Less is Less: When Are Snippets Insufficient for Human vs Machine Relevance Estimation?