On the Various Semantics of Similarity in Word Embedding Models