yes, got me there.

but i feel in the context (communication/meta-data inference) that is missing the trees for the forest

On the other hand, useful features should not be ignored, because other, almost unrelated things are hard.