Efficient LLM Adaptation Boosts Hate Speech Detection in Low-Resource Languages.

Detecting hate speech in languages with limited digital resources, such as Roman Urdu, presents significant challenges. These difficulties stem from a lack of annotated data, the informal nature of the language, and inconsistent spellings, making it hard for standard Large Language Models (LLMs) to perform effectively. This research focuses on addressing these issues for Roman Urdu, a widely used code-mixed language in South Asia. The study comprehensively evaluated various open-source LLMs, including Mistral, LLaMA, Falcon, and multilingual BERT, for hate speech detection in Roman Urdu. It compared their zero-shot inference capabilities against models fine-tuned using Parameter-Efficient Fine-Tuning (PEFT) with the Low-Rank Adaptation (LoRA) method. Utilizing the PURUTT dataset of over 72,000 annotated comments, the results demonstrated that while zero-shot models performed moderately (F1 = 0.56), fine-tuning even a small fraction of model parameters via PEFT dramatically improved classification performance, achieving an F1 score greater than 0.93. This highlights PEFT's effectiveness and computational efficiency for low-resource language processing.