Causality Guided Representation Learning for Cross-Style Hate Speech Detection

TL;DR

This work targets robust cross-style hate speech detection by reframing hate generation as a causal process and introducing CADET, a causal representation learning framework. CADET disentangles posts into four latent factors—creator motivation $M$, target $T$, style $S$, and context $U$—and employs confounder mitigation and latent counterfactual reasoning to isolate genuine hate intent from surface cues. Across four real-world datasets, CADET achieves superior cross-style generalization, with notable gains in explicit-to-implicit and implicit-to-explicit transfers and strong ablation results highlighting the contribution of each component. The approach offers interpretable, causally grounded insights and practical implications for safer online moderation by focusing on invariant hate signals rather than stylistic artifacts.

Abstract

The proliferation of online hate speech poses a significant threat to the harmony of the web. While explicit hate is easily recognized through overt slurs, implicit hate speech is often conveyed through sarcasm, irony, stereotypes, or coded language -- making it harder to detect. Existing hate speech detection models, which predominantly rely on surface-level linguistic cues, fail to generalize effectively across diverse stylistic variations. Moreover, hate speech spread on different platforms often targets distinct groups and adopts unique styles, potentially inducing spurious correlations between them and labels, further challenging current detection approaches. Motivated by these observations, we hypothesize that the generation of hate speech can be modeled as a causal graph involving key factors: contextual environment, creator motivation, target, and style. Guided by this graph, we propose CADET, a causal representation learning framework that disentangles hate speech into interpretable latent factors and then controls confounders, thereby isolating genuine hate intent from superficial linguistic cues. Furthermore, CADET allows counterfactual reasoning by intervening on style within the latent space, naturally guiding the model to robustly identify hate speech in varying forms. CADET demonstrates superior performance in comprehensive experiments, highlighting the potential of causal priors in advancing generalizable hate speech detection.

Figures (7)

• Figure 1: Distribution of hate speech styles and target groups across various social media platforms.
• Figure 2: Causal graph of hate-speech generation. An unobserved contextual environment (U) shapes creator motivation (M), target (T), and style (S). M, T, and S jointly produce the post (X). The hate label (Y) is determined by M and U.
• Figure 3: Framework of the proposed method. CADET firstly decomposes the factor in the latent space guided by the causal graph. Then, it mitigates the confounding effect of the platform, where various constraints are applied to enable the meaning representation learning. various. Finally, latent counterfactual reasoning is employed to learn style-invariant features.
• Figure 4: Training loss curve. The training process benefits from the loss-weighting curriculum schedule.
• Figure 5: Performance on LLM-transformed Hate Speech. CADET is robust under controlled style transformations.