95% confidence intervals overlap between Baseline NLI Diversity, Confidence NLI Diversity, SentBERT, and human judgments on conTest

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Confidence NLI Diversity achieves ρ=0.64 correlation with human diversity judgments on conTestfinding0.877
  Highest human correlation for semantic diversity metric
• Confidence NLI Diversity achieves state-of-the-art performance on measuring semantic diversityclaim0.829
  Main performance claim of the paper
• Confidence NLI Diversitymethod0.813
  Best-performing variant: aggregates softmax probability mass rather than binary class counts
• NLI Diversity is limited by the scope of the NLI model's training dataclaim0.792
  Limitation acknowledged in discussion section
• Neutral NLI predictions may capture lexical rather than semantic diversityhypothesis0.789
  Hypothesis proposed to explain Neutral NLI Diversity's high performance on decTest but low on conTest
• Baseline NLI Diversity – MNLI achieves Spearman ρ=0.59 on conTest diversity parameter correlationfinding0.776
  Comparable to top-performing automatic metric from Tevet and Berant 2021
• What is the role of neutral NLI predictions in measuring semantic diversity?question0.772
  Motivates the creation of Neutral NLI Diversity as an ablation
• Baseline NLI Diversitymethod0.765
  First variant: aggregates argmax NLI class predictions with contradiction=+1, entailment=-1, neutral=0