Fonduer has been successfully extended to perform information extraction 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 maintains the structure of the document using a tree data structure.
You'll need to install the Python3 Toolkit:
$ sudo apt install python3-tk
To install this package from PyPi:
$ pip install pdftotree
import pdftotree
pdftotree.parse(pdf_file, html_path=None, model_type=None, model_path=None, favor_figures=True, visualize=False):This is the primary command-line utility provided with this Python package. This takes a PDF file as input, and produces an HTML-like representation of the data:
usage: pdftotree [options] pdf_file
Script to extract tree structure from PDF files. Takes a PDF as input and
outputs an HTML-like representation of the document's structure. By default,
this conversion is done using heuristics. However, a model can be provided as
a parameter to use a machine-learning-based approach.
positional arguments:
pdf_file PDF file name for which tree structure needs to be
extracted
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 where tree structure should be saved. If none,
HTML is printed to stdout.
-f FAVOR_FIGURES, --favor_figures FAVOR_FIGURES
Whether figures must be favored over other parts such
as tables and section headers
-V, --visualize Whether to output visualization images for the tree
-d, --dry-run Run pdftotree, but do not save any output or print to
console.
-v, --verbose Output INFO level logging.
-vv, --veryverbose Output DEBUG level logging.
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.
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 --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.
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 Travis-CI.
To test changes in the package, you install it in editable mode locally in your virtualenv by running:
$ make dev
You will also need flake8 installed locally to run our code-style checks:
$ pip install flake8
Then you can run our tests:
$ make test