Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-Platform Hate Speech Detection with Weakly Supervised Causal Disentanglement

Paras Sheth, Tharindu Kumarage, Raha Moraffah, Aman Chadha, Huan Liu

TL;DR

The paper tackles cross-platform hate speech detection by addressing the evolving and platform-dependent nature of hate with a weakly supervised causal disentanglement approach called HATE-WATCH. It decomposes input representations into invariant causal cues $X_c$ and platform-target proxies $X_w$ using a VAE-based architecture fed by a RoBERTa encoder and a BART decoder, augmented with contrastive learning and confidence-based reweighting. Key contributions include a practical framework that operates without explicit hate-target labels, demonstrated via cross-platform experiments across four datasets, and evidence that the learned $X_c$ are invariant across platforms. The work advances scalable content moderation by reducing labeling demands while maintaining robust generalization, with broader implications for safer online communities and ethical deployment.

Abstract

Content moderation faces a challenging task as social media's ability to spread hate speech contrasts with its role in promoting global connectivity. With rapidly evolving slang and hate speech, the adaptability of conventional deep learning to the fluid landscape of online dialogue remains limited. In response, causality inspired disentanglement has shown promise by segregating platform specific peculiarities from universal hate indicators. However, its dependency on available ground truth target labels for discerning these nuances faces practical hurdles with the incessant evolution of platforms and the mutable nature of hate speech. Using confidence based reweighting and contrastive regularization, this study presents HATE WATCH, a novel framework of weakly supervised causal disentanglement that circumvents the need for explicit target labeling and effectively disentangles input features into invariant representations of hate. Empirical validation across platforms two with target labels and two without positions HATE WATCH as a novel method in cross platform hate speech detection with superior performance. HATE WATCH advances scalable content moderation techniques towards developing safer online communities.

Cross-Platform Hate Speech Detection with Weakly Supervised Causal Disentanglement

TL;DR

The paper tackles cross-platform hate speech detection by addressing the evolving and platform-dependent nature of hate with a weakly supervised causal disentanglement approach called HATE-WATCH. It decomposes input representations into invariant causal cues and platform-target proxies using a VAE-based architecture fed by a RoBERTa encoder and a BART decoder, augmented with contrastive learning and confidence-based reweighting. Key contributions include a practical framework that operates without explicit hate-target labels, demonstrated via cross-platform experiments across four datasets, and evidence that the learned are invariant across platforms. The work advances scalable content moderation by reducing labeling demands while maintaining robust generalization, with broader implications for safer online communities and ethical deployment.

Abstract

Content moderation faces a challenging task as social media's ability to spread hate speech contrasts with its role in promoting global connectivity. With rapidly evolving slang and hate speech, the adaptability of conventional deep learning to the fluid landscape of online dialogue remains limited. In response, causality inspired disentanglement has shown promise by segregating platform specific peculiarities from universal hate indicators. However, its dependency on available ground truth target labels for discerning these nuances faces practical hurdles with the incessant evolution of platforms and the mutable nature of hate speech. Using confidence based reweighting and contrastive regularization, this study presents HATE WATCH, a novel framework of weakly supervised causal disentanglement that circumvents the need for explicit target labeling and effectively disentangles input features into invariant representations of hate. Empirical validation across platforms two with target labels and two without positions HATE WATCH as a novel method in cross platform hate speech detection with superior performance. HATE WATCH advances scalable content moderation techniques towards developing safer online communities.
Paper Structure (19 sections, 10 equations, 3 figures, 3 tables)

This paper contains 19 sections, 10 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Preliminaries
  5. Disentangling Causal and Target representations
  6. Contrastive Learning Framework
  7. Confidence-Based Sample Reweighting
  8. Reconstructing the Input from the Disentangled components
  9. Model Training
  10. Experiments
  11. Dataset and Evaluation Metrics
  12. Experimental Settings
  13. RQ1. Performance Comparison
  14. RQ2. How Helpful Are The Weak-Labels?
  15. RQ3. Are the Disentangled Representations Truly Invariant?
  16. ...and 4 more sections

Figures (3)

  • Figure 1: The causal graph illustrates the hate speech detection mechanism, where $X_c$ denotes the causal factors predicting hate speech, $Y$ the hate label, $X_w$ the target, $X$ the input, and $P_l$ the latent platform variable affecting the target.
  • Figure 2: The HATE-WATCH architecture processes input $X$ via a RoBERTa to get initial representation $z$. This $z$ undergoes disentanglement into a causal component to identify invariant hate factors $X_c$, and a weakly supervised target component $X_w$ without true target labels. Both components' outputs, $X_c$ and $X_w$, are merged to form reconstructed embedding $\hat{z}$, which is decoded by BART to produce reconstructed input $\hat{X}$.
  • Figure 3: Visualizing the representations from different models to verify invariance across platforms. src (tgt) denote the source (target) platforms.