Efficient Adaptation of LLMs for Hate Speech Detection in Low-Resource Languages: A Comparative Study on Roman Urdu
Researchers have developed a method to improve the detection of hate speech in Roman Urdu, a low-resource language used by South Asians on social media. They fine-tuned large language models using a technique called Low-Rank Adaptation (LoRA), which showed significant improvement over zero-shot models. The study used a dataset of over 72,000 annotated comments and found that LoRA achieved an F1 score of greater than 0.93, outperforming other transformer models such as Mistral
Researchers have developed a method to improve the detection of hate speech in Roman Urdu, a low-resource language used by South Asians on social media. They fine-tuned large language models using a technique called Low-Rank Adaptation (LoRA), which showed significant improvement over zero-shot models. The study used a dataset of over 72,000 annotated comments and found that LoRA achieved an F1 score of greater than 0.93, outperforming other transformer models such as Mistral, LLaMA, and Falcon.
---
Why it matters: This matters to engineers because it provides a more efficient method for adapting large language models to low-resource languages, which is crucial for detecting hate speech in online platforms where such languages are commonly used.
Source: https://arxiv.org/abs/2608.18142
This article was originally published at: https://arxiv.org/abs/2608.18142