WordNet

WordNet is a lexical database of English developed at Princeton University. The project began in 1985 under the direction of psychologist George A. Miller with a specific ambition: to organize vocabulary the way it seems to be organized in the mind — not as an alphabetical list, but as a network of concepts. Nouns, verbs, adjectives and adverbs are grouped into sets of synonyms called synsets, each carrying a short definition (the gloss) and usage examples; synsets are linked to one another through labeled semantic and lexical relations. For decades it has been a building block of natural language processing and of AI applied to language.

According to Princeton’s official statistics, version 3.0 contains 155,287 unique words organized into 117,659 synsets, adding up to 206,941 word-sense pairs. The database is freely available for download.

How it is organized

The most frequently encoded relation is hypernymy and its inverse, hyponymy: the “is a kind of” link that arranges nouns into a lexical hierarchy whose branches converge on a single root node, “entity”. An armchair is a kind of chair; a chair, a kind of furniture. On top of that backbone sit meronymy (the part-whole relation: a backrest is part of a chair), troponymy among verbs (whispering is a way of talking) and the antonymy that structures adjectives (wet-dry). That is where it differs from a classic thesaurus, as Princeton’s own website explains: WordNet does not connect word forms but specific senses, and it explicitly labels every relation.

What it has been used for

For years it was the reference semantic resource in natural language processing: word sense disambiguation, semantic similarity measures computed over its hierarchy, and query expansion in information retrieval. The NLTK library ships it as one of its standard corpora. ImageNet, the celebrated image dataset of computer vision, is organized according to the WordNet hierarchy, as its official page states. The model also traveled to other languages: the EuroWordNet I and II projects under the Fourth Framework Programme, active from 1996 to 2000 with Piek Vossen as principal investigator, developed wordnets for eight European languages linked through an Inter-Lingual Index; Open Multilingual Wordnet currently reports 60 wordnets for 49 languages.

WordNet in the age of LLMs

Embeddings and large language models learn semantic relations directly from massive amounts of text, and they now cover much of WordNet’s classic workload. What the Princeton network still offers is different: explicit, inspectable relations curated by lexicographers, valuable wherever it matters to know why two words are related, not just how close their vectors sit. The original is frozen: Princeton’s website notes that, due to funding and staffing issues, there are no plans for new releases. The baton has passed to the community: Open English WordNet describes itself as a fork of Princeton WordNet developed through an open-source methodology and released its 2025 edition on December 31 of that year; the multilingual aggregator continues to add wordnets as they become available. Four decades after it started, Miller’s network is still the structured reference map of meaning in English.