A Transformer-Driven Bilingual Approach to Arabic Text Summarization

Abstract

This paper presents a novel bilingual approach for Arabic text summarization, addressing the unique academic and technical content’s linguistic challenges. The proposed pipeline solves the main problems associated with Arabic-specific models (e.g. AraBART, AraT5, and mBERT2mBERT) for summarization. Comparative evaluations using the XL-Sum dataset demonstrate that the proposed bilingual approach outperformed all tested models with a ROUGE-2 F1 score of 39.44% versus AraBART’s 33%, the most known one for Arabic text summarization. This research emphasizes the huge performance gap between English-language transformers and their Arabic counterparts, highlighting the potential of bilingual approaches to bridge this gap. By leveraging both English and Arabic models, the proposed pipeline significantly creates a concise and coherent summarization that can be used in many other fields, particularly in academia.