The Tail Wagging the Dog: Dataset Construction Biases of Social Bias Benchmarks

TL;DR

This paper investigates how dataset construction choices influence social bias scores in benchmarks, revealing that simple modifications can significantly alter bias measurements and highlighting the need for more robust bias evaluation methods.

Contribution

It demonstrates that superficial changes in dataset construction can substantially impact social bias scores, questioning the reliability of current benchmarks.

Findings

Shallow modifications affect bias scores across models

Dataset construction biases can distort social bias measurements

Current benchmarks may not reliably measure true social biases

Abstract

How reliably can we trust the scores obtained from social bias benchmarks as faithful indicators of problematic social biases in a given language model? In this work, we study this question by contrasting social biases with non-social biases stemming from choices made during dataset construction that might not even be discernible to the human eye. To do so, we empirically simulate various alternative constructions for a given benchmark based on innocuous modifications (such as paraphrasing or random-sampling) that maintain the essence of their social bias. On two well-known social bias benchmarks (Winogender and BiasNLI) we observe that these shallow modifications have a surprising effect on the resulting degree of bias across various models. We hope these troubling observations motivate more robust measures of social biases.