Illustration-aware analysis

The drawings are the only part of the manuscript we can read. Treating each page's illustration type as a topic label, we can ask which words are unusually attracted to which kind of picture. Word identity carries 1.017 bits of mutual information with illustration type — the text is not topic-neutral, which is the precondition for any semantic hypothesis.

observed countsdomain assignment

herbal

129 pages · 10,703 tokens

Full-page plant drawings with roots, stems and leaves; the largest section of the book.

wordsegmentationclassherePMI
cthycth·yC792 (88%)1.59
chorch·orC7146 (71%)1.30
cthorcth·orC738 (90%)1.64
shosh·oC788 (73%)1.33
qotchyqo·tch·yC751 (82%)1.50
ctholcth·olC747 (82%)1.51
kchyk·ch·yC626 (93%)1.68
cholch·olC7219 (58%)0.99
daiind·ai·inC1423 (53%)0.86
chych·yC788 (65%)1.16

stars

23 pages · 10,660 tokens

Dense short paragraphs each marked with a star in the margin.

wordsegmentationclassherePMI
lkeeyl·ke·eyC136 (88%)1.60
lkaiinl·kai·inC140 (82%)1.50
lkeedyl·ke·edyC133 (80%)1.48
lkainl·ka·inC129 (83%)1.52
cheeoch·ee·oC715 (94%)1.70
alalC4136 (54%)0.91
oteedyot·ee·dyC160 (61%)1.07
qokchedyqo·kch·edyC128 (72%)1.31
okeeyok·ee·yC196 (55%)0.93
qokeeyqo·ke·eyC1159 (52%)0.84

biological

20 pages · 6,862 tokens

Small figures in pools and connected channels of liquid.

wordsegmentationclassherePMI
qolqo·lC1115 (78%)2.06
shedysh·ed·yC1247 (58%)1.65
qokedyqo·ke·dyC1161 (61%)1.71
qokainqo·ka·inC1158 (57%)1.62
qokalqo·ka·lC1105 (56%)1.59
qokeedyqo·ke·edyC1150 (50%)1.42
ololC1233 (44%)1.25
olkainol·ka·inC125 (76%)2.03
chedych·ed·yC1208 (42%)1.17
solso·lC139 (58%)1.65

pharmaceutical

16 pages · 2,476 tokens

Rows of jars and detached plant parts, often with short labels.

wordsegmentationclassherePMI
okeolok·eo·lC730 (46%)2.78
qokeolqo·ke·olC723 (46%)2.78
qokeodyqo·keo·dyC214 (45%)2.75
ckheyckh·eyC713 (42%)2.64
cheolch·eo·lC137 (22%)1.73
qoorq·oo·rC56 (75%)3.48
okolo·ko·lC721 (27%)2.00
cheorch·eo·rC424 (24%)1.87
ctheolcthe·olC76 (60%)3.16
qokeorqo·ke·orC28 (38%)2.51

cosmological

12 pages · 2,460 tokens

Large foldout diagrams of nested circles and regions.

wordsegmentationclassherePMI
ytaiiny·tai·inC216 (39%)2.55
ykaiiny·kai·inC715 (34%)2.35
olkarol·ka·rC59 (47%)2.83
yteeyy·te·eyC410 (40%)2.58
ararC458 (17%)1.34
ytary·ta·rC58 (31%)2.21
oaroa·rC56 (38%)2.49
aiiinaii·inC410 (24%)1.87
shedaiinsh·ed·aiinC25 (38%)2.53
ytodyy·to·dyC24 (50%)2.91

text-only

5 pages · 1,612 tokens

Pages carrying text with no substantial illustration.

wordsegmentationclassherePMI
qokalyqo·kal·yC67 (47%)3.42
dalyd·al·yC67 (25%)2.52
chalch·alC19 (20%)2.16
ykeeodyy·keeo·dyC34 (36%)3.06
okalyok·al·yC65 (23%)2.38
ykaly·ka·lC74 (27%)2.61
dalda·lC123 (9%)1.12
qokairqo·kai·rC54 (24%)2.43
talta·lC14 (21%)2.27
yyC713 (9%)1.11

zodiac

12 pages · 1,292 tokens

Ringed diagrams with figures arranged around a central emblem.

wordsegmentationclassherePMI
oteosot·eo·sC415 (54%)3.94
oteodyot·eo·dyC215 (39%)3.49
okeosok·eo·sC46 (50%)3.84
otoot·oC75 (56%)3.99
oteesot·ee·sC45 (50%)3.84
oteeyot·ee·yC118 (13%)1.93
alalC428 (11%)1.67
okeoo·ke·oC75 (36%)3.35
oteyo·te·yC110 (18%)2.32
okalyok·al·yC66 (27%)2.96

astronomical

8 pages · 830 tokens

Circular diagrams with stars, suns and radiating rays.

wordsegmentationclassherePMI
daiird·aii·rC57 (32%)3.82
shess·he·sC44 (31%)3.77
daird·ai·rC47 (7%)1.64
sarsa·rC46 (7%)1.68
oteeyot·ee·yC18 (6%)1.40
cheoche·oC75 (8%)1.77
okeyo·ke·yC14 (6%)1.52
okolo·ko·lC74 (5%)1.19
darda·rC110 (3%)0.58
cheorch·eo·rC44 (4%)0.86

how this feeds translation

From picture to probabilistic semantics

An association is not a meaning. It narrows the semantic field a slot may draw from, and it does so with a number attached.

  1. 1. A word is counted on every page type it appears on. observed
  2. 2. Its PMI with each illustration type is computed; anything above ~0.35 bits is treated as a domain preference. observed
  3. 3. The word's induced class decides which kind of slot it fills — entity, action, quality or connective. inferred
  4. 4. Only then is a placeholder English word chosen from that domain and slot kind. speculative

Step 4 is deliberately reversible: change the placeholder vocabulary and every reading changes, while steps 1–3 and the predictive scores stay exactly the same. That is the boundary between this project's evidence and its speculation.