# CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models

**Authors:** Nikita Nangia, Clara Vania, Rasika Bhalerao, Samuel R. Bowman  
**arXiv:** [2010.00133](https://arxiv.org/abs/2010.00133)  
**Date:** 2020-09-30  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2010.00133v1/

## Key points

- Authors address cultural biases in pretrained language models
- Introduction of Crowdsourced Stereotype Pairs benchmark (CrowS-Pairs)
- Dataset comprises 1508 examples focusing on stereotypes related to nine types of bias
- Objective is to measure social biases in language models against protected demographic groups in the United States
- Each example presents a model with two sentences: one embodying a stereotype more strongly, and another with less stereotyping
- Evaluation of three widely-used MLMs using CrowS-Pairs reveals consistent favoring of sentences expressing stereotypes across all categories
- Urgent need for developing less biased language models highlighted
- CrowS-Pairs emerges as a valuable benchmark for assessing progress in mitigating social biases within language models, serving as a critical tool for researchers and practitioners aiming to build fairer and more inclusive NLP technologies

## Summary

In their paper "CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models," authors Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman address the issue of cultural biases present in pretrained language models. These models have shown success in various natural language processing tasks but also perpetuate harmful stereotypes due to the biases present in the training data. To combat this problem, the authors introduce the Crowdsourced Stereotype Pairs benchmark (CrowS-Pairs), a dataset comprising 1508 examples that focus on stereotypes related to nine types of bias such as race, religion, and age. The key objective of CrowS-Pairs is to measure social biases in language models against protected demographic groups in the United States. Each example in the dataset presents a model with two sentences: one that embodies a stereotype more strongly and another that contains less stereotyping. The dataset specifically targets stereotypes associated with historically disadvantaged groups and contrasts them with those related to advantaged groups. Through their evaluation of three widely-used MLMs using CrowS-Pairs, the authors make a significant finding – these models consistently favor sentences expressing stereotypes across all categories within the benchmark. This observation underscores the urgent need for developing less biased language models. As efforts continue towards creating more equitable AI systems, CrowS-Pairs emerges as a valuable benchmark for assessing progress in mitigating social biases within language models. The dataset serves as a critical tool for researchers and practitioners striving to build fairer and more inclusive NLP technologies.