임베딩이 의미를 포착하는 이유

Each word vector is a point in a high-dimensional space. Training arranges these points so geometry reflects meaning. 🌌

After training, words used in similar ways land near each other. The vectors for cat and dog end up close neighbors.

Words sharing the same contexts get pushed together, because the model learns to predict them in the same situations.

To compare two vectors, you use cosine similarity: the angle between them. A small angle means the words are alike.

Not only distance matters; the direction between vectors can encode relationships, like singular versus plural forms.

A classic result: the vector for king minus man plus woman lands near queen. Embeddings capture analogies as arithmetic.

Unlike one-hot vectors that treat every word as equally different, dense embeddings share structure, so learning transfers between similar words.

A handful of float dimensions can express thousands of subtle differences. This compression is why embeddings are so efficient.

You can load pretrained vectors trained on huge corpora, giving your model meaningful starting points for free.

Tools like t-SNE project embeddings down to 2D so you can see clusters of related words form clear neighborhoods.

No one hand-writes these vectors. The structure emerges purely from data and the training objective itself.

Embeddings put words in a space where distance and direction encode meaning, learned entirely from how words are used. ✅