They are embedded into a particular semantic vector space that is learned based on a model. Another feature vector could be hand rolled based on feature engineering, tidf ngrams etc. Embedding is typically distinct from feature engineering that is manual.