Here are some results accomplished this far on the WALLA.doc text file. These results are mainly about the inter-word relationships and word sequences. I will probably present further analyses later, but then mostly on the intra-word relationships (characters and character relationships). First of all I must comment on the limited size/length of the Walla corpus that can have impact on most of the statistics. I have not included the ’=’ (equal) sign which Walla used extensively, but instead treated it as a word-spacing marker.
Some basic statistics:
Average word length (counting each unique word only once): 7.29446 characters, standard deviation: 4.5279
Average word length (text aggregate): 6.62016 characters, standard deviation: 4.24014
Number of unique words: 1355
Total number of words (tokens): 1885
Number of lines of text analyzed: 974
First of all, the type-token ratio (number of unique words/total number of tokens) is extremely high (72%). This figure could possibly have decreased if the text would’we been running longer. But let's anyway jump to the results. I have run analysis to identify repeated sequences of 2-6 words in length. Most meaningful texts in any language contains a lot of these, but the most puzzling thing about the Voynich manuscript is that the repeated sequences 3-4 words or more in length drops of dramatically, while the number of repeated 2-word sequences (pairs) corresponds to what you find for many natural language texts (still a bit low in numbers, but not incompatibly so). Below is a graph over amount of repeated n-tuples (n = 2 to 6) in percentage for different texts. The files are some of the text files in Brian Cham’s corpus. I have also thrown in some files from the Project Gutenberg (ebook) website (prefixed by ”pg_”). One of these is an english poetry sample, and another is an encyclopedia sample in english. These were mostly used to illustrate the difference between various types of texts. The results from the Walla text is furthest right in the graph, and next to it is the Voynich Manuscript (VM). All of these texts that were longer than about 40 000 tokens have been truncated to the first 40 000 tokens, both to save time and to make their statistics more compatible with the VM which has this size.
As you can see there is large amount of variation in the repeated sequence-data. Although no text except the N-GRN (Guarani language) sample contains a complete lack of 5-6-tuples, if you disregard the one-letter/glyph-sequences (it is dubious to treat these glyphs as as individual words) in the circular diagrams in the VM, then the VM also completely lacks 5- or 6-tuples. It is difficult to see the low numbers in the diagrams but you can also have a look at the raw data (attached ’data.xls’ file). The N-GRN sample contains text that is of deliberate poor quality which is the probable reson for the absense of these n-tuples.
The Walla text could be seen as a bit unusual in a way. As expected, because of the high number of unique words, the number of repeated pairs is low, however contradictary to this is the comparatively high number of longer sequences, which is more compatible with texts of known natural languages than with the VM. Additionally, Walla wrote long sequences in all upper-case as well as all lower-case both, and since this analysis was case-sensitive you would expect an even lower number of all n-tuples than if the whole text were mostly either-case. So, even though Walla invented his own words which were both many and significantly longer than words in most languages, he managed to write identical longer sequences repeatedly. It should be mentioned that all of the repeated word-pairs/word-n-tuples in Walla’s text were repeated only once, except two pairs which were repeated twice each. These are ”Alois Walla” and ”Walla George” – both his name?
I am hesitant to draw any more conclusions as of now because of the limited text size, but I am still working on/thinking of what more I can do with it. Ben, more text would be very welcomed if you have time.
Some basic statistics:
Average word length (counting each unique word only once): 7.29446 characters, standard deviation: 4.5279
Average word length (text aggregate): 6.62016 characters, standard deviation: 4.24014
Number of unique words: 1355
Total number of words (tokens): 1885
Number of lines of text analyzed: 974
First of all, the type-token ratio (number of unique words/total number of tokens) is extremely high (72%). This figure could possibly have decreased if the text would’we been running longer. But let's anyway jump to the results. I have run analysis to identify repeated sequences of 2-6 words in length. Most meaningful texts in any language contains a lot of these, but the most puzzling thing about the Voynich manuscript is that the repeated sequences 3-4 words or more in length drops of dramatically, while the number of repeated 2-word sequences (pairs) corresponds to what you find for many natural language texts (still a bit low in numbers, but not incompatibly so). Below is a graph over amount of repeated n-tuples (n = 2 to 6) in percentage for different texts. The files are some of the text files in Brian Cham’s corpus. I have also thrown in some files from the Project Gutenberg (ebook) website (prefixed by ”pg_”). One of these is an english poetry sample, and another is an encyclopedia sample in english. These were mostly used to illustrate the difference between various types of texts. The results from the Walla text is furthest right in the graph, and next to it is the Voynich Manuscript (VM). All of these texts that were longer than about 40 000 tokens have been truncated to the first 40 000 tokens, both to save time and to make their statistics more compatible with the VM which has this size.
As you can see there is large amount of variation in the repeated sequence-data. Although no text except the N-GRN (Guarani language) sample contains a complete lack of 5-6-tuples, if you disregard the one-letter/glyph-sequences (it is dubious to treat these glyphs as as individual words) in the circular diagrams in the VM, then the VM also completely lacks 5- or 6-tuples. It is difficult to see the low numbers in the diagrams but you can also have a look at the raw data (attached ’data.xls’ file). The N-GRN sample contains text that is of deliberate poor quality which is the probable reson for the absense of these n-tuples.
The Walla text could be seen as a bit unusual in a way. As expected, because of the high number of unique words, the number of repeated pairs is low, however contradictary to this is the comparatively high number of longer sequences, which is more compatible with texts of known natural languages than with the VM. Additionally, Walla wrote long sequences in all upper-case as well as all lower-case both, and since this analysis was case-sensitive you would expect an even lower number of all n-tuples than if the whole text were mostly either-case. So, even though Walla invented his own words which were both many and significantly longer than words in most languages, he managed to write identical longer sequences repeatedly. It should be mentioned that all of the repeated word-pairs/word-n-tuples in Walla’s text were repeated only once, except two pairs which were repeated twice each. These are ”Alois Walla” and ”Walla George” – both his name?
I am hesitant to draw any more conclusions as of now because of the limited text size, but I am still working on/thinking of what more I can do with it. Ben, more text would be very welcomed if you have time.
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)