Linguistic Features Embedded in Arabic Transformers: A Feature Ablation Study for E-Commerce Deepfake Detection
Year: 2026
Type: article
Abstract: Traditionally, Arabic natural language processing pipelines apply aggressive preprocessing, including removing punctuation, normalizing characters, and eliminating emojis, based on the assumption that cleaner inputs yield better model performance. This study investigates the linguistic features embedded in Arabic transformer models for detecting synthetic product listings in e-commerce environments. We conducted a systematic ablation study to evaluate the impact of removing individual linguistic feature categories on the classification performance of four Arabic transformer models. Our results show that preprocessing Arabic e-commerce text is not simply a matter of removing noise: the punctuation functions as a consistent structural signal across all architectures, numeric characters and English text behave as model-dependent specification-carrying signals, and emojis have no measurable effect across the models. These findings provide a practical framework for understanding the linguistic features embedded in Arabic transformers for e-commerce deepfake detection and suggest that preprocessing pipelines should preserve domain-relevant structures, rather than aggressively applying uniform cleaning.
DOI: 10.64589/juri/231620
Language: English
Topic: Spam and Phishing Detection
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7233844303
