Evaluating the Examiner: The Perils of Pearson Correlation for Validating Text Similarity Metrics