Combating Hate Speech with Multi-Task Learning
Improving generalization for detecting harmful content targeting political figures
This research introduces a Multi-task Learning (MTL) framework that significantly improves hate speech detection across different domains and datasets.
- Creates a unified approach that overcomes inconsistent definitions and labeling criteria
- Demonstrates improved performance when generalizing to new domains
- Shows particular effectiveness for detecting harmful content targeting political figures
- Addresses the real-world challenge of cross-dataset performance
From a security perspective, this work provides essential tools for platforms to better protect vulnerable individuals and communities from online harm, helping maintain safer digital spaces while addressing the technical challenges of content moderation at scale.