Judging Credible and Unethical Statistical Data Explanations via Phrase Similarity Graph

Kang Zhang, Einoshin Suzuki

Abstract: We propose a graph-based method to judge credible and unethical statistical data explanations with the exploitation of human instincts proposed by Rosling et al. Our previous work proposes three categories of statistical data explanations and three corresponding judgment methods based on phrase embedding and carefully designed comparison conditions. However, we observe that the previous method β exhibits low accuracy in the explanations of (β) category due to its counter-intuitive semantic similarities between several phrases. To address this limitation and improve the performance, our new method β^2 constructs a Phrase Similarity Graph to generate additional comparison conditions and devises a credibility score to aggregate the conditions with their importance quantified by graph entropy. The experimental results show that our β^2 achieves over 81% accuracy while the previous method β achieves about 57%. Scrutiny reveals that our β^2 mitigates the problem of the counter-intuitive semantic similarities at a satisfactory level.