Here's the surprising connection: the inventor list on US20200410157A1 includes names long associated with applied-language research at scale, the kind of team that ships features into products used by hundreds of millions of people. That's the tell that a modest-sounding patent is actually production infrastructure.

The method is small and clean. Take a document, run semantic analysis to find entities and concepts worth linking, score candidate targets, and insert the links. The CPC tags, G06F 40/134, 40/166, 40/30, are the document-editing and semantic-analysis classes, not the heavy neural-network classes. This is applied NLP, the unglamorous layer that makes software feel smart.

“A system may use a machine-learned model to determine whether to classify a sequence of one or more words within a first document that is being edited as a candidate hyperlink based at least in part on context associated with the first document.”— U.S. Patent Application 2020/0410157 A1 source

Connect the dots to the present. Today's tools that summarize, cross-reference, and link documents are descendants of methods like this one. The 2020 version did with targeted NLP what a 2025 system might do with a language model, but the user-facing job is the same: read the text, decide what connects to what, act on it.

Why feature it? Because it's a clean example of the gap between hype and IP. Nobody held a keynote for automatic hyperlinking. But it's the kind of durable, shipped capability that quietly compounds, and the patent record is where you find it, not the press cycle.

The house caveat applies: this is a publication, not a grant, so it's a claim of method, not an enforceable right. What it documents is that by the end of 2020, the link-the-document task was mature enough to file. Small idea, long tail.