Arabic Text Summarization using Transformer Models

Abstract

This work presents an Arabic text summarization system based on Transformer models. The system was developed using Python, Streamlit, PyTorch, and the Hugging Face Transformers library, and supports abstractive, extractive, and hybrid summarization approaches. The experimental evaluation was conducted using Arabic texts from the XL-Sum dataset and two Transformer-based models: mT5 and AraT5. The generated summaries were evaluated using ROUGE metrics and human assessment criteria. The results showed that both models were capable of producing coherent and informative summaries. However, mT5 achieved better overall performance in terms of ROUGE scores, while AraT5 generated summaries faster. These findings demonstrate the effectiveness of Transformer models for Arabic text summarization and highlight the trade-off between summary quality and computational efficiency.