Identifying and Extracting Hierarchical Information from Business PDF Documents

Portable Document Format (PDF) is a popular choice for a secure communication and persistence of business information and is a universally accepted format by businesses choosing to become digital. PDF provides multiple ways to make the information visually appealing and readable, and device independent rendering. To achieve this, PDF stores metadata with individual text characters, graphic components and other layout elements. Such atomic component wise meta-data makes machine processing of information in the PDF format very challenging; the challenge is further extended due to the difficulty of stitching together the original semantics from the componentized information. We propose a generic approach for extracting the hierarchy of the document structure while separating the content from header and footer, and extracting metadata associated with checkboxes to annotate the business information contained in PDF for tasks like mining specifications and rules from the document. Our prototype is able to process real-life, large PDF documents each running into roughly 400 pages, with nearly 95% of the extraction requiring no human intervention.