Improving Arabic Relational Aggregated Search Using a Hybrid RAG and AraBERT Framework

Aggregated search is challenged by data heterogeneity, redundancy, and irrelevant information, particularly in Arabic because of its rich morphology and dialectal diversity. This study proposes an AI-driven framework to enhance Arabic aggregated search by integrating Retrieval-Augmented Generation (RAG) with semantic clustering. The framework employs RAG to retrieve relevant information from multiple search verticals, including web pages, images, news articles, and videos, and to generate contextually enriched textual representations. These representations are encoded with AraBERT to produce 768-dimensional semantic embeddings, which are then compressed into compact latent feature vectors through a stacked autoencoder. K-Means clustering is then applied to organize the compressed representations into semantically coherent topic groups. The proposed framework was evaluated on nine Arabic datasets spanning five application domains: education, sports, information technology, healthy food, and weather. Experimental results demonstrate that the proposed approach increases the average silhouette score from 0.59 to 0.68, representing a 15% improvement over clustering based on raw aggregated search results. Ablation experiments further indicate that both Retrieval-Augmented Generation and stacked autoencoder–based feature compression independently contribute to the observed performance gains. These findings demonstrate the effectiveness of integrating Retrieval-Augmented Generation with semantic representation learning to improve the organization and clustering of Arabic aggregated search results.

@article{Sedhom2026,
  title     = {Improving Arabic Relational Aggregated Search Using a Hybrid RAG and AraBERT Framework},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Sara Saad Sedhom and Ahmed Younes and Islam Elkabani and Ashraf Elsayed},
  doi       = {10.14569/IJACSA.2026.0170726},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170726}
}