Towards a Canonical and Structured Representation of PDF Documents through Reverse Engineering