New AI Tool Helps Spot Hate Speech in Roman Urdu
This could help protect Urdu speakers online from online abuse and misinformation.
This paper studies hate speech classification in Roman Urdu, a low-resource language marked by informal grammar, inconsistent sentence structures, and varied spellings. To find effective techniques under limited data, it compares prompt tuning, parameter-efficient fine-tuning with LoRA, and prompt engineering across four experiments: zero-shot LLM inference, PEFT with LoRA, prompt tuning with mixed and manually crafted prompts, and zero-shot and few-shot prompt engineering. The goal is to identify the most effective approach for classifying hate speech in this challenging setting.
- Roman Urdu is the informal Latin-alphabet version of Urdu used by millions online.
- Researchers used cheap, data-efficient AI tricks to detect hate speech in Roman Urdu.
- This could help social media platforms flag abuse faster and protect users.
Why It Matters
Makes online spaces safer for Urdu speakers by catching hate speech earlier and cheaper.