On the Relative Efficiency of Crowdsourced Rating Mechanisms: Experimental Evidence