Each word is converted into an embedding, which is a vector representation of our word in a features space, and is computed by a language model.
Then, to determine the similarity between two words, we compute the vector similarity between their embeddings.
Typically we do this using cosine similarity, that focuses on the angle between the vectors.