If identifying text structure in PDF documents is so difficult, how do PDF readers do it so well?

dave walker picture dave walker · Mar 27, 2014 · Viewed 18.1k times · Source

I have been trying to write a simple console application or PowerShell script to extract the text from a large number of PDF documents. There are several libraries and CLI tools that offer to do this, but it turns out that none are able to reliably identify document structure. In particular I am concerned with the recognition of text columns. Even the very expensive PDFLib TET tool frequently jumbles the content of two adjacent columns of text.

It is frequently noted that the PDF format does not have any concept of columns, or even words. Several answers to similar questions on SO mention this. The problem is so great that it even warrants academic research. This journal article notes:

All data objects in a PDF file are represented in a visually-oriented way, as a sequence of operators which...generally do not convey information about higher level text units such as tokens, lines, or columns—information about boundaries between such units is only available implicitly through whitespace

Hence, all extraction tools I have tried (iTextSharp, PDFLib TET, and Python PDFMiner) have failed to recognize text column boundaries. Of these tools, PDFLib TET performs best.

However, SumatraPDF, the very lightweight and open source PDF Reader, and many others like it can identify columns and text areas perfectly. If I open a document in one of these applications, select all the text on a page (or even the entire document with CTRL+A) copy and paste it into a text file, the text is rendered in the correct order almost flawlessly. It occasionally mixes the footer and header text into one of the columns.

So my question is, how can these applications do what is seemingly so difficult (even for the expensive tools like PDFLib)?

EDIT 31 March 2014: For what it's worth I have found that PDFBox is much better at text extraction than iTextSharp (notwithstanding a bespoke Strategy implementation) and PDFLib TET is slightly better than PDFBox, but it's quite expensive. Python PDFMiner is hopeless. The best results I have seen come from Google. One can upload PDFs (2GB at a time) to Google Drive and then download them as text. This is what I am doing. I have written a small utility that splits my PDFs into 10 page files (Google will only convert the first 10 pages) and then stitches them back together once downloaded.

EDIT 7 April 2014. Cancel my last. The best extraction is achieved by MS Word. And this can be automated in Acrobat Pro (Tools > Action Wizard > Create New Action). Word to text can be automated using the .NET OpenXml library. Here is a class that will do the extraction (docx to txt) very neatly. My initial testing finds that the MS Word conversion is considerably more accurate with regard to document structure, but this is not so important once converted to plain text.

Answer

David van Driessche picture David van Driessche · Mar 27, 2014

I once wrote an algorithm that did exactly what you mentioned for a PDF editor product that is still the number one PDF editor used today. There are a couple of reasons for what you mention (I think) but the important one is focus.

You are correct that PDF (usually) doesn't contain any structure information. PDF is interested in the visual representation of a page, not necessarily in what the page "means". This means in its purest form it doesn't need information about lines, paragraphs, columns or anything like that. Actually, it doesn't even need information about the text itself and there are plenty of PDF files where you can't even copy and paste the text without ending up with gibberish.

So if you want to be able to extract formatted text, you have to indeed look at all of the pieces of text on the page, perhaps taking some of the line-art information into account as well, and you have to piece them back together. Usually that happens by writing an engine that looks at white-space and then decides first what are lines, what are paragraphs and so on. Tables are notoriously difficult for example because they are so diverse.

Alternative strategies could be to:

  • Look at some of the structure information that is available in some PDF files. Some PDF/A files and all PDF/UA files (PDF for archival and PDF for Universal Accessibility) must have structure information that can very well be used to retrieve structure. Other PDF files may have that information as well.
  • Look at the creator of the PDF document and have specific algorithms to handle those PDFs well. If you know you're only interested in Word or if you know that 99% of the PDFs you will ever handle will come out of Word 2011, it might be worth using that knowledge.

So why are some products better at this than others? Focus I guess. The PDF specification is very broad, and some tools focus more on lower-level PDF tasks, some more on higher-level PDF tasks. Some are oriented towards "office" use - some towards "graphic arts" use. Depending on your focus you may decide that a certain feature is worth a lot of attention or not.

Additionally, and that may seem like a lousy answer, but I believe it's actually true, this is an algorithmically difficult problem and it takes only one genius developer to implement an algorithm that is much better than the average product on the market. It's one of those areas where - if you are clever and you have enough focus to put some of your attention on it, and especially if you have a good idea what the target market is you are writing this for - you'll get it right, while everybody else will get it mediocre.

(And no, I didn't get it right back then when I was writing that code - we never had enough focus to follow-through and make something that was really good)