the exception of three Herbal A pages), just with a narrower gap (in that specific projection):

   

So that opens the whole "labelese" vs. running text issue. Looking at the Zodiac diagrams (f70v2 - f73v), here are the 20 most frequent bigrams in the labels:

 kgram:     AE     OP     OF     AR     CC     CO     89     PA     FC     OE
  Rank:      1      2      3      4      5      6      7      8      9     10
REFreq: 0.0741 0.0599 0.0537 0.0476 0.0435 0.0422 0.0388 0.0320 0.0299 0.0286

 kgram:     FA     EA     PC     O8     RA     E9     AM     AJ     8A     C9
  Rank:     11     12     13     14     15     16     17     18     19     20
REFreq: 0.0252 0.0245 0.0231 0.0218 0.0211 0.0204 0.0197 0.0170 0.0143 0.0136

Here are the 20 most frequent bigrams for the non-"L" lines on those pages:

 kgram:     OP     CO     CC     PC     AE     OF     C9     FC     OE     SC
  Rank:      1      2      3      4      5      6      7      8      9     10
REFreq: 0.0698 0.0604 0.0589 0.0546 0.0529 0.0425 0.0391 0.0354 0.0345 0.0339

 kgram:     AR     89     AM     8A     O8     ZC     C8     SO     EA     FA
  Rank:     11     12     13     14     15     16     17     18     19     20
REFreq: 0.0319 0.0313 0.0282 0.0256 0.0227 0.0175 0.0172 0.0167 0.0167 0.0138

To pick just a couple examples, 

* "SC" is the 10th most frequent glyph bigram on non-label lines at 3.39%, it doesn't even make the top 20 for labels

* "C9" is 3.91% of glyph bigrams for non-label lines of text, 1.36% for labels

* "AR" is 3.19% for non-label lines, 4.76% for labels

* "CO" is 6.04% for non-label lines, 4.22% for labels

Again, this is all within the zodiac plages. I think that makes a case that labelese and running text are different animals.

Finally, there is also this point: while Currier posited two different languages, intermediate forms (which he had not seen or at least not mentioned) exist. However, the amount of text of these intermediate forms is clearly less than both the 'clearly' A and 'clearly' B material. This also affects the appearance of the point clouds.

There are clearly "dialects" within the A & B languages, but I don't think the plots without labels included in my earlier posts in the thread support claims of intermediate forms between the two. There are two separable sets of pages (with ~3 exceptions as discussed in an earlier post), with the Herbal A, Pharma, Zodiac, and Astro pages falling on one side of the wall and the Herbal B, Bio, Rose foldout, and Starred paragraph pages falling on the other side. The Zodiac labels appear to be intermediate between them in a statistical sense, but given that they differ from the non-label text on the same pages I'm not sure I buy Zodiac labelese as an intermediate form between A language running text and B language running text. Your mileage may differ.