A plain summary, so you can get the gist here without leaving.
In 2013, a Google team found a fast way to turn words into lists of numbers that capture meaning, so that simple arithmetic on those numbers could answer analogies like king minus man plus woman lands near queen.
What it is
Word2Vec, led by Tomas Mikolov and colleagues at Google, is a method for learning vectors for words. A vector here is just a long list of numbers. Each word gets its own vector, and the trick is that the numbers are arranged so that words used in similar ways end up close together.
The model learns these vectors by reading huge amounts of text and trying to predict which words appear near each other. No human labels the meanings. The patterns of usage do the teaching.
The core idea
A word is known by the company it keeps. Words that show up in similar contexts, like coffee and tea, get similar vectors. What surprised people was that relationships between words showed up as consistent directions in this number space.
The famous example is that taking the vector for king, subtracting man, and adding woman lands you near queen. The model was never told this. It fell out naturally from learning how words are used, which hinted that meaning has a kind of geometry.
Why it matters
Word2Vec showed that meaning could be captured in a compact, computable form, and it did so efficiently enough to run on ordinary hardware. That made it practical and popular.
The idea of turning things into vectors that capture meaning, now often called embeddings, runs through modern AI. Search, recommendations, and the way large language models handle text all build on this foundation. If you work with AI, you meet embeddings constantly, and this paper is where the idea became widely usable.
- Published in 2013 by Mikolov and colleagues at Google.
- Turns each word into a vector of numbers that reflects its meaning.
- Learns from word context in raw text, with no human labels.
- Relationships appear as directions, so king minus man plus woman lands near queen.
- Launched the embeddings idea that underpins search, recommendations, and LLMs.
Mikolov et al., Google
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.