[Paper Review] What Can We Learn from Collective Human Opinions on Natural Language Inference Data?

Yixin Nie, Xiang Zhou+1 more|arXiv (Cornell University)|Oct 7, 2020

Topic Modeling37 references4 citations

TL;DR

This paper introduces ChaosNLI, a large-scale dataset with 464,500 human annotations across 4,645 NLI examples, revealing that high human disagreement persists in widely used NLI benchmarks like SNLI, MNLI, and αNLI. It demonstrates that state-of-the-art models perform nearly at random on low-agreement examples—where most errors occur—highlighting the limitations of relying solely on majority-label evaluation and advocating for model evaluation against the full distribution of human opinions instead of a single gold label.

ABSTRACT

Despite the subjective nature of many NLP tasks, most NLU evaluations have focused on using the majority label with presumably high agreement as the ground truth. Less attention has been paid to the distribution of human opinions. We collect ChaosNLI, a dataset with a total of 464,500 annotations to study Collective HumAn OpinionS in oft-used NLI evaluation sets. This dataset is created by collecting 100 annotations per example for 3,113 examples in SNLI and MNLI and 1,532 examples in Abductive-NLI. Analysis reveals that: (1) high human disagreement exists in a noticeable amount of examples in these datasets; (2) the state-of-the-art models lack the ability to recover the distribution over human labels; (3) models achieve near-perfect accuracy on the subset of data with a high level of human agreement, whereas they can barely beat a random guess on the data with low levels of human agreement, which compose most of the common errors made by state-of-the-art models on the evaluation sets. This questions the validity of improving model performance on old metrics for the low-agreement part of evaluation datasets. Hence, we argue for a detailed examination of human agreement in future data collection efforts, and evaluating model outputs against the distribution over collective human opinions. The ChaosNLI dataset and experimental scripts are available at https://github.com/easonnie/ChaosNLI

Motivation & Objective

• To investigate the distribution of human opinions on standard NLI datasets, challenging the assumption that majority labels represent ground truth.
• To assess whether state-of-the-art NLU models can capture the full distribution of human annotations rather than just predicting the majority label.
• To evaluate whether model performance correlates with human agreement levels and whether low-agreement examples represent the primary source of model errors.
• To advocate for a shift in evaluation from majority-label accuracy to modeling the distribution of collective human judgments, especially in pragmatic, subjective NLU tasks.

Proposed method

• Collect 100 human annotations per example for 3,113 examples in SNLI and MNLI-matched, and 1,532 examples in αNLI, resulting in 464,500 total annotations.
• Compute entropy of the human label distribution per example to quantify human agreement levels, using this as a proxy for example difficulty.
• Train and evaluate state-of-the-art models (e.g., BERT, RoBERTa, XLNet, ALBERT, DistilBERT, BART) on the same NLI tasks using standard classification heads.
• Measure model performance using both standard accuracy and distributional metrics like Jensen-Shannon divergence (JSD) and Kullback–Leibler (KL) divergence between model softmax outputs and the empirical human label distribution.
• Partition the dataset into bins based on entropy (i.e., agreement level) and compare model accuracy and distributional similarity across these bins.
• Analyze shared errors across models to identify patterns in failure modes, particularly on low-agreement examples.

Experimental results

Research questions

• RQ1 How prevalent is high human disagreement in standard NLI datasets like SNLI, MNLI, and αNLI when annotated with significantly more human judgments?
• RQ2 To what extent can state-of-the-art NLU models predict the full distribution of human annotations rather than just the majority label?
• RQ3 Does model performance correlate with the level of human agreement, and are models significantly worse on examples with low human agreement?
• RQ4 Do the most common errors made by state-of-the-art models occur predominantly on examples with low human agreement, suggesting that these are the most challenging cases?
• RQ5 Can evaluating models against the distribution of collective human opinions provide a more reliable and informative benchmark than relying on majority-label accuracy?

Key findings

• High human disagreement persists in 10% to 31% of examples across αNLI, SNLI, and MNLI-matched, even with 100 annotations per example, challenging the validity of the majority label as ground truth.
• State-of-the-art models achieve near-perfect accuracy (up to 0.9) on high-agreement examples but perform at or below random chance (around 0.5) on the lowest-agreement bins, indicating a fundamental failure to generalize to ambiguous cases.
• Models fail to capture the distribution of human opinions, as measured by high Jensen-Shannon divergence (JSD) and KL divergence, even when they achieve high accuracy on the majority label.
• The shared errors of state-of-the-art models are predominantly concentrated in the low-agreement subset of the data, suggesting that this subset represents the core challenge in current NLU evaluation.
• The level of human agreement is a strong indicator of model difficulty and is correlated with uncertainty in model predictions, supporting its use as a diagnostic tool for model calibration and reliability.
• Evaluating models against the full distribution of human opinions—not just the majority label—reveals significant gaps in model capabilities, especially in pragmatic reasoning and subjective inference tasks.