I've been running Scholarcy through its paces for the last few months, primarily as a research accelerator for my work in attribution modeling and customer journey analytics. The ability to quickly extract key claims and methodologies from academic papers is invaluable. However, I've hit a consistent and fairly disruptive snag when dealing with a specific type of source material.
My issue is with academic PDFs formatted in a standard two-column layout. Scholarcy's highlight feature, which is crucial for me to verify the AI's summary against the original text, seems to completely lose the reading order. When I click on a highlight in the summary card to jump to the source, it very often takes me to the wrong column, or even to a completely different page. The extracted text in the summary itself sometimes appears as a nonsensical jumble of phrases from across the two columns.
Has anyone else in the community encountered this? I'm trying to determine if this is:
* A known limitation with a known workaround (like pre-processing the PDF somehow).
* A bug that's specific to certain PDF generators or publishing houses.
* Just the current state of the technology for this particular document format.
For context, I'm not just talking about complex tables or figures—this happens with plain text paragraphs in standard ACM, IEEE, or other journal templates. It significantly impacts my trust in the extraction for data-heavy sections, which is where I need it most. I find myself constantly having to manually locate the correct text in the PDF, which defeats a lot of the time-saving purpose.
I'd be very curious to hear:
* If others have seen this and what, if anything, you've done to mitigate it.
* Whether the browser extension, Word add-in, or other integration points handle two-column PDFs more reliably.
* Any insights into whether Scholarcy's team has acknowledged this as a priority.
The tool is otherwise fantastic for building a literature matrix, but this particular flaw creates a notable hidden cost in manual verification time for a large segment of academic literature.
—Jen
—Jen
Oh man, I feel your pain on this one. I've been running similar tests with Scholarcy and a few other AI summarizers for my own marketing analytics research, and the two-column PDF layout is basically kryptonite for them. I've seen the exact same thing - highlights jumping to the wrong column or even the wrong page entirely. It's like the parser treats the text flow as a single horizontal strip instead of respecting the column break.
I've got a hunch this is more of a "current state of the tech" issue than a per-publisher bug. The underlying PDF extraction libraries (like PDFMiner or similar) often struggle with multi-column layouts because they don't have a concept of reading order. They just grab text positions and try to reconstruct the sequence, and when columns are close together, the algorithm guesses wrong. I've noticed it's worse with PDFs that have dense footnotes or sidebars too.
One workaround I've stumbled on for my own sanity: if the PDF isn't too locked down, I pre-process it with a tool like OCRmyPDF or even Adobe Acrobat's "export to HTML" option. That often flattens the columns into a single logical text stream. Then I upload that version to Scholarcy. It's a pain, but it reduces the jumble. No guarantee it'll fix every case though.
Are you seeing this across all publishers, or more common with certain ones? I've had better luck with IEEE and Springer PDFs than with, say, older ACM conference proceedings.
If it's not measurable, it's not marketing.