WordNet

WordNet is a large lexical database of English developed at Princeton University beginning in 1985. It groups nouns, verbs, adjectives, and adverbs into sets of synonyms called synsets, each expressing a distinct concept, and connects those synsets through labeled semantic relations such as hypernymy (is-a), meronymy (part-whole), and antonymy ^[1]. The result is a machine-readable network of word meanings rather than word forms: the word "bank" appears in separate synsets for the financial institution and the riverside, and each sense has its own definition, example sentences, and position in the network. WordNet 3.0, released in 2006, contains 155,287 words organized into 117,659 synsets ^[2].

The project was started by psychologist George A. Miller, one of the founders of cognitive psychology, and was designed to test psycholinguistic theories of how humans store and retrieve word meanings ^[3][4]. It became something its creators did not initially plan: the default lexical resource of classic natural language processing. For roughly two decades, word-sense disambiguation research used WordNet as its standard sense inventory ^[5], semantic similarity between words was computed as distance in the WordNet graph ^[6], and when computer vision researchers needed a category system for a giant image dataset, they built ImageNet directly on top of WordNet's noun hierarchy ^[7].

Princeton stopped releasing new versions after WordNet 3.1 in 2011, citing staffing and funding constraints ^[1][8]. Since 2019 the database has been maintained as the community-run Open English WordNet, which publishes updated editions roughly annually on GitHub ^[9][10]. Although learned representations such as word embeddings and large language models have displaced hand-built lexicons in most production NLP, WordNet remains in active use for evaluation, linguistic research, and as the hub that links wordnets in dozens of other languages ^[10][11].

Origins at Princeton

In 1985, a group of psychologists and linguists at Princeton University's Cognitive Science Laboratory began building a lexical database informed by psycholinguistic research on the mental lexicon. The initial idea was modest, an aid for searching dictionaries conceptually rather than alphabetically, but the project soon developed its own principles and goals ^[3]. Miller, then a Princeton psychology professor best known for the 1956 paper "The Magical Number Seven, Plus or Minus Two," directed the effort; he received the National Medal of Science in 1991 and died in 2012 ^[4]. Christiane Fellbaum, a co-developer of the database from its early years, later directed the project ^[8]. The work was funded by U.S. government research agencies, initially the Office of Naval Research and later DARPA and the National Science Foundation, among others ^[8].

The design premise was that the mental lexicon is organized by meaning, not spelling. Standard dictionaries scatter related concepts across the alphabet; WordNet instead makes the semantic relations among concepts explicit and navigable. Miller's team argued that WordNet "instantiates hypotheses based on results of psycholinguistic research" and could be read as a dictionary built on psycholinguistic principles ^[3]. By 1990 the database already held roughly 95,600 word forms organized into about 70,100 synsets ^[3].

Two publications anchored WordNet in the research literature: Miller's 1995 overview in Communications of the ACM ^[12] and the 1998 MIT Press volume WordNet: An Electronic Lexical Database, edited by Fellbaum ^[13].

How WordNet is organized

The basic unit is the synset: an unordered set of words that are interchangeable in some context, such as {shut, close} or {car, automobile}. Each synset carries a short definition (a "gloss") and, in most cases, example sentences. A word form with several meanings appears in one synset per meaning, so every word-sense pair is unique ^[1]. Because the four parts of speech have different semantic organization, WordNet is really four sub-networks, one each for nouns, verbs, adjectives, and adverbs, with a smaller number of cross-part-of-speech pointers ^[1].

Synsets are connected by a small inventory of relations ^[1][3]:

Scroll sideways for more →

The hypernym relation gives nouns a tree-like taxonomy. Every noun hierarchy ultimately leads up to the single root synset {entity}, and the relation is transitive: an armchair is a kind of chair, a chair is a kind of furniture, so an armchair is a kind of furniture ^[1]. The hierarchies are shallow by design; Miller reported that they seldom exceed about a dozen levels, with coverage divided among twenty-five topical "lexicographer files" for nouns ^[3]. WordNet also distinguishes types from instances: armchair is a type of chair, while specific people, countries, and geographic entities are instances and always sit at the leaves of their hierarchies ^[1].

This structure is what separates WordNet from a thesaurus. A thesaurus groups word forms by loose similarity; WordNet links specific senses and labels every relation, so software can traverse the graph and reason about it ^[1].

Size, versions, and license

Official statistics for WordNet 3.0, the last major release, break down as follows ^[2]:

Scroll sideways for more →

Most words are monosemous (128,391 of them have exactly one sense), while 26,896 polysemous words account for 79,450 senses ^[2]. WordNet 3.0 came out in 2006 ^[9][8]; a 3.1 update followed in 2011 and was the last Princeton release ^[8]. The project's website states that, due to limited staffing, there are no plans for future releases and the team can no longer accept comments and suggestions ^[1].

