How Neighborly works

The concept

Every clue is one of the hidden word’s neighbors in meaning. The first clue is a distant neighbor, and each wrong guess brings a closer one, until the last clue is the hidden word’s very nearest neighbor.

But how does a computer know which words are “near” each other? That’s the fun part.

Words as places on a map

Nobody typed in a dictionary of meanings. Instead, a program read billions of words of ordinary text and noticed which words turn up in similar sentences. “The ___ chased the ball” could be puppy or kitten; it’s rarely van. Linguists put it neatly: you shall know a word by the company it keeps.

From that, the program gives every word a position, like a place on a map. Words used in similar ways end up close together; unrelated words end up far apart.

Not 2 directions, but 300

A paper map has two directions: east–west and north–south. Two isn’t nearly enough for language. Puppy is close to kitten (young animals), but also to cute (feelings) and pony (pets). So the program uses 300 directions at once.

That means every word becomes a list of 300 numbers, something like:

puppy → 0.21, −0.43, 0.08, 0.66, … (296 more)

That list is called the word’s embedding. No single number means anything obvious on its own; meaning lives in the whole pattern. You can’t picture 300 directions (nobody can), but a computer handles them as easily as two.

Measuring closeness

Think of each word’s list of numbers as an arrow pointing somewhere in that 300-direction space. Two words are close in meaning when their arrows point the same way.

The game measures this with a score from about −1 to 1, called cosine similarity:

• near 1: the arrows point almost the same way (cello and viola)
• near 0: unrelated (cello and pudding)

Nearest neighbors

To make a puzzle, the computer takes the hidden word and scores every other word in its vocabulary against it, then sorts them from closest to farthest. Those are its nearest neighbors.

Your clues come from that list: the 30th-closest word first, then the 22nd, 16th, 11th, 7th, 4th, 2nd, and finally the closest of all. Before choosing, it throws out:

• giveaways, like plurals or words sharing the answer’s root (writers for writing)
• very rare words and filler words (really, which)
• anything offensive, using a blocklist

Hot and cold

When you guess a word that isn’t the answer, the game ranks your guess against the answer the same way: out of about 27,800 common English words, how close is yours? Ranked in the top 10 is 🔥 Hot; top 100 is ☀️ Warm; and further out it cools to ❄️ Cold.

Why it’s sometimes weird

The map was learned from Wikipedia and news articles written up to about 2014, so it reflects how words were used there, which isn’t always how you’d use them.

• News habits leak in. In news writing, bitter mostly appears in bitter rivalry and bitter dispute, so its neighbors are about fighting, not taste.
• Opposites are neighbors. Win and lose, early and late: opposites appear in the same kinds of sentences, so they land close together. That surprises almost everyone.
• Data carries human bias. Text written by people reflects people’s prejudices, and a map learned from it does too. That’s one reason the clues go through a blocklist and were reviewed by hand.

Your puzzle’s map

Here is a puzzle you’ve finished, with the answer and its eight clues. To draw it, the 300 directions are squeezed down to the 2 that spread these nine words out the most (a technique called principal component analysis). Squeezing loses information, so treat it as a sketch: words that look close here are usually close, but not always.

It all runs in your browser

There’s no server and no AI service behind the game. The puzzles and a compressed copy of the word map (96 numbers per word instead of 300, about 2.7 MB) are downloaded once, and every comparison happens on your device. Nothing you type is sent anywhere. Your stats are kept only in this browser.

Credits

• GloVe word vectors (glove.6B, 300 dimensions) by Jeffrey Pennington, Richard Socher and Christopher D. Manning, Stanford NLP. Public domain under theODC Public Domain Dedication and License 1.0.
• SCOWL word lists by Kevin Atkinson and contributors, used to keep the vocabulary to real, common English words. Copyright 2000–2020; used under its permissive license.
• List of Dirty, Naughty, Obscene, and Otherwise Bad Words(English) by Shutterstock, used as a blocklist. LicensedCC BY 4.0.
• NLTK Snowball stemmer, Apache 2.0, used to filter out giveaway clues.