Word embedding is a technique that converts words into numerical vectors, positioning each word as a point in a high-dimensional space where semantic relationships are captured as geometric ones. Words used in similar contexts end up close together in that space, so a well-trained embedding places "king" and "queen" near each other and further from an unrelated word like "bicycle," even though nothing in the raw text explicitly states that relationship.
This numerical form is what makes language usable by neural networks in the first place, since models operate on numbers, not raw text. Beyond that basic function, embeddings famously support a kind of vector arithmetic: subtracting the vector for "man" from "king" and adding "woman" lands close to "queen," a property that reflects how consistently the model has learned real relationships between concepts rather than just memorizing word co-occurrence.