WordNet's license is one reason for its ubiquity. Princeton distributes the database and software under a short BSD-style license that permits use, copying, modification, and redistribution for any purpose, including commercial applications, without fee, requiring only that the copyright notice be preserved and that Princeton's name not be used in advertising ^[14]. This let WordNet be embedded in commercial software, search engines, and downstream resources without negotiation, decades before "open data" became a common practice.

Open English WordNet

With Princeton's releases frozen, a group of researchers forked the database in 2019 as English WordNet (now Open English WordNet), developed in the open on GitHub. The founding paper, by John P. McCrae, Alexandre Rademaker, Francis Bond, Ewa Rudnicka, and Fellbaum herself, noted that despite WordNet being one of the most widely used language resources in NLP, there had been only one update since version 3.0 in 2006, and proposed an open-source quality-assurance model in which anyone can propose fixes, with the project aiming to be a drop-in replacement for Princeton WordNet ^[9].

Editions have appeared roughly annually since then ^[10]:

Scroll sideways for more →

Starting with the 2025 edition, the project split proper nouns into a companion resource, Open English Namenet, derived largely from Wikidata; the core 2025 wordnet of common words contains 135,969 words in 107,519 synsets, while the "Plus" version adds manually validated proper nouns ^[10]. Open English WordNet is licensed under CC BY 4.0 on top of the original WordNet license, is browsable at en-word.net, and is accessible programmatically through the wn Python library ^[10][15]. Alignment with earlier versions and with wordnets in other languages is maintained through the Collaborative Interlingual Index (CILI) ^[10].

Uses in classic NLP

WordNet became the workhorse of meaning-related tasks in pre-neural NLP. It has been described as the most commonly used computational lexicon of English for word-sense disambiguation (WSD), the task of picking the contextually correct sense of an ambiguous word ^[5][8]. Sense-annotated corpora and shared evaluation tasks standardized on WordNet's sense inventory, which made results comparable across systems but also exposed a recurring criticism: WordNet's sense distinctions are unusually fine-grained, sometimes finer than human annotators can reliably distinguish, which caps inter-annotator agreement and system performance alike ^[8].

The taxonomy also supports quantitative measures of semantic similarity. Path-based measures score two senses by the shortest path between them in the hypernym graph; refinements such as Wu-Palmer and Leacock-Chodorow similarity factor in taxonomy depth. These measures ship in the NLTK toolkit's WordNet corpus reader, alongside lookup of synsets, hypernyms, and the other WordNet relations ^[6]. Beyond WSD and similarity, WordNet has been used for query expansion in information retrieval, text classification, summarization, and machine translation ^[8]. Derived resources tagged additional information onto its synsets, including SentiWordNet, which scored every WordNet 3.0 synset for positivity and negativity to support sentiment analysis ^[8].

ImageNet and computer vision

WordNet's most consequential downstream use arguably happened outside NLP. ImageNet, the image dataset that catalyzed the deep learning era, is explicitly "an image dataset organized according to the WordNet hierarchy": each category is a WordNet synset, and the project set out to attach on the order of 1,000 photographs to each of tens of thousands of noun synsets ^[7]. The 2009 paper introducing ImageNet, from Fei-Fei Li's group, describes building the dataset on WordNet's backbone ^[16]; the full ImageNet grew to 14,197,122 images indexed under 21,841 synsets ^[7].

The ImageNet Large Scale Visual Recognition Challenge (ILSVRC), run annually from 2010, used a 1,000-category subset of these synsets for its classification task ^[17][18]. Those 1,000 WordNet synsets are the classes that AlexNet learned to recognize in 2012, and they remain the label set behind countless pretrained convolutional neural network checkpoints. A lexicon hand-built by psycholinguists in the 1980s thus quietly defined the category system for a decade of computer vision benchmarks.

Multilingual wordnets

WordNet's design proved portable across languages. EuroWordNet, an EU-funded project that ran from March 1996 to June 1999, built wordnets for Dutch, Italian, Spanish, German, French, Czech, and Estonian, structured like the Princeton original and connected through an Inter-Lingual Index so that a concept in one language could be mapped to its counterparts in the others ^[19]. The Global WordNet Association, whose co-founders and co-presidents are Piek Vossen and Fellbaum ^[11], has coordinated this ecosystem since the first Global WordNet Conference, held in Mysore, India, in January 2002, and maintains a directory of wordnet projects covering dozens of languages ^[21]. The Open Multilingual Wordnet distributes 60 openly licensed wordnets covering 49 languages in a common format, accessible through the same NLTK interface as English WordNet ^[20][6].

