Balancing Efficiency and Accuracy: Lottery Sampling for Arabic Claim Detection

The rapid spread of misinformation highlights the need for automated fact-checking systems, where claim detection serves as a key initial step. However, developing such systems for low-resource languages like Arabic remains challenging due to limited annotated data and computational constraints. This paper proposes a three-stage approach for Arabic claim detection using the Arabic News Stance (ANS) Corpus. First, we evaluate small, medium, and large Large Language Models (LLMs) in zero-shot settings to establish baselines. Then, parameter-efficient fine-tuning with QLoRA is applied to smaller models. Finally, a lottery-inspired adaptive sampling method enhances supervised fine-tuning by focusing on informative samples and preventing overfitting. The best-performing model, LLaMA-۳.۱-۸B with lottery sampling, achieved ۷۴.۷۸% accuracy, surpassing previous methods with lower resource costs. This work demonstrates that medium-scale LLMs with efficient fine-tuning can provide practical, scalable solutions for claim detection in Arabic.

کلیدواژه ها:

Claim detection ، Arabic language ، Large language models ، Small language models ، Lottery Sampling.