If we do this before the parsing we can leverage that semantics in the matching phase.
as_property
Probably this can be improved upon if the data is later analyzed with it's similars.
tree_nlu