WordNet in the LLM era

WordNet's centrality in NLP declined as learned representations took over. Distributional models such as word2vec and GloVe derive word similarity from raw text statistics rather than curated graphs, and contextual models from BERT onward handle sense ambiguity implicitly, without consulting a sense inventory. The symptoms of the shift were visible at Princeton itself: the funding and staffing that once sustained regular releases dried up after 2011 ^[1], while the field's benchmarks moved from sense-labeled corpora toward end-to-end tasks.

But the database did not become irrelevant; it changed roles. Open English WordNet continues to publish yearly editions, and its 2025 release added a Wikidata-derived name inventory, effectively bridging the classic lexicon with modern knowledge graphs ^[10]. WordNet remains standard equipment for linguistic annotation, cross-lingual linking through CILI, WSD evaluation, and the many wordnets in other languages that inherit its structure ^[10][20]. It also persists inside the plumbing of machine learning: every model evaluated on ImageNet classes is, at one remove, using WordNet synsets. Like Cyc, WordNet is hand-curated knowledge infrastructure from the symbolic AI era that outlived the research program it was built for; its permissive license and simple file format helped it get absorbed into the tooling of later paradigms, which is why a forty-year-old psycholinguistics project still ships in Python packages today ^[6][14].

See also

• ImageNet
• Natural language processing
• Word embedding
• Knowledge graph
• Word2vec
• Cyc

References

1. ^Princeton University. "What is WordNet?" WordNet: A Lexical Database for English. wordnet.princeton.edu
2. ^Princeton University. "wnstats(7WN): WordNet 3.0 database statistics." WordNet documentation. wordnet.princeton.edu/...wnstats7wn
3. ^Miller, George A.; Beckwith, Richard; Fellbaum, Christiane; Gross, Derek; Miller, Katherine. "Introduction to WordNet: An On-line Lexical Database" (revised August 1993). Princeton University. wordnetcode.princeton.edu/5papers.pdf
4. ^Wikipedia. "George Armitage Miller." en.wikipedia.org/...George_Armitage_Miller
5. ^Navigli, Roberto. "Word Sense Disambiguation: A Survey." ACM Computing Surveys 41(2), 2009. dl.acm.org/...1459352.1459355
6. ^NLTK Project. "Sample usage for wordnet." NLTK documentation. nltk.org/...wordnet
7. ^ImageNet. "About ImageNet." image-net.org/about
8. ^Wikipedia. "WordNet." en.wikipedia.org/...WordNet
9. ^McCrae, John P.; Rademaker, Alexandre; Bond, Francis; Rudnicka, Ewa; Fellbaum, Christiane. "English WordNet 2019 - An Open-Source WordNet for English." Proceedings of the 10th Global Wordnet Conference, 2019. aclanthology.org/2019.gwc-1.31
10. ^Global Wordnet Association. "Open English Wordnet." GitHub repository README. github.com/...english-wordnet
11. ^Global WordNet Association. "The Global WordNet Association." globalwordnet.github.io/...gwa
12. ^Miller, George A. "WordNet: A Lexical Database for English." Communications of the ACM 38(11), 1995. dl.acm.org/...219717.219748
13. ^Fellbaum, Christiane (ed.). WordNet: An Electronic Lexical Database. MIT Press, 1998. direct.mit.edu/...etAn-Electronic-Lexical-Database
14. ^Princeton University. "License and Commercial Use of WordNet." wordnet.princeton.edu/license-and-commercial-use
15. ^Open English Wordnet website. en-word.net
16. ^Deng, Jia; Dong, Wei; Socher, Richard; Li, Li-Jia; Li, Kai; Fei-Fei, Li. "ImageNet: A Large-Scale Hierarchical Image Database." IEEE CVPR 2009. image-net.org/...imagenet_cvpr09.pdf
17. ^ImageNet. "Large Scale Visual Recognition Challenge 2012 (ILSVRC2012)." image-net.org/...2012
18. ^Russakovsky, Olga; Deng, Jia; Su, Hao; et al. "ImageNet Large Scale Visual Recognition Challenge." arXiv:1409.0575, 2014. arxiv.org/...1409.0575
19. ^ILLC, University of Amsterdam. "EuroWordNet: Building a multilingual database with wordnets for several European languages." archive.illc.uva.nl/EuroWordNet
20. ^Open Multilingual Wordnet. omwn.org
21. ^Global WordNet Association. Homepage (conference history and Wordnets in the World directory). globalwordnet.github.io