Domain: Social Media/Content Moderation
Description: Automated content moderation systems on major platforms show systematic bias in flagging and removing content, disproportionately affecting marginalized communities and non-English content.

Ethical Challenges:

• Cultural Bias: Western-centric training data leading to misclassification of cultural content
• Language Bias: Lower accuracy for non-English content and code-switching
• Political Bias: Inconsistent enforcement across political spectrum
• Context Insensitivity: Inability to understand cultural or historical context

Public Datasets:

• HatEval Dataset: Multilingual detection of hate speech against immigrants and women
• URL: https://competitions.codalab.org/competitions/19935
• OLID Dataset: Offensive Language Identification Dataset with three-level annotation
• URL: https://sites.google.com/site/offensevalsharedtask/
• Founta et al. Dataset: 4-class hate speech detection dataset from Twitter
• URL: https://github.com/ENCASEH2020/hatespeech-twitter