Merriam-Webster and Unstructured Data Processing (www.georgeho.org)

Kory Stamper’s account of how Merriam‑Webster is made — skimming magazines and web text to “read and mark” interesting usages, augmenting that with corpora, then having editors manually revise or write definitions — reads like a blueprint for successful unstructured‑data projects. The workflow is threefold: collect and curate raw text, structure it through focused human effort (editors typically spend ~15 minutes per word), and expose ancillary datasets or features (etymologies, pronunciations, usage dates) that amplify product value. Notably, much of the highest-value work is low‑tech and human‑intensive (index cards still appear), underscoring that better code or fancier ML doesn’t automatically produce more value. For the AI/ML community the lessons are practical: invest in human‑in‑the‑loop curation and simple structuring pipelines before chasing modeling complexity; treat structured corpora as both supplements and starting points; and think of datasets as products that can pivot toward high‑value ancillary services. Real‑world examples like Google Search or a cryptic‑crossword dataset mirror the pattern: large raw collection, pragmatic structuring, then productized features on top. The takeaway: structuring is itself the core value proposition, and clever downstream features — not necessarily better models — often determine product‑market fit.