Skip to content

UVA Data Science Capstone project for Internet Archive. This project aimed to classify PDFs as research or non-research documents using an image and text-based approach. For the image-based models, we leveraged CNN transfer learning and used XGBoost for text-based approach.

Notifications You must be signed in to change notification settings

yte9pc/Internet-Archive-PDF-Capstone

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

55 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Internet-Archive-PDF

Problem Description

As new effort in our mission towards “Universal Access to All Knowledge”, the Internet Archive is attempting to collect and provide perpetual access to the “scholarly web”: the public record of research publications and datasets available on the world wide web, interlinked by both hyperlinks (URLs) and citations. We have a specific focus on “long-tail” open access works (which may be published in non-English language, outside North America or Europe, in nonSTEM disciplines, from small or informal publishers, not assigned DOIs, not archived in existing preservation networks, etc).

Implementation and training of a fast PDF identification tool, which can score files on their likelihood of being a research publication, and what stage of publication (eg, abstract, manuscript, camera ready, OCR scan) the file represents. Ideally the tool would process hundreds of millions of files and be network (as opposed to CPU) bound.

Environment

  • Python Version 3.7.4
  • Tensorflow Version 2.3.1
  • Keras Version 2.4.3

How To

To run the text based approach follow these steps

  • First run createDB.py or download googletranslate.db
    • Recommend downloading googletranslate.db
  • Second run multiprocessingMetaText.py
    • python multiprocessingMetaText.py -d "path to input directory of pdfs" -o "path to output directory to store csv" -f "file name to save as (needs to be pkl)" -p "# of processes to use (or let the script decide)"
      • This script will call metaText.py and create a dataset that is used for the model

To run the image based approach follow these steps

  • First run multiprocessingPDFtoImages.py
    • python multiprocessingMetaText.py -d "path to input directory of pdfs" -o "path to output directory to store png" -p "# of processes (or let the script decide)"
      • This script will call pdfToImage.py and create an image of the first page of each PDF
  • Second if a VGG16_V4 model does not exist create a training data set and run cnn.py else skip this step
    • Modify line 52 to point to training set
    • python cnn.py
      • This will train a VGG16 model and save the model
  • Third run multiprocessingCNNPredict.py
    • Modify line 15 in cnnPredict.py to point to saved model
    • python multiprocessingCNNPredict.py -d "path to png files" -o "path to save result" -f "resultName.csv" -p "# of processes (or let the script decide)"
      • This script will call cnnPredict.py and return a csv file with the prediction of each PDF

About

UVA Data Science Capstone project for Internet Archive. This project aimed to classify PDFs as research or non-research documents using an image and text-based approach. For the image-based models, we leveraged CNN transfer learning and used XGBoost for text-based approach.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published