I was expecting something about double-entry accounting and how it can (initially) be non-intuitive whether to mark a given entry as a debit (-100) or credit (+100), especially for liabilities (when you enter a credit card payment in your ledger, the signs aren't what you might intuitively expect when you haven't wrapped your head around the system yet).
Instead I get a nice blog post about a knotty little problem caused by someone's abuse of semantic markup for style purposes. Shows that HTML doesn't have a monopoly on that issue. PDFs are already bad enough as a format for extracting data from, we don't need dumb HTML-inspired tricks on top of that.
Nice job finding the issue and explaining it clearly, OP. Thanks for an interesting writeup.
Why not skipping all the minus signs unless they do not immediately precede a number. I see all the minus signs are after the number, not before, so just skip if they are not before. Preferably in the same BT/ET block as the number.
Some trailing minus signs are true negatives, but it depends on the color which ones are and which ones aren't (and someone got paid for that. Amazing).
That's the problem: what you see and what is actually in the PDF are not the same thing: the actual text in the PDF has a minus sign on both. The entry that you see as "$110.00-" is withdrawal, and its minus sign is very much real. However, the entry that you see as "$1,527.57" also has a minus sign. but you don't see it because it's been given the same font colour as the background.
So we have two identically formatted piece of data, where only the text colour offers a clue on what's real and what isn't. Absolutely insane.
Is it not dangerous to ignore text based on its color? For this particular case it might work but looks like you have a generic service that is supposed to work with any bank statement.
Yes it’s not a safe general rule. I can do it because I first classify the statement into a document_type and then run special code for specific document_types.
It's the ground truth, and it's not like it's more complex than parsing a pdf, at this point, the technology for OCR might even be better than PDF parsing, which is full of accidental instead of natural complexity.
As the article stated, in this specific case the bank used an "invisible" character (a minus sign) to assist in alignment .. which resulted in errors when parsing the text nodes by layout.
OCR of a digital image generated from those PDF draw commands only operates on the visible portions that are rendered.
Another case to be considered .. FOI requests for documents from "pain in the arse" government departments or law firms can result in "digital documents" that are deliberately generated to be difficult to process (go figure) .. eg: filled with circular and spaghetti draw commands that "print correctly" but are a nightmare horror show to parse.
( "technically compliant" is a legally compliant "fuck you" )
This is a great example of why a PDF’s text layer is not necessarily the same as its visible content.
I’d be cautious about making “non-black means invisible” the general rule. A possible middle ground between raw text extraction and full OCR would be to render the page, map each extracted glyph’s bounding box back to the rendered pixels, and discard glyphs that have almost no contrast with their local background.
That would preserve the exact embedded text for visible characters while using rasterization only to determine visibility.
I agree with you, it's not safe to have a rule saying "non-black means invisible".
I should have made it clear in the blog post, but the first thing I do when processing a PDF is I identify the document_type. Then I run special rules for that document_type.
So the "non-black means invisible" rule only applies when processing this document_type.
I was expecting something about double-entry accounting and how it can (initially) be non-intuitive whether to mark a given entry as a debit (-100) or credit (+100), especially for liabilities (when you enter a credit card payment in your ledger, the signs aren't what you might intuitively expect when you haven't wrapped your head around the system yet).
Instead I get a nice blog post about a knotty little problem caused by someone's abuse of semantic markup for style purposes. Shows that HTML doesn't have a monopoly on that issue. PDFs are already bad enough as a format for extracting data from, we don't need dumb HTML-inspired tricks on top of that.
Nice job finding the issue and explaining it clearly, OP. Thanks for an interesting writeup.
I always liked 99.57CR
Why not skipping all the minus signs unless they do not immediately precede a number. I see all the minus signs are after the number, not before, so just skip if they are not before. Preferably in the same BT/ET block as the number.
The trailing minus signs are true negatives.
Some trailing minus signs are true negatives, but it depends on the color which ones are and which ones aren't (and someone got paid for that. Amazing).
That's the problem: what you see and what is actually in the PDF are not the same thing: the actual text in the PDF has a minus sign on both. The entry that you see as "$110.00-" is withdrawal, and its minus sign is very much real. However, the entry that you see as "$1,527.57" also has a minus sign. but you don't see it because it's been given the same font colour as the background.
So we have two identically formatted piece of data, where only the text colour offers a clue on what's real and what isn't. Absolutely insane.
A little disappointed Timecube isn’t somehow referenced. Same vibe.
Does this end on a cliffhanger?
Did they try one method, and what was the feedback?
I did try method number two, and it worked. I should update this post.
Is it not dangerous to ignore text based on its color? For this particular case it might work but looks like you have a generic service that is supposed to work with any bank statement.
Yes it’s not a safe general rule. I can do it because I first classify the statement into a document_type and then run special code for specific document_types.
OCR the PDF.
It's the ground truth, and it's not like it's more complex than parsing a pdf, at this point, the technology for OCR might even be better than PDF parsing, which is full of accidental instead of natural complexity.
I don’t really understand how OCRing a PDF could be more accurate than reading the text nodes.
Another thing to consider is OCR works well for English but not so well for other languages.
As the article stated, in this specific case the bank used an "invisible" character (a minus sign) to assist in alignment .. which resulted in errors when parsing the text nodes by layout.
OCR of a digital image generated from those PDF draw commands only operates on the visible portions that are rendered.
Another case to be considered .. FOI requests for documents from "pain in the arse" government departments or law firms can result in "digital documents" that are deliberately generated to be difficult to process (go figure) .. eg: filled with circular and spaghetti draw commands that "print correctly" but are a nightmare horror show to parse.
( "technically compliant" is a legally compliant "fuck you" )
This is a great example of why a PDF’s text layer is not necessarily the same as its visible content.
I’d be cautious about making “non-black means invisible” the general rule. A possible middle ground between raw text extraction and full OCR would be to render the page, map each extracted glyph’s bounding box back to the rendered pixels, and discard glyphs that have almost no contrast with their local background.
That would preserve the exact embedded text for visible characters while using rasterization only to determine visibility.
Has anyone tried this kind of hybrid approach?
I agree with you, it's not safe to have a rule saying "non-black means invisible".
I should have made it clear in the blog post, but the first thing I do when processing a PDF is I identify the document_type. Then I run special rules for that document_type.
So the "non-black means invisible" rule only applies when processing this document_type.