Baseline NLI Diversity

First variant: aggregates argmax NLI class predictions with contradiction=+1, entailment=-1, neutral=0

Neighborhood — ranked by edge-count

paper

method

NLI Diversity
extendsrelated_to
Novel metric proposed in this paper using NLI predictions to score semantic diversity of a response set
Neutral NLI Diversity
extendsrelated_to
Variant weighting neutral predictions equally to contradictions to test if neutrals capture lexical diversity
Confidence NLI Diversity
extends
Best-performing variant: aggregates softmax probability mass rather than binary class counts

cosine ≥ 0.65 · no typed edge

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

NLI Diversity is limited by the scope of the NLI model's training dataclaim0.815
Limitation acknowledged in discussion section
What is the effect of more generalized NLI training data on NLI Diversity performance?question0.793
Future work question arising from the finding that Combined model did not outperform MNLI model
Confidence NLI Diversity achieves state-of-the-art performance on measuring semantic diversityclaim0.779
Main performance claim of the paper
95% confidence intervals overlap between Baseline NLI Diversity, Confidence NLI Diversity, SentBERT, and human judgments on conTestfinding0.765
Indicates lack of statistically significant differences between top methods
What is the role of neutral NLI predictions in measuring semantic diversity?question0.762
Motivates the creation of Neutral NLI Diversity as an ablation
Neutral NLI predictions may capture lexical rather than semantic diversityhypothesis0.758
Hypothesis proposed to explain Neutral NLI Diversity's high performance on decTest but low on conTest
Natural trait (expressed at baseline)concept0.748
A trait already strongly expressed at baseline (alpha=0) without steering intervention
Diversity Thresholdconcept0.743
Target minimum diversity level (e.g., 10 contradictions) that DTG iterates toward