"I'm sorry to hear that": Finding New Biases in Language Models with a Holistic Descriptor Dataset

TL;DR

This paper introduces HolisticBias, a comprehensive dataset with nearly 600 descriptors across 13 demographic axes, enabling more inclusive bias detection in language models through over 450,000 prompts, revealing previously undetectable biases.

Contribution

The creation of HolisticBias, a participatorily assembled, extensive bias measurement dataset that covers more demographic axes than existing datasets, facilitating improved bias detection in NLP models.

Findings

HolisticBias uncovers biases in token likelihoods of language models.

It detects biases in offensiveness classifiers.

The dataset enables identification of previously undetectable biases.

Abstract

As language models grow in popularity, it becomes increasingly important to clearly measure all possible markers of demographic identity in order to avoid perpetuating existing societal harms. Many datasets for measuring bias currently exist, but they are restricted in their coverage of demographic axes and are commonly used with preset bias tests that presuppose which types of biases models can exhibit. In this work, we present a new, more inclusive bias measurement dataset, HolisticBias, which includes nearly 600 descriptor terms across 13 different demographic axes. HolisticBias was assembled in a participatory process including experts and community members with lived experience of these terms. These descriptors combine with a set of bias measurement templates to produce over 450,000 unique sentence prompts, which we use to explore, identify, and reduce novel forms of bias in several generative models. We demonstrate that HolisticBias is effective at measuring previously undetectable biases in token likelihoods from language models, as well as in an offensiveness classifier. We will invite additions and amendments to the dataset, which we hope will serve as a basis for more easy-to-use and standardized methods for evaluating bias in NLP models.