Turning words into numbers

What this lesson shows

To do math on meaning, turn each word into a position: a vector.

The question it opens with

Is 'cat' more like 'dog' or like 'car'? With only letters, you can't compute it.

The walkthrough, in the lesson's own words

• On the scrambled map, tap the word you'd put next to cat: dog or car? Gut call.
• Watch the model sort every word: similar meanings drift together.
• Tap dog or car, whichever you'd place next to cat. There's no wrong answer yet.
• Watch the model place every word. Words that mean similar things drift into the same cluster.
• But the model can't see meaning, only letters. Do these share any?
• A hand-drawn map in the model's style; real maps are learned, see 2.5.
• You search a help center for “can't sign in” and it surfaces an article titled “login troubleshooting”, even though you typed none of those words. How?
• Each phrase is turned into a vector, a position by meaning, and the search matches on nearby positions, not shared letters. That is why “can't sign in” lands next to “login” even with zero words in common.
• No letter appears in both. By letters alone, cat and dog look unrelated.
• Scrambled at first, now watch them sort by meaning.
• Similar meanings cluster together. Near positions = similar meaning.
• Turning words into positions is what lets the model compare them by meaning at all, the missing piece behind the ranked guesses you saw in 0.1. Each position is a pair of numbers, called a
• Now you try: drag the mystery word anywhere and watch its nearest neighbour update.

Key takeaway

I placed 'cat' near 'dog' and far from 'car', and the machine agreed.

What you can do after this lesson

You can explain why words must become vectors for a machine to compare meaning.

Where it leads: But a flat map can't separate every nuance… we need more dimensions.

This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.