On crowdsourcing relevance magnitudes for information retrieval evaluation