Aligning Implied Statements for Implicit Hate Speech Generalizability with Context-Bounded Semi-hard Negative Mining

The paper introduces ImpSH, a triplet-based framework designed for classifying implicit hate speech by aligning posts with implied statements and utilizing context-bounded semi-hard negative mining. Evaluations on datasets such as IHC, SBIC, and DynaHate show that ImpSH, when paired with BERT and HateBERT, outperforms standard supervised contrastive methods, particularly in cross-domain scenarios. This approach enhances representation stability and reduces misclassification due to domain shifts, providing a more robust solution for practitioners addressing the complexities of implicit hate speech detection.