I believe this is a side of the question we have not yet considered in this thread: which words appear immediately before and after bench-words. As a first experiment, I have looked at the most common couple chedy/shedy. I hope these results are reasonably accurate, but double-checking would be appreciated.
I have only considered pages in which at least one of the two words appears. For each of the most frequent words in these pages, I have counted the number of occurrences of:
- W.chedy
- chedy.W
- W.shedy
- shedy.W
The counts have then been divided by the number of occurrences of W in these pages, giving the % of occurrences of W that appear immediately before or after chedy/shedy. The expected frequency is about 2% for combinations with chedy and 1.7% for combinations with shedy. Words are sorted by decreasing frequency, so data near the right side of the graph are less significant.
Once again, the two words appear to behave quite similarly. The greatest difference could be that 'daiin' follows chedy but never follows shedy. Another difference is that shedy.chedy is considerably more frequent than chedy.shedy (8 vs 3 occurrences); but we already discussed how chedy tends to appear nearer the end of the line than shedy. Third, here it seems that shedy is more often followed by -edy words: this could be something worth testing for more general cases.
[
attachment=3729]
For comparison, I have also looked at the behaviour of qokeedy, which occurs 305 times in this pages (1.2% frequency). In this case, the differences are quite large, look for instance at the combination with ol, daiin, qokain, chey, or. In general, one can observe that (in this set of data) qokeedy never follows -in, while chedy/shedy often do. Given the Voynichese tendency to quasi-reduplication, it seems also interesting that chey never appears before chedy/shedy (while it is frequent before qokeedy).
If Voynichese words are meaningful, it seems that chedy and shedy must have a closely related meaning/function, while qokeedy appears to be quite different. On the other hand, the differences can be largely explained by the general dislike of q-words to follow -in and the opposite preference of bench-words to do so.
Picking up on this message:
(18-11-2019, 03:54 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.I did not yet list the percentages of chedy and Shedy occurences for the different sections.
They are:
chedy shedy
----- -----
Herbal B: 1.933 1.132
Biological: 2.979 3.485
Cosmo: 1.155 0.880
Stars (all): 1.732 1.048
Stars-Bio: 1.984 1.297
Stars (rest): 1.355 0.677
Total: 2.092 1.780
[font=Sans-serif]This shows that there is a general trend, but still significant variability between the different sections.[/font]
[font=Sans-serif]It is mainly in the Biological section that there is also an internal correlation.[/font]
[font=Sans-serif]I wonder if this could be related to the hypothesis that this quire is a mixture of two original quires. There has been quite some speculation about that [/font]
Looking back at the information at Ciphermysteries, the proposed split of Q13 is:
Q13A : fol 76+83 , 77+82 , 79+80
Q13B : fol 75+84 , 78+81
This is a split of bifolios into two groups based mainly on the illustrations. Q13A is presumed to be the 'earlier' one since it starts with a text-only page.
This leaves only 12 'observations' for Q13A and 8 for Q13B.
Computing the statistics using nablator's approach shows correlations of:
+0.718 for Q13A
+0.463 for Q13B
Is this significant? If so, it is certainly only marginal, but it would be interesting to see how a selection based on the illustrations shows up in the text.
What is mostly visible is that the variation in fractions of
Shedy per page is similar in both parts, while the variation of
chedy is much less in Q13B. The ratios of
Shedy /
chedy for the two parts are:
1.11 for Q13A
1.29 for Q13B
Again this seems rather marginal.
(24-11-2019, 03:28 PM)MarcoP Wrote: You are not allowed to view links. Register or Login to view.I believe this is a side of the question we have not yet considered in this thread: which words appear immediately before and after bench-words. As a first experiment, ...
I did that for every letter and also for every word in the text. (don't know if I published this, but probably I did)
Also I compared it with the many variations that one can perform. The final observations are not different than what we already know: the letters do not behave as normal letters in other European languages. From there I went on and formed new conclusions, which are out of scope.
MarcoP Wrote:
If Voynichese words are meaningful, it seems that chedy and shedy must have a closely related meaning/function, while qokeedy appears to be quite different. On the other hand, the differences can be largely explained by the general dislike of q-words to follow -in and the opposite preference of bench-words to do so.[font=Tahoma, Verdana, Arial, sans-serif][/font]
Your final conclusion is important, because that is exactly how I looked at the text: if there are "vords" before, after (or inside) the "thing" we are looking at,
and those "vords" or parts (letters or combined letters) behave the same, we can create a map with what I call dependencies. These dependencies are rigid.
Rigid as in semi-absolute: these
can occur, where other "parts"/"vords"
can not occur on that position.
