Case study · 02
Haqiqa
Agentic Arabic News Verification
Haqiqa extracts claims, searches for supporting evidence, ranks sources, evaluates evidence, and produces reasoned verification results with citations. It combines an agentic verification workflow with evaluated Arabic news classification models.
- Arabic NLP
- Agentic AI
- Machine Learning
Problem
Arabic news verification requires more than classifying text in isolation. Useful results need evidence, source comparison, reasoning, and citations that a reader can inspect.
Agentic workflow
The system uses a LangGraph-based five-stage verification workflow. Each stage transforms or evaluates evidence before returning a final decision.
- Claim extraction
- Web search
- Evidence ranking
- Evidence verification
- Final decision with reasoning and citations
Evaluated machine-learning models
Earlier model development used a dataset of 13,750 Arabic news articles. AraBERT reached 96.22% F1, 96.22% accuracy, and 99.57% AUC. XGBoost reached 94.50% F1, 94.51% accuracy, and 98.94% AUC.
Architecture
The web application connects a Next.js interface to a FastAPI backend and LangGraph workflow. It includes web search, semantic evidence retrieval, authentication, persistence, and a Telegram interface.
Related research
Arabic NLP · AI Evaluation · Intelligent Systems