Randomness versus specifics for word-frequency distributions