AraXLM: Evaluating Arabic Diacritization Tools for Cross-Language Plagiarism Detection

In recent years, plagiarism detection systems have evolved from basic lexical matching and n-gram overlap methods to Deep Learning (DL) models capable of capturing semantic relationships between texts.While these DL-based approaches have achieved notable success across various languages, their effectiveness in Arabic remains limited due to inherent linguistic ambiguities, particularly the omission of diacritical marks.This absence hinders accurate semantic interpretation and limits the ability of models to detect paraphrased or semantically obfuscated content in Arabic texts.This paper presents an evaluation of Arabic Text Diacritization (ATD) tools as the initial phase of a plagiarism detection framework designed for Arabic-English cross-lingual model text analysis (AraXLM).It describes the first stage of the framework, which focuses on assessing the performance of state-of-the-art ATD tools.An empirical analysis was conducted on six ATD models using Word Error Rate (WER), Diacritic Error Rate (DER), both with and without case endings (CE), and Bilingual Evaluation Understudy (BLEU) metrics.The results show that tools such as Shakkelha produced lower DER and high BLEU values, indicating high accuracy in diacritic restoration, while Fine-Tashkeel demonstrates the lowest WER and highest BLEU, reflecting best word-level performance.In contrast, CAMeL Tools and Mishkal display comparatively higher error rates across both metrics.These findings suggest that incorporating accurate diacritization models into Arabic NLP tasks, such as Machine Translation (MT) and Plagiarism Detection (PD), improves text normalisation and the quality of semantic embeddings.Thus, the AraXLM framework, supported by effective diacritization pre-processing, enhances linguistically aware detection of plagiarism involving Arabic text, where precise semantic alignment between languages is essential.