So proper old-school AI then? Perl and regexes?
(Have you _seen_ the internals of Perl's regex code? I swear it has more inexplicable random interconnectedness than most neural nets...)
(which is random, but you need two columns in the PDF). Try select from column 1 near the bottom, across to column 2.
PDFs are very diverse. Our algorithm does not always work, alas, but it is significantly better than what existed before.
In fact, for some reason, it works well for this document on page 1 and 3 but not so well on page 2! But I no longer have access to the low level tools I used to use to figure out what went wrong.
PDFs are more "baked" layouts rather than high level semantic markup, so it's not easy to automatically determine this. The AI is probably a set of heuristics.
Just a bit of an aside. I'm still somewhat bitter about it :-p
This is quite similar to iOS text selection in practice: mostly works, but when it doesn't you're screwed.
Unfortunately it doesn't work correctly for the PDFs I get from my accountant.
It's one of those features that are really obvious, but probably surprisingly hard to implement.
My colleague and I (we co-developed the "little bit of AI") never saw this comment (and the applause) live: we were still lining up outside of Moscone West. As Apple employees we weren't allowed in until all the guests and paying attendees had entered. Still, we watched the video afterwards and got a huge kick out of the mention.
BTW, the AI we added did not use neural nets or deep learning.