How AI understands words: a map of meanings
Big idea: AI turns words into numbers, so that words with similar meanings sit close together.
The story in short
On a rainy day, Happy tipped out Dadi’s 300 handmade word cards, in English, Hindi and Kannada. Sorting by colour or by first letter made no sense. Sorting by meaning worked: “mango”, “आम” and “ಮಾವಿನ ಹಣ್ಣು” ended up side by side on the floor. Kabir chachu explained that AI gives every word an address made of numbers, learned from which words fill the same blanks: “I ate a ripe, juicy ___”. Then came the arrow trick: raja − boy + girl ≈ rani. But Riya found that some old word maps put “doctor” closer to “he” than “she”.
The idea
- Embedding: a list of numbers that gives a word an address on a map of meanings.
- Vector and dimensions: the list of numbers; real embeddings use hundreds or thousands.
- Context: the words around a word tell you how it is used, and so what it means.
- Multilingual: similar meanings land together, even across languages.
- Vector arithmetic: moving along the same arrow (boy → girl) turns raja into rani.
- Bias: unfair patterns in the training text get copied into the map.
- Semantic search: searching by meaning, not exact words. The society app found the water-tanker notice even though nobody typed “tanker”.
Where you see it
Search boxes that understand what you mean, translation apps, and chatbots all use embeddings.
Remember: a word is known by the company it keeps, and so is a machine that learns from words.
Check yourself
You search a shopping app for "something to keep tea hot" and it shows flasks. What makes that possible?
How does an AI learn where to place "mango" on its word map?
An old word map puts "nurse" near "she" and "engineer" near "he". Where did this come from?
More for you: 🎒 Students · 🧑🏫 Parents and teachers