Compare Python PDF extraction libraries with sample files

[mgorenstein]

Write up mini-paper comparing performance of various text-extractors on a document with available plaintext (possibly a particular edition of the bible).

• [ ] Find popular samples with clean and accurate plain text.
  • [ ] For each sample, find versions of varying quality.
  • One version should be a PDF I generate from the plain text.
  • One version should be a rich PDF where text is queryable.
  • One version should require OCR.
  • [ ] Trim extra info like page number and Project Gutenberg header, if possible.
  • An approach could be to locate the first and last sentence of the text, consider only between these.
  • Or, just leave it be? All extractors will pick up this noise and it's expected in typical use cases.
• [ ] Determine which PDF extractor libraries to test.
  • Definitely PDFMiner since I've already reverse-engineered it programatically.
• [ ] Determine the measures of extraction accuracy.
  • I've used various measures of string difference in the FuzzyWuzzy library with some success.
  • Measure speed.
  • Many of these libraries do layout analysis--is this helpful or not? Surely has an effect on speed.
• [ ] Run the conversions and calculate accuracy.
  • [ ] Create a testing suite that determines setting regimes with uniformly better accuracy and use those settings as the benchmark for a particular library.

[grahamsack]

Hi Mark -- I experimented a bit with PyPDF and PDFMiner for syllabus extraction. PyPDF seemed to be smoother to work with. I tried extracting syllabi into plain text and html. The extractors captured most of the text correctly but, in cases where the formatting was complicated or had lots of tables, it jumbled the order.

Best,

Graham

On Thu, May 22, 2014 at 12:41 PM, Mark Gorenstein notifications@github.comwrote:

Write up mini-paper comparing performance of various text-extractors on a document with available plaintext (possibly a particular edition of the bible).

— Reply to this email directly or view it on GitHubhttps://github.com/dhcolumbia/opensyllabus/issues/30 .

[denten]

We are planning to do a more formal comparison. Stay tuned.

[grahamsack]

If you want to leverage it, I put my code for the extractor in the opensyllabus/Classifiers folder.

[mgorenstein]

Thanks, Graham.

[mgorenstein]

Libraries

• [x] PDFMiner
  • Slate seems like a good wrapper for it, though is possibly outdated.
  • Tutorial for annotation extraction in case we decide to go this route.
• [x] pyPDF2
• [x] PDFBox
• [x] PDFTextStream
  • Worth checking out. Some sites claim Python support, but if not the Java version is still a tenable option.
  • Just use os.system() to execute through Python.
• [ ] GFX
  • Requires installation of SWF Tools
• [ ] pdfextract
  • two modes--line preservation, no preservation
• [x] xpdf

Source Texts

• Pride and Prejudice
  • Project Gutenberg plain text
  • Project Gutenberg PDF generated from plain text
  • Planet Books PDF
  • Google scan of old edition

[mgorenstein]

I'm going to move ahead with V1 given these resources. I'll make the platform flexible enough to support the addition of other PDF extractors in case we come across any serious contenders that I've missed.

Graham and Dennis: let me know if you have any suggestions, especially with the selection of source texts. I went with P&P because it's in the public domain, was written in English, and has a range of released PDFs.

[grahamsack]

I had read about slate while looking into PDFMiner and I thought it sounded very good and comparatively user-friendly, but I wasn't able to get it working due to a dependency issue I was never able to resolve. If you can it working, that's great as it sounds like a good library.

On Sun, May 25, 2014 at 1:29 PM, Mark Gorenstein notifications@github.comwrote:

I'm going to move ahead with V1 given these resources. I'll make the platform flexible enough to support the addition of other PDF extractors in case we come across any serious contenders that I've missed.

Graham and Dennis: let me know if you have any suggestions, especially with the selection of source texts. I went with P&P because it's in the public domain, was written in English, and has a range of released PDFs.

— Reply to this email directly or view it on GitHubhttps://github.com/dhcolumbia/opensyllabus/issues/30#issuecomment-44140539 .

[mrenoch]

This could be worth checking out, to unify and maybe simplify text extraction:

http://datascopeanalytics.com/what-we-think/2014/07/27/extract-text-from-any-document-no-muss-no-fuss

https://github.com/deanmalmgren/textract

[samzhang111]

Jumping in after not contributing very much... I'm familiar with some of the people who are maintaining Apache Tika out at NASA JPL. It is a project that has a strong core team of developers, and has overlapping goals with textract. The advantage to Tika (and textract) is that you don't need separate logic for each document format, and you also get standard metadata for each document.

Tika wraps around pdfbox for pdf documents, which performed a 6 second extraction in the benchmarking stats file. I bet the slowness was caused by the bootup time of the JVM, though. If you separate the JVM initialization code with the conversion, I imagine it's more in the range of the pure-python extractors. This is how I've used Tika in python in the past: www.hackzine.org/using-apache-tika-from-python-with-jnius.html

Cheers, Sam

[chrismattmann]

Thanks @samzhang111 yep, happy to provide any info here on Tika if it helps

[chrismattmann]

Coming back here, just FYI we have a fully supported Tika port to Python using the JAX-RS REST server. FYI: https://github.com/chrismattmann/tika-python