Social Bias Frames: Reasoning about Social and Power Implications of Language

Abstract

Warning: this paper contains content that may be offensive or upsetting. We then establish baseline approaches that learn to recover SOCIAL BIAS FRAMES from unstructured text. We find that while stateof-the-art neural models are effective at highlevel categorization of whether a given statement projects unwanted social bias (80% F 1), they are not effective at spelling out more detailed explanations in terms of SOCIAL BIAS FRAMES. Our study motivates future work that combines structured pragmatic inference with commonsense reasoning on social implications.

Multi-Dimensional Gender Bias Classification

“…Most relevant to this work, Sap et al (2020) proposed a framework for modeling pragmatic aspects of many social biases in text. Our work and theirs focus on complementary aspects of a larger goal-namely, making NLP safe and inclusive for everyone-but the two approaches differ in several ways.…”

“…We treat statistical gender bias in human or model generated text specifically, and in detail. Sap et al (2020) proposed a different but compatible perspective, and aimed to situate gender bias within the broader landscape of negative stereotypes in social media text, an approach that can make parallels apparent across different kinds of harmful text. Moreover, they considered different pragmatic dimensions than we do: they targeted negatively stereotyped commonsense implications in arguably innocuous statements, whereas we investigate pragmatic dimensions that straightforwardly map to conversational roles (i.e., topics, addressees, and creators of text).…”

Unsupervised Discovery of Implicit Gender Bias

“…Our work differs from prior work on bias detection in NLP in that we infer bias from data in an unsupervised way, whereas prior work relies on crowd-sourced annotations (Fast et al, 2016;Bolukbasi et al, 2016;Wang and Potts, 2019;Sap et al, 2020). This work typically focuses on specific types of bias, such as condescension (Wang and Potts, 2019) or microaggressions (Breitfeller et al, 2019) and involves carefully constructed annotations schemes that are difficult to generalize to other data sets or types of bias.…”

“…They also often rely on human-defined "known" stereotypes, such as lists of traditionally male and female occupations obtained through crowd-sourcing, which restricts analysis to a narrow surface-level domain. Similarly, supervised approaches can provide insight into carefully defined types of bias (Wang and Potts, 2019;Breitfeller et al, 2019;Sap et al, 2020), but they rely on human annotations tasks, which are difficult to design or generalize to other domains, especially because social concepts differ across contexts and cultures (Dong et al, 2019).…”