Attention Mapping for Hallucination Reduction in Arabic Financial Using Retrieval-Augmented Generation Systems

This research introduces a customized attention visualization framework for mitigating hallucinations in Arabic Retrieval-Augmented Generation (RAG) systems tailored for financial document analysis. The proposed architecture extends MarBERT with a dual-stage attention supervision mechanism and a hallucination-aware loss formulation, trained on a newly constructed dataset of 7,000 annotated Arabic financial query-context pairs. A grounding alignment score is computed over attended tokens, and generated responses are rejected when falling below a dynamically adjusted precision-aware threshold,serving as the core decision-making approach for hallucination detection. The system achieves 95.04% classification accuracy, 95.84% precision, 94.12% recall, and an F1 score of 94.97%, outperforming AraELECTRA and mBERT-RAG baselines. For hallucination detection, it reaches an 87.63% detection rate with only 8.91% false alarms and a detection latency of 3.2 tokens, showcasing reliability in identifying ungrounded financial claims. Visualizations using cross-attention matrices and token contribution scores reveal consistent alignment with critical Arabic financial terms, temporal expressions, and entity references. An interactive attention interface further empowers non-technical users to identify hallucinations with 83.4% success in usability tests. These results validate the effectiveness of integrating interpretable attention modeling with domain-specific retrieval, offering a scalable approach for risk-sensitive natural language processing in low-resource languages.