Roman Urdu Hate Speech Classification Explored

The proliferation of toxic and hateful content online poses significant societal challenges, particularly in low-resource languages like Roman Urdu, which is characterized by informal grammar and inconsistent spellings. This research investigates and compares various state-of-the-art techniques for effectively classifying hate speech in such challenging linguistic environments where data is scarce. The study designed four experiments to evaluate different approaches. It began by assessing the zero-shot understanding of Roman Urdu by Large Language Models (LLMs) without fine-tuning. Subsequent experiments explored parameter-efficient fine-tuning (PEFT) using LoRA, which minimizes computational costs by updating only a small subset of parameters. Prompt tuning, utilizing both mixed and manually crafted prompts with minimal training examples, was also investigated for its computational efficiency. Finally, the research applied prompt engineering through zero-shot and few-shot learning, relying solely on carefully designed instruction prompts for classification without further model training. The comparative analysis aims to pinpoint the most effective and computationally efficient methods for hate speech detection in low-resource languages, contributing to safer online interactions for diverse communities.