Hate speech detection in low resource languages using large language models

Articles in Press

Document Type : Original Article

Authors

Department of Mathematics and Computer Science, Amirkabir University of Technology (Tehran Polytechnic), Iran

Abstract

The prevalence of hate speech on social media platforms is increasing, prompting considerable attention from the research community to detect such harmful content. Recent studies have focused on refining language models (LMs) to effectively identify hate speech, resulting in notable advancements in performance. Nonetheless, the majority of these studies are confined to identifying hate speech exclusively in English, disregarding the vast amount of hateful content produced in other languages, notably those considered low-resource languages. Constructing a classifier capable of effectively detecting hate speech in a low-resource language with limited data poses a formidable challenge. To address the existing gap, we perform comparative study for five large language models and three Parameter-Efficient Fine-Tuning Methods, to determine which model and method excel in detecting hate speech proficiently on two languages that have limited linguistic resources available. Specifically, we evaluate three approaches: Sequence Classification based fine-tuning (SEQ_CLS), Causal language modeling-based fine-tuning (CLM), and In-context learning approach (ICL). Our findings emphasize the ability of generative models to address the challenges of data scarcity and enhance model performance through these methods and approaches.

Keywords

Main Subjects

AUT Journal of Mathematics and Computing

Articles in Press, Accepted Manuscript
Available Online from 20 January 2026

Files

Share

How to cite

Statistics