Evaluating Gender Bias in Natural Language Inference

TL;DR

This paper introduces a new evaluation method to detect gender bias in natural language inference models, revealing that popular models exhibit significant gender stereotypes, which can be mitigated with dataset balancing techniques.

Contribution

It proposes a challenge task for measuring gender bias in NLI models and evaluates state-of-the-art models, highlighting the effectiveness of debiasing strategies.

Findings

BERT, RoBERTa, BART models show gender bias in predictions.

Debiasing by gender-balanced datasets reduces bias.

Models trained on MNLI and SNLI are prone to gender stereotypes.

Abstract

Gender-bias stereotypes have recently raised significant ethical concerns in natural language processing. However, progress in detection and evaluation of gender bias in natural language understanding through inference is limited and requires further investigation. In this work, we propose an evaluation methodology to measure these biases by constructing a challenge task that involves pairing gender-neutral premises against a gender-specific hypothesis. We use our challenge task to investigate state-of-the-art NLI models on the presence of gender stereotypes using occupations. Our findings suggest that three models (BERT, RoBERTa, BART) trained on MNLI and SNLI datasets are significantly prone to gender-induced prediction errors. We also find that debiasing techniques such as augmenting the training dataset to ensure a gender-balanced dataset can help reduce such bias in certain cases.