This is the material and text for my comment on Bischof and Airoldi JASA 2016.

The data tokens just come from the available reuters tokens, parsed as in code/reutersParse.ipynb

The data used here should match those grels files, which are just simple triplet representations.  The hierarchy information is implicit in these matrix representations.  e.g., in the first document E11 is a child of ECAT and both are listed:

E11 2286 1
ECAT 2286 1
M11 2286 1
M12 2286 1
MCAT 2286 1