How Embeddings Store Meaning as Geometry?

Hinzugefügt: 2026-06-06

[00:00:00]How can embeddings capture relationships between words? In embedding space, words are not just points. They are directions. Here, queen is related to woman, and similarly, king is related to man. Now, watch the vector operation.

[00:00:14]Take queen, subtract woman, and add man.

[00:00:17]The result moves toward king. So, the relationship of queen to woman is similar to king to man. The heat map shows this with cosine similarity.

[00:00:25]Brighter cells mean the vectors are more aligned. This is why embeddings are powerful. They don't just store words, they store relationships as geometry.