“I’m sorry to hear that”: Finding New Biases in Language Models with a Holistic Descriptor Dataset

By Eric Michael Smith, Melissa Hall, Melanie Kambadur, Eleonora Presani, Adina Williams. 2022

207 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

Llama 2: Open Foundation and Fine-Tuned Chat Models (2023, methodology): “We also analyze the representation of different demographic groups in the pretraining data bymeasuring rates of usage of demographic identity terms from theHolisticBias dataset (Smith et al., 2022) as a proxy.”
Scaling Instruction-Finetuned Language Models (2022, background): “Additionally, we highlight that improvements in this work related to toxic language and identity groups remain limited to a biased subset of identity terms and in English only (Smith et al., 2022; Bhatt et al., 2022; Dev et al., 2021b).”
PaLM 2 Technical Report (2023, background): “Additionally, evaluations of classification bias remains limited to a biased subset of identity terms in English (Smith et al., 2022; Bhatt et al., 2022; Dev et al., 2021b).”
Gender bias and stereotypes in Large Language Models (2023, background): “However, as is well known, language models perpetuate and occasionally amplify biases, stereotypes, and negative perceptions of minoritized groups in society [10, 13, 14, 66, 69, 84, 85, 90].”
Large Language Model Alignment: A Survey (2023, background): “, 2023) extends the HolisticBias (Smith et al., 2022) to up to 50 languages, emphasizing the universality and diversity of biases in a multilingual environment.”
Marked Personas: Using Natural Language Prompts to Measure Stereotypes in Language Models (2023, methodology): “For this classification, we anonymize the data and then remove punctuation, capitalization, pronouns, and any descriptors that are explicit references to gender, race, or ethnicity using the list of holistic descriptions provided by Smith et al. (2022).”
Bias and Fairness in Large Language Models: A Survey (2023, methodology): “, Responding "I’m sorry to hear that" to "I’m an autistic dad" conveys a negative misrepresentation of autism (Smith et al., 2022) Stereotyping Negative, generally immutable abstractions about a labeled social group e.”