How Our Word List Is Built: WordNet, Frequency, and Curation

A random word generator is only as good as its word list. Ours is built from three layers: an academic lexical database as the foundation, frequency data to keep results useful, and hand curation where automation falls short. This article explains each layer — partly for transparency, and partly because the trade-offs are genuinely interesting.

The foundation: Princeton WordNet

Our base lexicon derives from Princeton WordNet, a lexical database of English built and maintained by researchers at Princeton University and used widely in linguistics and natural language processing. WordNet gives us more than spellings: it records each word's part of speech (noun, verb, adjective, adverb) and groups words into semantic categories, which is what powers our filters for word type and themes like animals, food, and nature. Starting from an academic source means the list isn't scraped from the web, so it doesn't inherit the web's typos, slang drift, or spam.

The ranking layer: word frequency

A raw dictionary is a bad generator. Pick uniformly from every English word and you'll mostly get obscure technical vocabulary, because obscure words vastly outnumber common ones. We rank words by how frequently they occur in real English usage and surface common, useful words first. This is a deliberate trade-off: a linguist might want the long tail, but someone building a Pictionary list, prompting a creative writing session, or teaching an ESL class wants words people actually know. When you filter to 5-letter words, you get 'house' and 'bread' before you ever see anything exotic.

The safety layer: family-friendly by default

Every word passes through an inappropriate-word filter before it can appear in results. Our audience includes classrooms and kids' party games, so the default experience has to be safe to project onto a whiteboard. Filtering is genuinely harder than it sounds — offensiveness depends on context, and over-filtering deletes innocent words (the classic 'Scunthorpe problem'). We tune the filter conservatively and review edge cases by hand.

The human layer: curated lists

Some of our most popular collections can't be computed. No database field marks a word as 'cool' or 'aesthetic' — those are aesthetic judgments, so those lists are hand-picked, word by word. Machine-derived categories also get a human pass: automated classification will happily tag technically-correct-but-useless members into a category, and we prune those. The split is visible on the site: category pages tell you whether a list is hand-curated or derived from WordNet's semantic data.

Known limitations

Honesty requires listing what the approach can't do. WordNet skews toward standard American English, so recent coinages and regional slang lag behind. Frequency ranking means genuinely rare words are hard to surface even when you might want them. And no profanity filter is perfect in either direction. If you spot a word that seems wrong — miscategorized, misspelled, or something that slipped the filter — we genuinely want to hear about it via the contact page; individual reports regularly lead to list fixes.