Computational Social Science Forum — An Empirical Social Science Approach to Auditing a Hate Speech Model

CSS Training Program

October 26, 2021
4:00pm to 5:00pm
Virtual Participation


Computational Social Science Forum
Date: Tuesday, October 26, 2021
Time: 4:00-5:00 PM Pacific Time
Location: Virtual Participation – Register to attend via Zoom

An Empirical Social Science Approach to Auditing a Hate Speech Model

Speakers: Renata BarretoGraduate Student Researcher, Berkeley Law, and Pratik Sachdeva, Postdoc, Social Sciences D-Lab
Abstract : Over the last decade, researchers in academia and the tech industry have developed a number of natural language processing (NLP) models for the detection of hate speech. Hate speech models have many limitations and biases, such as decreased performance on African American Vernacular English (AAVE), difficulties differentiating sentiment of posts that include identity terms, and flagging counter speech as hate speech. To mitigate these biases, scholars have employed diverse methodologies, yet very few studies examine the role of labelers’ demographic characteristics on the labels themselves. Instead, most machine learning audits focus on class imbalance in the training data. Bringing a computational social science perspective to ML audits, this paper examines how annotators from different groups label hate speech data. Unlike previous scholarship that infers demographic characteristics, our data collected rich, granular, and self-identified groups from our annotators. Furthermore, we do not only look at one identity group along a binary axis (i.e. gender: male / female or race: black /white), and instead draw on annotators’ categorical, self-described characteristics: sexual orientation, race, gender, citizenship, ideology, and disability. We find that annotators from the same identity being targeted in a post are more likely to give it a higher score. These findings have major implications for the collection of data for training hate speech models by social media companies and academic researchers. 

The Computational Social Science Forum is an informal setting for the interdisciplinary exchange of ideas and scholarship at the intersection of social science and data science. Participants engage in a variety of activities such as presentations of work in progress, discussions and critiques of recent papers, introductions to new tools and methods, discussions around ethics, fairness, inequality, and responsible conduct of research, as well as professional development. This Forum is organized as part of the Computational Social Science Training Program, and weekly meetings are hosted by researchers from BIDS and D-Lab. The group welcomes social scientists and researchers with interests in data science methods and tools, and data scientists with applications or interests in public policy, social, behavioral, and health sciences. Participants include graduate students, postdocs, staff, and faculty, and members are encouraged to attend regularly in order to foster community around improving computational social science research, supporting the development and research of group members, and fostering new collaborations. Interested UC Berkeley community members are invited to use this registration form to receive the schedule and access links. Please contact for more information or if you are interested in presenting current research for an upcoming session.


Renata Barreto

Graduate Student Researcher, Berkeley Law

Pratik Sachdeva

Postdoc, Social Sciences D-Lab