pdftotree

WARNING: pdftotree is experimental code and is NOT stable. It is not integrated with or supported by Fonduer.

Fonduer performs knowledge base construction from richly formatted data such as tables. A crucial step in this process is the construction of the hierarchical tree of context objects such as text blocks, figures, tables, etc. The system currently uses PDF to HTML conversion provided by Adobe Acrobat. However, Adobe Acrobat is not an open source tool, which may be inconvenient for Fonduer users.

This package is the result of building our own module as replacement to Adobe Acrobat. Several open source tools are available for pdf to html conversion but these tools do not preserve the cell structure in a table. Our goal in this project is to develop a tool that extracts text, figures and tables in a pdf document and returns them in an easily consumable format.

Up to v0.4.1, pdftotree's output was formatted in its own "HTML-like" format. From v0.5.0, it conforms to hOCR, an open-standard format for OCR results.

Dependencies

pdftotree depends on the following native libraries:

• ImageMagick 6+ (for Wand)
• Java 8+ (for tabula-py)

Installation

To install this package from PyPi:

$ pip install pdftotree

Usage

pdftotree as a Python package

pdftotree

This is the primary command-line utility provided with this Python package. This takes a PDF file as input and produces an hOCR file as output:

usage: pdftotree [options] pdf_file

Convert PDF into hOCR.

positional arguments:
  pdf_file              Path to input PDF file.

optional arguments:
  -h, --help            show this help message and exit
  -mt {vision,ml,None}, --model_type {vision,ml,None}
                        Model type to use. None (default) for heuristics
                        approach.
  -m MODEL_PATH, --model_path MODEL_PATH
                        Pretrained model, generated by extract_tables tool
  -o OUTPUT, --output OUTPUT
                        Path to output hOCR file. If not given, it will be
                        printed to stdout.
  -V, --visualize       Whether to output visualization images
  -v, --verbose         Output INFO level logging.
  -vv, --veryverbose    Output DEBUG level logging. Use this if tabula should not be silent.

extract_tables

This tool trains a machine-learning model to extract tables. The output model can be used as an input to pdftotree:

usage: extract_tables [-h] [--mode MODE] --model-path MODEL_PATH
                      [--train-pdf TRAIN_PDF] --test-pdf TEST_PDF
                      [--gt-train GT_TRAIN] --gt-test GT_TEST --datapath
                      DATAPATH [--iou-thresh IOU_THRESH] [-v] [-vv]

Script to extract tables bounding boxes from PDF files using machine learning.
If `model.pkl` is saved in the model-path, the pickled model will be used for
prediction. Otherwise the model will be retrained. If --mode is test (by
default), the script will create a .bbox file containing the tables for the
pdf documents listed in the file --test-pdf. If --mode is dev, the script will
also extract ground truth labels for the test data and compute statistics.

optional arguments:
  -h, --help            show this help message and exit
  --mode MODE           Usage mode dev or test, default is test
  --model-path MODEL_PATH
                        Path to the model. If the file exists, it will be
                        used. Otherwise, a new model will be trained.
  --train-pdf TRAIN_PDF
                        List of pdf file names used for training. These files
                        must be saved in the --datapath directory. Required if
                        no pretrained model is provided.
  --test-pdf TEST_PDF   List of pdf file names used for testing. These files
                        must be saved in the --datapath directory.
  --gt-train GT_TRAIN   Ground truth train tables. Required if no pretrained
                        model is provided.
  --gt-test GT_TEST     Ground truth test tables.
  --datapath DATAPATH   Path to directory containing the input documents.
  --iou-thresh IOU_THRESH
                        Intersection over union threshold to remove duplicate
                        tables
  -v                    Output INFO level logging
  -vv                   Output DEBUG level logging

PDF List Format

The list of PDFs are simply a single filename on each line. For example:

1-s2.0-S000925411100369X-main.pdf
1-s2.0-S0009254115301030-main.pdf
1-s2.0-S0012821X12005717-main.pdf
1-s2.0-S0012821X15007487-main.pdf
1-s2.0-S0016699515000601-main.pdf

Ground Truth File Format

The ground truth is formatted to mirror the PDF List. That is, the first line of the ground truth file provides the labels for the first document in corresponding PDF list. Labels take the form of semicolon-separated tuples containing the values (page_num, page_width, page_height, top, left, bottom, right). For example:

(10, 696, 951, 634, 366, 832, 653);(14, 696, 951, 720, 62, 819, 654);(4, 696, 951, 152, 66, 813, 654);(7, 696, 951, 415, 57, 833, 647);(8, 696, 951, 163, 370, 563, 652)
(11, 713, 951, 97, 47, 204, 676);(11, 713, 951, 261, 45, 357, 673);(3, 713, 951, 110, 44, 355, 676);(8, 713, 951, 763, 55, 903, 687)
(5, 672, 951, 88, 57, 203, 578);(5, 672, 951, 593, 60, 696, 579)
(5, 718, 951, 131, 382, 403, 677)
(13, 713, 951, 119, 56, 175, 364);(13, 713, 951, 844, 57, 902, 363);(14, 713, 951, 109, 365, 164, 671);(8, 713, 951, 663, 46, 890, 672)

One method to label these tables is to use DocumentAnnotation, which allows you to select table regions in your web browser and produces the bounding box file.

Example Dataset: Paleontological Papers

A full set of documents and ground truth labels can be downloaded here: PaleoDocs. You can train a machine-learning model to extract table regions by downloading this dataset and extracting it into a directory named data and then running the command below. Double check that the paths in the command match wherever you have downloaded the data:

$ extract_tables -v --train-pdf data/paleo/ml/train.pdf.list.paleo.not.scanned --gt-train data/paleo/ml/gt.train --test-pdf data/paleo/ml/test.pdf.list.paleo.not.scanned --gt-test data/paleo/ml/gt.test --datapath data/paleo/documents/ --model-path data/model.pkl

The resulting model of this example command would be saved as data/model.pkl.

For Developers

We are following Semantic Versioning 2.0.0 conventions. The maintainers will create a git tag for each release and increment the version number found in the version file accordingly. We deploy tags to PyPI automatically using GitHub Actions.

Tests

To test changes in the package, you install it in editable mode locally in your virtualenv by running:

$ make dev

This will also install all the tools we use to enforce code-style.

Then you can run our tests:

$ make test

Release

Follow the below steps to release

1. Make commits with the following changes:

   1. Update the CHANGELOG
   2. Change the version at pdftotree/_version.py to 0.X.Y.

2. Submit the commits as a pull-request
3. Once the pull-request is merged, add a tag v0.X.Y (don't forget "v" at the beginning) and push it

4. Pushing the tag triggers GitHub Actions workflow that

   1. Creates a pre-release on GitHub
   2. Publishes a package to PyPI

5. Edit the pre-release and release it
6. Increment the version to 0.X.(Y+1)+dev