grep isn’t really designed as a natural language search tool
My understanding of GREP history is that Ken Thompson created grep to do some textual analysis on The Federalist Papers, which to me sounds like it was designed for processing natural language. But it was on a PDP-11 which had resource constraints. Lines of text would be more uniform to manage than sentences given limited resources of the 1970s.
Thanks for the PERL code. Though I might favor sed or awk for that job. Of course that also means complicating emacs’ grep mode facility. And for PDFs I guess I’d opt for pdfgrep’s limitations over doing a text extraction on every PDF.