Natural Language Processing

Head-Gated Dynamic Decoupling for Effective Implicit Hate Speech Detection

Implicit hate speech detection relies heavily on contextual reasoning with often scattered linguistic clues. In mixed-data training, existing models suffer from a dominance of explicit samples during parameter updates, which suppresses the capture of complex implicit semantics. We attribute this asymmetric performance degradation to representation competition and gradient starvation, where strong explicit gradients hinder the effective learning of implicit representations. To address this, we propose the head-gated dynamic decoupling (HGDP) framework. Architecturally, HGDP introduces a sample-aware sparse gating mechanism that constructs specialized computational subgraphs by dynamically activating selective attention heads for explicit versus implicit samples. Optimization-wise, we design a conditional gradient flow (CGF) strategy to structurally block gradient interference from strong-signal samples onto decoupled pathways. Empirical evaluations demonstrate that HGDP yields substantial gains in implicit detection benchmarks without compromising performance on explicit samples. These results effectively validate the framework's capacity to alleviate gradient starvation and enhance overall model robustness.