Few-shot learning with enhanced calibration for language- and harm-aware multilingual hate speech detection (FLEC-HS)

Multilingual hate speech detection remains challenging in low-resource settings, where labelled data are limited, particularly for identity-targeted and context-dependent content. We present Few-Shot Learning with Enhanced Calibration for Language- and Harm-Aware Multilingual Hate Speech Detection (FLEC-HS). This modular framework combines few-shot prompting with a frozen GPT-4o model and a lightweight Feature-Conditioned Calibration Layer (FCL) for post-hoc probability calibration. The primary evaluation uses HateDay (approximately 160,000 posts across eight languages), while HateCheck and HateXplain provide complementary functional and cross-dataset diagnostics. The term few-shot refers specifically to the prompting stage, which uses up to K = 16 demonstrations per language. The complete system additionally uses approximately 400 labelled development instances per language to train the FCL and select operating thresholds. Thresholds are selected by five-fold cross-validation on development data and frozen before test evaluation. On HateDay, FLEC-HS achieves a macro-F1 of 0.64, compared with 0.56 for the fully fine-tuned XLM-R-large. At the development-selected operating point requiring hate-class precision of at least 95%, hate recall increases from 0.48 to 0.61. Relative to the fully fine-tuned XLM-R-large baseline (48% review share), FLEC-HS reduces the share of posts sent to human review to 28% under the evaluated entropy-routing policy. Removing the FCL reduces macro-F1 to 0.59 and increases the protected-group false-negative rate from 5.1 to 10.4%. Protected-group fairness is operationalised narrowly through false-negative rates derived from dataset metadata; no broader fairness guarantee is claimed. The findings are limited to the evaluated benchmarks and configuration and remain subject to domain shift, commercial API dependencies, and incomplete coverage of dialects and protected groups.