ToxiGAN: Toxic Data Augmentation via LLM-Guided Directional Adversarial Generation

TL;DR

ToxiGAN addresses the challenge of imbalanced toxic-language data by combining adversarial generation with semantic guidance from LLMs. It uses LLM-generated neutral exemplars as semantic ballast and a two-step alternating directional learning process to push generated samples toward toxicity while preserving label fidelity and sub-mode coverage, mitigating mode collapse and semantic drift. The framework employs K class-specific generators and a multi-head discriminator, with an adaptive neutral ballast pool to stabilize training and maintain domain coherence. Across four hate-speech benchmarks, ToxiGAN achieves the best average Macro-F1 and Hate-F1 compared with GAN- and LLM-based baselines, with ablations confirming the value of semantic ballast and directional training. These results demonstrate a scalable, robust augmentation approach for toxicity classifiers in low-resource settings, while also engaging with ethical considerations around content safety and deployment.

Abstract

Augmenting toxic language data in a controllable and class-specific manner is crucial for improving robustness in toxicity classification, yet remains challenging due to limited supervision and distributional skew. We propose ToxiGAN, a class-aware text augmentation framework that combines adversarial generation with semantic guidance from large language models (LLMs). To address common issues in GAN-based augmentation such as mode collapse and semantic drift, ToxiGAN introduces a two-step directional training strategy and leverages LLM-generated neutral texts as semantic ballast. Unlike prior work that treats LLMs as static generators, our approach dynamically selects neutral exemplars to provide balanced guidance. Toxic samples are explicitly optimized to diverge from these exemplars, reinforcing class-specific contrastive signals. Experiments on four hate speech benchmarks show that ToxiGAN achieves the strongest average performance in both macro-F1 and hate-F1, consistently outperforming traditional and LLM-based augmentation methods. Ablation and sensitivity analyses further confirm the benefits of semantic ballast and directional training in enhancing classifier robustness.

Figures (7)

• Figure 1: ToxiGAN with k toxic generators, one neutral texts provider, and one multi-class discriminator.
• Figure 2: Illustration of Two-Step Alternating Directional Learning in Embedding Space. The black arrow shows the initial generation after pretraining. Gray arrows represent updates during alternating optimization: shifting toward toxicity and authenticity directions by penalizing unexpected directional evaluations.
• Figure 3: Average toxicity scores of real and synthetic samples across datasets.
• Figure 4: Training curves of ToxiGAN and ablations on OR. (ToxiGAN w/o semantic ballast degrades to SentiGAN.)
• Figure 5: Sensitivity to Various Augmentation Data Ratio on DC. Ratios refer to proportion of original data; remainder is filled with synthetic/duplicated toxic data.