Zero- and few-shot Arabic text summarization using retrieval-augmented generation

Summarization of Arabic text is not an easy task due to the complicated morphology of the language and the lack of annotated data. The paper will suggest a Retrieval-Augmented Generation (RAG) model to enhance zero-shot and few-shot Arabic summarization. The approach incorporates external knowledge with the help of multilingual retriever architecture fine-tuned on large Arabic corpora, which allows covering a larger semantic space and enhancing coherence. Experiments indicate that the model is regularly better than traditional sequence-to-sequence baselines with zero-shot gains of +4.2 ROUGE-1 and few-shot gains of +3.5. Ablation experiments illustrate the effects of quality of retrieval on performance. All in all, the approach proves that it is possible to create more informative and factually sound Arabic summaries with the help of RAG under low-resource settings.