Graph Converter

The Graph Converter is a tool for creating a graph representation out of the content of PDFs. A graph representation can act as the basis for further document processing steps. Geometric relationships are encapsulated. By those, a document structure can be retrieved.

The tool works independent of different document layouts. The graph construction can be controlled via parameter settings mentioned subsequently. Furthermore, layout-based optimizations without the need parameter tweaks are supported using a regression estimation based on document layout characteristics.

The processing of PDF documents is done using the PDFContentConverter library.

How-to

• Pass the path of the PDF file which is wanted to be converted to GraphConverter.
• Call the function convert(). The document graph representations are returned page-wise as a list of networkx graphs.
• Media boxes of a PDF can be accessed using get_media_boxes(), the page count over get_page_count()

Example call:

converter = GraphConverter(pdf)
result = converter.convert()

A file is the only parameter mandatory for a graph construction. Beside the graph conversion, media boxes of a document can be accessed using get_media_boxes() and the page count over get_page_count(). General document layout characteristics are stored in a converter.meta object.

A more detailed example usage is also given in Tester.py.

Example

The following image shows a resulting document graph representation when using the GraphConverter. TODO

Settings

General parameters:

• file: file name
• merge_boxes: indicating if PDF text boxes should be graph nodes, based on visual rectangles present in documents.
• regress_parameters: indicating if graph parameters are regressed or used as a priori optimized default ones.

Edge restrictions:

• use_font: differing font size
• use_width: differing width
• use_rect: nodes contained in differing visual structures
• use_horizontal_overlap: indicating if horizontal edges should be built on overlap. If not, default deltas are used.
• use_vertical_overlap: indicating if vertical edges should be built on overlap. If not, default deltas are used.

Edge thresholds:

• page_ratio_x: maximal relative horizontal distance of two nodes where an edge can be created
• page_ratio_y: maximal relative vertical distance of two nodes where an edge can be created
• x_eps: alignment epsilon for vertical edges in points if use_horizontal_overlap is not enabled
• y_eps: alignment epsilon for horizontal edges in points if use_vertical_overlap is not enabled
• font_eps_h: indicates how much font sizes of nodes are allowed to differ as a constraint for building horizontal edges when use_font is enabled
• font_eps_v: indicates how much font sizes of nodes are allowed to differ as a constraint for building vertical edges when use_font is enabled
• width_pct_eps: relative width difference of nodes as a condition for vertical edges if use_width is enabled
• width_page_eps: indicating at which maximal width of a node the width should act as an edge condition if use_width is enabled

Project Structure

• GraphConverter.py: contains the GraphConverter class for converting documents into graphs.
• util:
  • constants:
  • StorageUtil: store/load functionalities
• Tester.py: Python script for testing the GraphConverter
• pdf: example pdf input files for tests

Output Format

As a result, a list of networkx graphs is returned. Each graph encapsulates a structured representation of a single page.

Edges are attributed with the following features:

• direction: shows the direction of an edge.
  • v: Vertical edge
  • h: Horizontal edge
  • l: Rectangular loop. This represents a novel concept encapsulating structural characteristics of document segments by observing if two different paths end up in the same node.
• length: Scaled length of an edge
• lengthx_phys: Horizontal edge length
• lengthy_phys: Vertical edge length
• weight: Scaled total length

All nodes contain the following content attributes:

• id: unique identifier of the PDF element
• page: page number, starting with 0
• text: text of the PDF element
• x_0: left x coordinate
• x_1: right x coordinate
• y_0: top y coordinate
• y_1: bottom y coordinate
• pos_x: center x coordinate
• pos_y: center y coordinate
• abs_pos: tuple containing a page independent representation of (pos_x,pos_y) coordinates
• original_font: font as extracted by pdfminer
• font_name: name of the font extracted from original_font
• code: font code as provided by pdfminer
• bold: factor 1 indicating that a text is bold and 0 otherwise
• italic: factor 1 indicating that a text is italic and 0 otherwise
• font_size: size of the text in points
• masked: text with numeric content substituted as #
• frequency_hist: histogram of character type frequencies in a text, stored as a tuple containing percentages of textual, numerical, text symbolic and other symbols
• len_text: number of characters
• n_tokens: number of words
• tag: tag for key-value pair extractions, indicating keys or values based on simple heuristics
• box: box extracted by pdfminer Layout Analysis
• in_element_ids: contains IDs of surrounding visual elements such as rectangles or lists. They are stored as a list [left, right, top, bottom]. -1 is indicating that there is no adjacent visual element.
• in_element: indicates based on in_element_ids whether an element is stored in a visual rectangle representation (stored as "rectangle") or not (stored as "none").

The media boxes possess the following entries in a dictionary:

• x0: Left x page crop box coordinate
• x1: Right x page crop box coordinate
• y0: Top y page crop box coordinate
• y1: Bottom y page crop box coordinate
• x0page: Left x page coordinate
• x1page: Right x page coordinate
• y0page: Top y page coordinate
• y1page: Bottom y page coordinate

Future Work

• The GraphConverter will be extended using OCR processing for images in order to support more unstructured types than solely PDFs.

Acknowledgements

• Example PDFs are obtained from the ICDAR Table Recognition Challenge 2013 https://roundtrippdf.com/en/data-extraction/pdf-table-recognition-dataset/.

Authors

• Michael Benedikt Aigner
• Florian Preis