🤖 AI Summary
Existing hate speech detection models exhibit poor cross-dataset generalization, primarily due to inconsistent annotation criteria and severe bias transfer stemming from sparse, imbalanced samples involving political figures. To address this, we propose the first multi-task learning framework that explicitly models speaker identity sensitivity, jointly optimizing three complementary tasks: hate speech detection, stance classification, and political figure attribute prediction. Our architecture employs a shared BERT encoder with task-specific output heads, augmented by adversarial debiasing and gradient normalization to mitigate label sparsity and bias propagation. Experimental results demonstrate substantial improvements: F1 scores increase by 4.2–9.7% across multiple political figure subsets; zero-shot cross-figure transfer accuracy significantly surpasses single-task baselines; and model robustness and fairness are notably enhanced.
