Harmful Speech Detection by Language Models Exhibits Gender-Queer Dialect Bias

Overview

Watch a research seminar exploring bias in content moderation systems and their impact on LGBTQ+ online expression. Examine findings from a study investigating how language models handle reclaimed slurs in gender-queer dialect, using the novel QueerReclaimLex dataset with 109 curated templates. Learn how five off-the-shelf language models perform in assessing potential harm in texts, and discover the challenges in accurately moderating content authored by gender-queer individuals. Understand the implications of current content moderation practices that disproportionately flag posts from transgender and non-binary users as toxic. Explore potential solutions through chain-of-thought prompting to help language models better consider author identity context. Presented by Rebecca Dorn, a PhD candidate at USC's Information Science Institute, whose research focuses on AI fairness, natural language processing, and computational social science, particularly examining how NLP systems interact with marginalized communities' dialects.

Syllabus

Harmful Speech Detection by Language Models Exhibits Gender-Queer Dialect Bias

Taught by

USC Information Sciences Institute

Reviews

Start your review of Harmful Speech Detection by Language Models Exhibits Gender-Queer Dialect Bias

Taught by

Is AI Sexist? Examining Gender Bias in Large Language Models

Examining Gender Bias in Large Language Models - Is AI Sexist?

AI for Good - Detecting Harmful Content at Scale

Social Biases in Text Representations and Their Mitigation

Preventing Toxicity and Unconscious Biases Using Large Language and Deep Learning Models

Multilingual Large Language Models - Advances in Generative AI

Never Stop Learning.