To answer, words are not extracted in isolation. The model sees a context of about 20 words around it always.