Based on those dependencies, I created a kind of map of possibilities. If you then compare that with the "normal plain text alphabetical" dependencies, you can see that they indeed seem to behave somewhat normal. I did not move on from that point forward, because the amount of information is too big for me to consume and map.
(25-11-2019, 06:38 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.Looking back at the information at Ciphermysteries, the proposed split of Q13 is:
Q13A : fol 76+83 , 77+82 , 79+80
Q13B : fol 75+84 , 78+81
This is a split of bifolios into two groups based mainly on the illustrations. Q13A is presumed to be the 'earlier' one since it starts with a text-only page.
This leaves only 12 'observations' for Q13A and 8 for Q13B.
Computing the statistics using nablator's approach shows correlations of:
+0.718 for Q13A
+0.463 for Q13B
Is this significant? If so, it is certainly only marginal, but it would be interesting to see how a selection based on the illustrations shows up in the text.
This was not so difficult to check. Trying a few other potential splits results in some cases where the two correlations are much more similar, but also others where the two correlations were even further apart.
There is therefore no evidence that this particular split is "special" from the text point of view.
At least from this very specific aspect.
(25-11-2019, 11:52 PM)Torsten Wrote: You are not allowed to view links. Register or Login to view.The correlation between <chedy> and <qokeedy> is therefore even stronger as the correlation between <qokeedy> and <qokedy> or the correlation between <okeedy> and <okedy>.
Pearson's Correlation(chedy[501], shedy[426]) : +0.84 (n=225)
Pearson's Correlation(qokeedy[305],chedy[501]) : +0.71 (n=225)
...
These are again the absolute values, and it was already demonstrated that these are influenced to a large extent by the variations in the page sizes. It would be more meaningful to give the corrected values.
(05-05-2020, 12:59 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
Torsten, yes, I think that the self-citation method does not adequately explain the text in the Voynich MS, and I am also convinced that this is not 'how it was done'.
That is my opinion of course, just like your opinion is the opposite.
However, like you, I do my best to explain on what my opinion is based.
It is possible to discuss with the goal to increase the knowledge about a subject. However it also possible to discuss with the goal to campaign for some opinion. Let's check if René does indeed his best to explain his point of view. Let us use this short thread as an example. Here we discuss just some observations for sh_ and ch_ words. It should be easy to come to an agreement.
I calculated the correlation coefficient for some sh_ words and the corresponding ch_ words and argued that the they imply that the chance for a 'sh'-word to occur on a folio increases as more often the corresponding 'ch'-word appears on that folio (see You are not allowed to view links. Register or Login to view.).
At first René did not denied my calculations but argued that they wouldn't mean much since the results would be dominated by the varying page length:
(14-11-2019, 08:44 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
To illustrate the fact that varying page length is dominating these results, lets look at the hypothetical case that the Voynich manuscript text were composed by picking words arbitrarily from a hat.
This causes a positive correlation between ch-words and sh-words, without any causality.
As René writes himself this is just a hypothetical scenario. He didn't provide any evidence that ch_ and sh_ word tokens are indeed randomly distributed within the Voynich manuscript.
Let's illustrate the hypothetical scenario with a simplified example. Let's assume that we have four pages with 100, 200, 100, and 300 words. On each page every tenth word is a 'Word A' and every fifth word is a 'Word B':
| page | word count | count(Word A) | count(Word B) |
| 1 | 100 | 10 | 20 |
| 2 | 200 | 20 | 40 |
| 3 | 100 | 10 | 20 |
| 4 | 300 | 30 | 60 |
I
n this hypothetical example it is possible to use the word count for a page to calculate the number of occurrences for 'Word B' but it is also possible to use the counts for 'Word A' for this purpose. The correlation coefficient between 'Word A' and 'Word B' as well as between the word count and 'Word B' is 1.0:
Pearson's Correlation(count(Word A), count(Word B)): +1.00
Pearson's Correlation(count(page), count(Word B)): +1.00
The cause for the correlation values are the rules for generating the sample text. If every tenth word is a 'Word A' and every fifth word is a 'Word B' it is possible to use the count for one word to calculate the count for the other one no matter what the page sizes are.
René uses his hypothetical scenario to explain the correlation coefficient for the word pair 'chol' / 'shol' (First René writes chol / chor instead of chol / shol. Maybe since Currier used 'chol' and 'chor' as examples for word types typical for language A (You are not allowed to view links. Register or Login to view.). The correlation coefficient for 'chol' / 'chor' is 0.42 (Pearson's Correlation(chol[396],chor[219]): 0.42)).
(14-11-2019, 08:44 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
Because of all this, it is also interesting to look at the pair chol / chor.
