Vord paradigm tool

[MarcoP]

TLDR: Voynchese word ordering is dominated by line effects and repetition. A model based on the alternation of two classes 0,1,0,1,0,1... cannot be accurate, because lines are typically made of an "anomalous" line-initial word, a sequence of words with repetitions and line-positional preferences and a differently anomalous line-final word. In my opinion, a model is likely to require at least 4 classes; lines could be something like S,C,C,N,C,C,C,N,E (S=start, C=occurring consecutively, N=non-repeating, E-end) or maybe S,L,L,L,R,L,R,R,R,E (L=left-preference, R=right-preference).

---

After the ideas put forward in this thread, I ran a series of experiments along the lines discussed in You are not allowed to view links. Register or Login to view.. I used Alexander Clark's implementation of the Ney-Essen algorithm, which classifies words on the basis of their behaviour only, ignoring morphology.

I used Quire 20 from the Zanbergen-Landini transliteration, where benches and benched-gallows were replaced with upper-case characters and 'z' correspond to unreadable or rare characters. Uncertain spaces were ignored. As a first experiment, I only considered the first character of each word and configured the algorithm to find two classes. In this graph, the thickness of the arrows corresponds to the frequency of transitions (only the most frequent words for each class appear in this and the following graphs).

   

Class:1 includes the line separator (nlnl) and the  characters that typically appear at the beginning of a line. This class is typically followed by a word belonging to the other class. Class:0 includes all the most frequent word initials and this class tends to loop on itself, meaning that words belonging to this class tend to appear consecutively.

So a typical line could be something like:

nlnl 1 0 0 0 1 0 0 0 nlnl

This is the parsing of
<f103r.10,+P0>    daiin.ol.oain.okeol.chol.okam.chety.shedy.otaiin.shedy.teolshy.oteedy.sor,ain
1:nlnl nlnl
1:d daiin
0:o ol
0:o oain
0:o okeol
0:C Col
0:o okam
0:C Cety
0:S Sedy
0:o otaiin
0:S Sedy
1:t teolSy
0:o oteedy
1: s sorain
1:nlnl nlnl

So this result is dominated by the line-initial effects that make the first character of line initial words unusual in other positions.

---

I then tried a similar experiment considering both the first and last character for each word.

   

These results are more difficult to interpret. Class:0 only includes the following 11 words. It could be that they share the property of only rarely repeating identically, but I am not sure:

nlnl 1086
Cy 850
Sy 445
an 220
ar 167
Sl 85
Co 42
So 24
Sd 23
Cd 23
lo 19

---

Things get more interesting if the number of classes is increased. With 4 classes, the line-break takes a whole class (C:0); C:3 contains line-initial words; C:2 contains the most frequent words that also tend to occur repeatedly; C:3 contains words that do not repeat and all -m words (which typically are line-final).

   

---

Four classes only based on the first character of words cannot clearly identify line-final words (which have peculiar endings). But sh- and q- words tend to avoid the right-side of lines (see You are not allowed to view links. Register or Login to view.) and here they get their own class (C:1) which is not likely to directly lead to a line-break. C:3 groups line-initial characters and C:0 is the bulk of ordinary characters.