Designing Toxic Content Classification for a Diversity of Perspectives
summary
A survey of 17,280 participants produces about 108,000 labels showing people who are LGBTQ+, young, or previously targeted mark more comments as toxic while bystanders mark fewer, so identity and lived experience systematically shift labels. Off‑the‑shelf classifiers (e.g., Perspective API) score about precision 0.35 → 0.6 and accuracy 0.37 → 0.68 when tuned per user, and uploader analysis reports personalized model tuning improves accuracy by 86% on average, with 72% of participants seeing gains.
AI-written summary · about
chapters (4)
- 0:00Start
- 1:48Survey Participants (US only)
- 2:21Identities, Experiences, and Toxicity
- 3:38Fine Tuning Toxicity Classifiers
related talks
57mTackling toxicity: How to build a thriving player community
30mBeyond Labels: Evolving Data Classification
22mExploring User Perceptions of Discrimination in Online Targeted Advertising
25mEmbracing Real-World Pluralism: Moving beyond "gold standard" truth in AI evaluation
40mResponsible AI Practices: Fairness in ML
18mMultimodal Classification of Moderated Online Pro-Eating Disorder Content
29mLabeling against all odds
29mAnnotating data the right way