DGDS: Reliability-Gated Neuro-Symbolic Learning for Roman Urdu Hate Speech Detection

Roman Urdu hate and offensive language detection is difficult because informal social-media posts combine non-standard spelling, code-mixing, local abusive expressions, and severe label imbalance. Existing transformer-based detectors improve contextual modeling, but they often leave lexical threat evidence implicit and rarely address whether model confidence is reliable when large language model teacher signals are used in low-resource training. We propose Dynamic Gated Distillation with Symbolic Evidence (DGDS), a reliability-gated neuro-symbolic learning framework for Roman Urdu hate and offensive language detection. DGDS combines a cross-lingual semantic encoder, a context-modulated symbolic threat profiler, and a reliability-gated dual-stream learner. The key idea is to treat the large language model teacher output as a noisy structured soft-label record rather than as a replacement for gold labels. DGDS therefore uses the teacher probability assigned to the gold class, the current student confidence, and a teacher reliability gate to form a dynamic importance weight for supervised training. Under a leakage-aware evaluation protocol on the RUHSOLD benchmark, the central contribution of DGDS is confidence reliability: it holds macro-F1 at parity with the strongest dual-stream baseline on both the binary and five-class settings while substantially reducing the expected calibration error from 0.0512 to 0.0134 on Task 1 (a 74% relative reduction) and from 0.0813 to 0.0268 on Task 2 (a 67% relative reduction). These calibration gains are statistically significant (p<0.001, paired stratified bootstrap) and robust to teacher replacement (GPT-4o-mini yields ECE 0.0158 vs. the 0.0512 baseline on Task 1). These results support a conservative conclusion: large language model soft labels are most useful for Roman Urdu harmful-content detection when they are filtered as uncertain training evidence and deployed through a local calibrated student model.