EXPLORE WORD EMBEDDINGS
wordmath
What happens when you subtract one meaning and add another?
CLOSEST MATCH
| Rank | Word or phrase | Similarity |
|---|
Similarity compares the direction of the vectors. It is not a probability or a mark for correctness. The three input terms are excluded from the results.
Show the numbers behind this calculation
Each input becomes numbers. Here are the first eight coordinates; the search uses all of them. We subtract and add each coordinate, then find vectors pointing in a similar direction.
How to explore
Try a country swap, a family relationship, or a change in word form. Change just one input and watch the neighbours change. Find one result that fits your expectation and one that surprises you.
What is being calculated?
Words and phrases are represented by lists of numbers called embeddings. The calculator uses OpenAI’s text-embedding-3-large. It computes A − B + C using unit-length input vectors and searches the answer collection by cosine similarity.
This is an approximate relationship, so read the result as “close to”, not “equals”. Patterns in language, ambiguity, and the available answer collection all affect what comes back. New phrases can be inputs even when they are absent from the answer collection.
Inputs are lowercased and extra spaces removed. Previously unseen inputs are sent to OpenAI to obtain an embedding and cached on this server. Please use public words and concepts rather than private information.
Which words can appear as answers?
Loading the answer collection…
The collection includes the complete English large wordlist from wordfreq, including names and uncommon words. The wordfreq list alone omits many multiword landmarks and names. The vocabulary reflects language use and may include offensive terms or stereotypes. It is not a dictionary of approved answers.
This collection also includes Wikipedia’s Vital Articles, adding multiword names and concepts across subjects.
Sources: wordfreq by Robyn Speer, with data from SUBTLEX (Brysbaert and colleagues), OpenSubtitles, Wikipedia, and other corpora. Source details and attribution.