For the choice of a previous word in the "self-citation" (with modification) method, even if the selection of words is non-deterministic and more or less random, humans don't act like computers, so there would certainly be a human bias in selection... detectable or not in the text of the VM? Would the detection of a bias be good evidence for the "self-citation" (with modification) method or not? I'm not sure.
Likely patterns for the selection of source-target words could be: use two (or more) consecutive words together for the next two generated words (in any order), or skip one word (or two) while reading, because it's easier to read several close words together than just one and then go to a totally different area of the page and read the next word.
Type 1a:
... source1 source2 ...
...
... target1 target2 ...
Type 1b:
... source2 source1 ...
...
... target1 target2 ...
Type 2a:
... source1 skipped source2 ...
...
... target1 target2 ...
Type 2b:
... source2 skipped source1 ...
...
... target1 target2 ...
I've only tested these patterns, with source words on the same line, target words on the same line, all on the same page. They create many more "hits" (possible source-target locations by small Levenshtein distance) than in a word-shuffled page. But we already know that word order is not random in the VM, so I'm not sure if the statistics really show bias in selection of source-target or if the word ordering biases (such as the known y_q affinity and others, known and unknown) get transferred to the similar but modified words. I would appreciate advice on how to resolve the issue.
A more general human bias for the selection of the next source word(s) would be the close proximity on the page, not necessarily on the same line: proximity between source and target or proximity between sources.
Note to self: I need to try various patterns, expected and unexpected, and explain better which hypotheses I'm trying to test.
Likely patterns for the selection of source-target words could be: use two (or more) consecutive words together for the next two generated words (in any order), or skip one word (or two) while reading, because it's easier to read several close words together than just one and then go to a totally different area of the page and read the next word.
Type 1a:
... source1 source2 ...
...
... target1 target2 ...
Type 1b:
... source2 source1 ...
...
... target1 target2 ...
Type 2a:
... source1 skipped source2 ...
...
... target1 target2 ...
Type 2b:
... source2 skipped source1 ...
...
... target1 target2 ...
I've only tested these patterns, with source words on the same line, target words on the same line, all on the same page. They create many more "hits" (possible source-target locations by small Levenshtein distance) than in a word-shuffled page. But we already know that word order is not random in the VM, so I'm not sure if the statistics really show bias in selection of source-target or if the word ordering biases (such as the known y_q affinity and others, known and unknown) get transferred to the similar but modified words. I would appreciate advice on how to resolve the issue.
A more general human bias for the selection of the next source word(s) would be the close proximity on the page, not necessarily on the same line: proximity between source and target or proximity between sources.
Note to self: I need to try various patterns, expected and unexpected, and explain better which hypotheses I'm trying to test.