Causality-Guided Data Augmentation for Cross-Platform Hate Speech Detection

Tianming Jiang

doi:10.1111/exsy.70283

Return

Causality-Guided Data Augmentation for Cross-Platform Hate Speech Detection

2026-05-04

PRE

DOI：10.1111/exsy.70283

Original

Original request for help

Save

Abstract

En 中文

Modern social media platforms empower users to share opinions widely but also facilitate the spread of hate speech. Deep learning models achieve strong results on in-domain data yet often degrade sharply when applied to content from different platforms. This cross-platform challenge calls for methods that can generalise across domains. Recent approaches attempt to address this by modelling causal relationships, aiming to remove superficial patterns and focus on the true drivers of hate speech. However, these methods usually depend on manually defined causal cues, which are difficult to obtain and may not transfer across platforms. In this paper, we propose Causality-guided Data Augmentation (CDA), a training framework that improves cross-platform robustness without requiring predefined causal cues, target labels, or additional supervision. CDA is motivated by causal invariance, the intuition that relationships stable across environments are more likely to generalise than environment-specific correlations. Operationally, CDA approximates this idea through intervention-inspired masking, which perturbs high-impact source-domain features to reduce shortcut reliance and encourage the model to discover alternative signals that transfer better across platforms. Experiments on four real-world hate speech datasets from different platforms show that CDA consistently outperforms state-of-the-art cross-platform detection methods, improving average macro-F1 by approximately 5%. These results highlight CDA's potential to enable more reliable and scalable hate speech detection across diverse online communities.

Keywords:

causality

cross-platform

data augmentation

hate speech detection