The reason is that these words are relatively infrequent on the longer pages, which are all in Currier-B. As a result, the variation of absolute counts per page is much more limited.
The correlation for this case is given as +0.44
This seems like a high positive correlation but for people used to working with correlation coefficients, it is not.
The hypothetical scenario is based on the assumption that the tokens are randomly distributed. However René argues that chol / shol are not randomly distributed. Moreover he points out that all longer folios are in Currier-B and that 'chol' as well as 'shol' are infrequently used in Currier B. With other words René is arguing that his hypothetical scenario doesn't fit for this two words. This doesn't come as a surprise. There exists in fact not a single word type in the manuscript that is randomly distributed.
Moreover, Currier even described two 'languages' Currier A and B using different sets of words: "There are two different series of agglomerations of symbols or letters, so that there are in fact two statistically distinguishable 'languages'" (You are not allowed to view links.
Register or
Login to view.
). Typical for Currier A are for instance the words like 'chol' and 'shol' and typical for Currier B are for instance the words like 'chedy' and 'shedy'.
There were some answers to Renés post.
MarcoP suggested to use percentages and to remove pages with zero matches to 'correct' the results:
(15-11-2019, 10:42 AM)MarcoP Wrote: You are not allowed to view links. Register or Login to view.
Torsten measured a correlation coefficient of 0.84
By removing pages at 0,0, correlation drops to 0.74
By also considering % instead of absolute counts, correlation is 0.57
Nablator suggested to use the expected number of words to 'correct' the results:
(14-11-2019, 02:37 PM)nablator Wrote: You are not allowed to view links. Register or Login to view.
To discount this linear relation (on average) between the two, the number of words of a certain type on a page can be compared to the expected number of words (from the global frequency ratio and the total number of words on the page) to see whether there are more or less words of this type than expected on the page.
I suggested to check the results for folios with similar word counts:
(14-11-2019, 10:31 PM)Torsten Wrote: You are not allowed to view links. Register or Login to view.
To exclude an effect of the folio size I have calculated the correlation coefficient for folios containing less text (herbal folios in quire 1 to 7) and for folios containing lot of text (Quire 13 and 20).
René completely ignored my post and argued:
(16-11-2019, 06:51 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
The effect of varying page length, which causes an artificial positive correlation, seems to be effectively eliminated by two different methods:
1) the one proposed by @nablator, by subtracting the expected values based on some average
2) by computing the percentages
René argued further:
(17-11-2019, 03:25 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
I have no satisfactory explanation for the phenomenon, but the artefacts from varying page length are important, and I think sufficiently demonstrated. Any pair of words that occur consistently throughout a meaningful text will show this type of correlation if the text is cut in unequal pieces and the correlation is computed over these pieces.
One can normalise in several different ways. Two have been used here.
But there is no need to normalize correlation values. The correlation coefficient didn't depend on the cause. There is also no upper limit of how many correlations for a given observation can exist. Therefore it doesn't make any sense to use one correlation value to 'correct' another one. Moreover, René didn't provide any evidence that it is possible to correct the correlation coefficient or that the suggested methods can be used to do so. He accepts the 'corrected' correlation values only since lower values are favorable to his thoughts. But since something different was calculated the results also mean something different.
Let's use the simple example again to illustrate how the used 'normalization' methods work. By using the expected counts we get only zero values and by using percentages we get always 10 % for 'Word A' and 20 % for 'Word B'. In both cases the values didn't differ and it is therefore not possible to calculate a correlation for them. In order to get a result some variation is necessary for both words. Let us therefore modify the count for 'Word B' on page 2 from 40 to 39 and the count for 'Word A' on page 4 from 30 to 29. By using the absolute values the correlation coefficient only drops from 1.0 to 0.9986.
However, if the expected word counts are used the result becomes unpredictable:
| page | word count | count(Word A) | count(Word B) | expected(Word A) | expected(Word B) | delta(Word A) | delta(Word B) |
| 1 | 100 | 10 | 20 | 9.86 | 19,86 | 0.14 | 0.14 |
| 2 | 200 | 20 | 39 | 19.71 | 39.71 | 0.29 | -0.71 |
| 3 | 100 | 10 | 20 | 9.86 | 19.86 | 0.14 | 0.14 |
| 4 | 300 | 29 | 60 | 29.57 | 59.57 | -0.57 | 0.42 |
| sum | 700 | 69 | 139 | 69 | 139 | 0 | 0 |
By using the delta to the expected word counts we got -0.71 as result. The reason is that the delta values are defined as distributed around the mean values. Therefore the sum of all delta values must be zero. This way small differences become important.
