13-05-2026, 03:52 PM
In the recipe section, I annotated the stars showing a tail and the stars without a tail.
I then trained a simple statistical model to predict whether a paragraph maps to a “tail” star using only internal text features (no external information, no hand labeling tricks).
The model correctly identifies tail paragraphs about 86% of the time (balanced accuracy), which is far above chance (50%).
(I also ran permutation tests where we randomly reshuffled the data within each folio, preserving structure but destroying signal. In those tests, performance dropped to about 52%, which is basically chance. This suggests the model is not just exploiting simple artifacts like folio grouping or class imbalance.)
Is anyone aware of similar studies? I'd be interested to know if star morphometrics correlate to textual contents.
Thanks!
I then trained a simple statistical model to predict whether a paragraph maps to a “tail” star using only internal text features (no external information, no hand labeling tricks).
The model correctly identifies tail paragraphs about 86% of the time (balanced accuracy), which is far above chance (50%).
(I also ran permutation tests where we randomly reshuffled the data within each folio, preserving structure but destroying signal. In those tests, performance dropped to about 52%, which is basically chance. This suggests the model is not just exploiting simple artifacts like folio grouping or class imbalance.)
Is anyone aware of similar studies? I'd be interested to know if star morphometrics correlate to textual contents.
Thanks!