NELL (Never-Ending Language Learner)

NELL runs as a continuous experiment: it reads web pages, proposes new facts about entities and how they relate, and uses what it already knows to judge and improve the next round of extractions. Over years of running from 2010 onward, this has grown into a large store of candidate beliefs, organised into categories (this thing is a city, a company, an athlete) and relations (this company is headquartered in this city), each with a confidence score reflecting how sure the system is.

The knowledge base is published for download from Carnegie Mellon's site, typically as tab-separated files that list each belief along with its confidence, the iteration in which it was learned, and the textual patterns that produced it. That provenance is genuinely useful: you can trace where a fact came from, and you can filter aggressively by confidence to keep only the beliefs the system is most sure about before turning them into text for retrieval.

For a RAG project, NELL is best treated as raw web-derived facts and as a case study rather than a polished knowledge source. If you want a broad, messy set of entity-and-relation statements to experiment with, or you are researching how structured knowledge can be learned from text at scale, it is a rich resource. Flattening high-confidence beliefs into short sentences and embedding them can seed an entity-aware system.

The caveats are significant. Because the beliefs come from automatic reading rather than human curation, quality varies a lot, and low-confidence rows include plenty of errors and nonsense, so the confidence scores are not optional. The data is offered for research use, so commercial reuse is restricted. And the project has been largely static for years, so the facts reflect the web of the 2010s, not now. For clean, current, human-curated knowledge, Wikidata, DBpedia, and ConceptNet are the better foundations; NELL's value is its scale, its provenance, and its honesty about uncertainty.