Tag: nlp

When the Defense Technology Information Center needed 1,400 new thesaurus terms filed against 272 fixed categories (Field & Group) our approach accomplished more than a person could have in a year. A single WordNet run made roughly 380,800 semantic comparisons in about four hours, collapsing one to two man-years of manual judgment into an afternoon plus a focused human review. Here is how it worked, what it saved, and where it goes next.

Daydreamer generates associations without being prompted. Given a lexical model of English — WordNet, VerbNet, SentiWordNet, and the NRC Emotion Lexicon — it draws links between ideas, uses the oppositions among them as a signal, and uncovers the concepts that unify them: concepts its input never named.

The complete DOJ Epstein release is 1,378,652 documents. Instead of reading headlines about them, we indexed, embedded, and machine-read the entire corpus — full-text search, semantic vectors, and computer vision — measuring 46 topics, 48 named individuals, and every image, to establish what is actually in the files and what is not.