Toxic language detection is an important task for online content moderation. However, existing models often associate neutral group-related terms—such as identity, regional, or demographic expressions—with toxic labels, producing false-positive bias in which benign texts containing such terms are misclassified as toxic. To address this, we propose NISC, a neuron-level iterative debiasing and semantic compensation framework for Chinese toxic language detection. Starting from a trained detector, NISC first identifies bias sources from high-confidence false-positive samples without a manually predefined lexicon. It then builds paired samples with and without each bias source and compares neuron-level toxic contributions between them. Guided by this contribution difference, NISC locates neurons that drive false-positive predictions and weakens them through iterative pruning. To offset the semantic degradation that pruning may cause, NISC uses a large language model to generate benign non-toxic samples that contain the identified bias sources but express no attack, insult, discrimination, or hatred, and applies them for lightweight compensation training with the pruning mask fixed. Experiments on multiple Chinese toxic language datasets show that NISC lowers the false-positive rate on group-related neutral texts while largely preserving conventional detection performance.

Fine-grained toxic language detection aims to assess the harmfulness, expression style, and attack target of textual content, supporting refined content moderation on online platforms. Most existing methods model each subtask independently, failing to exploit semantic correlations among tasks, while conventional contrastive learning is prone to introducing false negative samples under multi-task, multi-label settings. To address these limitations, we propose a Prompt-Enhanced Multi-Task Contrastive Learning framework (PEMTCL). The framework employs a dynamic prompt generation mechanism to construct instance-level soft prompts, enabling the encoder to adaptively capture fine-grained semantic cues required by different subtasks. It then jointly optimizes four subtasks—toxicity identification, toxic type discrimination, expression type detection, and targeted group detection—achieving inter-task knowledge complementarity within a shared representation space. Furthermore, a label-aware multi-head contrastive learning module leverages multi-task label relationships for reliable negative sample selection and incorporates a sample reweighting mechanism, effectively mitigating false negatives and class imbalance. Experiments on two Chinese datasets, ToxiCN and FG-COLD, demonstrate that PEMTCL achieves the best performance on most subtasks, with notable improvements on fine-grained tasks. Ablation studies, training data scale analysis, and case studies further validate the effectiveness and robustness of the proposed model.