They are looking at the continuous embeddings, not at the discrete words inferred from them