Pearson's Correlation(delta(Word A),delta(Word B)): -0.71 (n=4 folios)
By using the percentages the result is -0.33:
| page | word count | count(Word A) | count(Word B) | Word A in % | Word B in % |
| 1 | 100 | 10 | 20 | 10 | 20 |
| 2 | 200 | 20 | 39 | 10 | 19.5 |
| 3 | 100 | 10 | 20 | 10 | 20 |
| 4 | 300 | 29 | 60 | 9,67 | 20 |
Pearson's Correlation(Word A in %,Word B in %): -0.33 (n=4 folios)
The resulting percentages are very similar to each other. As pointed out in You are not allowed to view links. Register or Login to view. the Pearson's Correlation cannot be used to calculate the correlation coefficient for percent values. However René ignored this argument as well.
For both calculation methods small changes have a high impact. Both methods become unstable if the values are indeed randomly distributed and therefore close to the expected values. However, this is not what we see for the Voynich manuscript. Since the counts are not randomly distributed in the manuscript the 'corrected' results are only smaller. But what does this smaller values mean?
They mean that the results does not depend on the page size and that it is still necessary to find an explanation for the positive correlation values. Moreover, if it is possible to demonstrate a correlation even for the delta values for words like 'chedy' / 'shedy' this indicates a rather close relation between this two words.
However, René argues the opposite way:
(17-11-2019, 03:25 PM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
I have no satisfactory explanation for the phenomenon, but the artefacts from varying page length are important, and I think sufficiently demonstrated.
Now René even argues that it would be wrong to calculate the correlation coefficient correctly by using the absolute values:
(27-11-2019, 05:55 AM)ReneZ Wrote: You are not allowed to view links. Register or Login to view.
These are again the absolute values, and it was already demonstrated that these are influenced to a large extent by the variations in the page sizes. It would be more meaningful to give the corrected values.
Why does this discussion matter? It matters since what is typical for sh_ and ch_ words is in fact typical for all words in the Voynich manuscript. 'Isolated' sh_ words (i.e. were the corresponding ch_ word doesn't exist) occur rarely, while for all 'sh'-word-types with at least 4 tokens also the corresponding 'ch'-word exists. This happens since this is just an example of a general rule for the Voynich manuscript. 'Isolated' words (i.e. without any similar word type) usually appear just once, while for the most frequent types also many similar words exists (Timm & Schinner 2020, p. 6). Since the Voynich manuscript behaves You are not allowed to view links.
Register or
Login to view.
in this regard, this can be observed for a single folio as well as for Currier A and B and also for the whole manuscript.
Decide yourself if Renés posts were indeed helpful to increase our knowledge about ch_ and sh_ words.
(05-11-2019, 08:56 PM)Davidsch Wrote: You are not allowed to view links. Register or Login to view.Based on EVA.
1.
I assume that sh_ and ch_ words are the same, because they behave the same and have same contacts
2.
If sh... words are always compacter than ch..words in the entire text,
3.
could this strongly signal that sh...words are compacter because the diacritic mark above signals an abbreviation?
4.
if the answer is yes, can we assume that sh... words are abbreviated versions on the same ch... words and what letters could be compacted there?
CH and SH have similar function and similar sound, but they are not the same. In Slovenian, CH represents the sound Č which in the Middle Ages was spelled as CH or ZH. Slavic languages distinguished two Č sounds - soft and hard. Croatians and Serbians still use two different letters: ć and č. Since they were hard to tell apart, Slovenians in 16th century adapted only one Č for both Slavic sounds. There are also two other Slavic letters that fall into this category: the letters Š (sh) and Ž (similar to J in joy) . In some words, Š sounds more like Č, in others more like Ž. The differentiation of these sounds is still causing spelling mistakes in Slovenian. As a separate words, CH and SH are dialectal phonetic spelling of the words CHE and SHE. In fast speech, the vowels, and some consonants were often dropped, and sometimes such short words were spelled together with the next word. The word CHE means 'if', and 'še' means 'yet'. Voynich EVA words RCHY and RSHY (assuming the E as a second letter is inserted to replace unwritten semi-vowel, and Y changed to I, since Y does not exist in Slovenian), are two distinct words: RECHI means 'things' or 'you say!' and REŠI means 'save', 'protect'. To add to the confusion, sometimes C was used for sound Č, or CI for ČI, S for Š or Ž. On the other hand, ROŽA (flower) was often spelled as ROSA, but since this word also means 'a dew', some spelled it as ROSHA - ROŠA. I noticed the author of the VM had a lot of problems with this letters. This was further complicated because the VM-h had the same shape as C (Latin K), and Z had similar shape as C ( as EVA-h), because the connecting line in cursive letters was an extended straight line of C.