Smarter, Not Bigger: Unveiling Implicit Hate Speech with Scalable Language Models

ACL ARR 2025 February Submission7396 Authors

Abstract:

The detection of implicit hate speech is one of the critical challenges faced by the Natural Language Processing community as it requires the use of indirect and vague language-a step that traditional approaches find impossible. An exhaustive study of implicit hate detection has been carried out in this paper, concentrating on a subclass of tasks known as Implicit Target Span Identification, aimed at identifying spans of text that perform less explicit targeting of protected groups.

We systematically evaluate three Masked Language Models—BERT, RoBERTa, and Hate-BERT—alongside two Small Language Models ModernBERT and SmolLM2 and, as well as Large Language Models with LLama 3.2B and GPT-3.5. Our approach considers both zero-shot and fine-tuning methodologies while examining the effects of instruction tuning and Low-Rank Adaptation (LoRA) to assess their impact on detection tasks.

The results indicate that ModernBERT with only 149M parameters outperforms instruction-tuned larger models such as LLaM 3.2B, ModernBERT achieved F1 scores of 72.2 and 75.1 on IHC and SBIC datasets, respectively, while LLaM 3.2B attained 70.8 and 74.2 F1 scores for IHC and SBIC, respectively. RoBERTa-Large remains the best overall, scoring 72.5 F1 on the IHC dataset and 75.8 on the SBIC dataset. Compared to it, SmolLM2-135M attained an F1 score of 69.0 on IHC and 71.5 on SBIC, still showing competitive performance notwithstanding its smaller size.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: Implicit Hate Speech Detection, Bias and Fairness in NLP, Ethical AI, MLMs ,LLMs, Instruction Tuning, LoRA

Contribution Types: Approaches to low-resource settings, Approaches low compute settings-efficiency, Data resources

Languages Studied: English

Submission Number: 7396