My first thought is to think of this as a tree structure.
Not that my first thoughts are always right ...
Suppose you have a hash of root tree nodes. The key to the
root node hash is one of your matches. Each node is itself
a hash, containing several things: The line number, a
(possibly undefined) reference to the next node in the
tree, and anything else you may need. As you parse each
element (each line), you take each match of that element
and look it up by starting at the top of the tree: You do
a lookup in the root node hash. If you find a hit, you
follow the next node references down to the end, and add a
new node. If you don't find a hit, you add a new element
to the root hash. I believe this could also be optmized by
keeping a last node hash. To get your relationships, you
traverse each root node down when you're done.
Update: Actually, I'm not sure this truly handles all
the relationships that elegantly. At a minimum, each node
would have to be an array of other (sub)nodes to get all
the relationships. It seems like there should be a simpler
way to do that.