Publications

Nguyen Cong Phat, et al.

Hate speech detection on Vietnamese social media remains challenging due to tonal ambiguity of the language, pervasive teencode, and culturally encoded implicit attacks that evade surface-level classifiers. Existing approaches largely ignore why content is hateful (the underlying stereotypes, harm mechanisms, and attack targets), limiting their ability to discriminate between offensive insults and dehumanizing hate. To address this, we adapt the HARE rationale-augmented training framework to Vietnamese targeted hate speech detection on ViTHSD, a 10,001-comment corpus with 11 multi-label target-severity classes. A two-stage QLoRA pipeline fine-tunes Qwen2.5-3B-Instruct: Stage 1 trains on 7,540 oversampled examples; Stage 2 continues on 1,221 LLM-generated, human-reviewed Chain-of-Thought rationale tuples encoding target identification, harm-mechanism reasoning, and severity justification. Multi-seed evaluation shows HARE achieves the best F1-Macro among compared models (35.93±0.63%) and reveals an asymmetric boundary redistribution, with disproportionate Offensive-level gains (+6.84 avg F1) alongside modest Hate-level improvement (+2.77 avg F1), not previously characterized in Vietnamese hate speech literature. These findings demonstrate that structured rationale supervision can improve fine-grained severity discrimination in low-resource, culturally complex settings, while highlighting the need for further disentangling rationale content from data-exposure effects.