Probabilistic Soft Logic for Toxic Intent Prediction in Conversation: a Moral Foundations Theory-Driven Neural-Symbolic Framework

Transformer-based models (RoBERTa, DeBERTa, ToxicBERT) achieve high accuracy but suffer critical limitations: they operate as black boxes, classify content based on surface features regardless of toxic intent, and exhibit biases when profanity is present. We present a neural-symbolic framework that integrates Probabilistic Soft Logic with Moral Foundations Theory, enabling accurate prediction of toxic intent and moral reasoning. Our approach improves prediction accuracy, provides interpretable explanations, and demonstrates synergistic effects. We evaluated 748,283 toxic instances (from a total of 1,268,154 conversations) across Telegram war discourse and Reddit climate debates. The results show a 4.9 percentage point improvement over the strongest transformer baseline with statistical significance. The framework provides interpretable explanations for 91.3\% of predictions through PSL rules grounded in moral psychology. Ablation analysis confirms genuine neural-symbolic synergy with full integration, achieving F1=0.847, significantly outperforming individual neural and symbolic components. This advances ethical AI and conversation moderation.