Explainable, Evidence-Based Verification of Arabic Claims via Multi-Source Retrieval and Cross-Lingual NLI

We present a training-free, explainable system for verifying Arabic-language claims that combines Arabic Named-Entity Recognition (NER), parallel multi-source evidence retrieval, dense semantic reranking, and cross-lingual Natural Language Inference (NLI) under a single weighted verdict aggregator. Entities are extracted with CAMeLBERT-mix-NER and used to bias a parallel search over trusted Arabic RSS feeds, Google News, a verified-account X (Twitter) endpoint, and DuckDuckGo. Retrieved snippets are reranked by a multilingual-E5 encoder and scored by an XLM-RoBERTa-large checkpoint fine-tuned on XNLI/ANLI; per-source entailment and contradiction probabilities are combined through a weighted aggregator with multiplicatively capped priors over source authority, learned domain reputation, author credibility, and recency. We evaluate on the AraFacts benchmark and make the following contributions, each of which a reader can rely on: (i) a corrected, openly unit-tested aggregator that lets all retrieved evidence—not only official sources—drive the verdict; (ii) a rigorous, reproducible baseline study showing that AraFacts’s natural class imbalance (94% of claims are false-labelled) makes accuracy misleading and that even a well-tuned classical text-only classifier reaches only 0.40 macro-F1; and (iii) an explainable system packaged for deployment as a Streamlit application, a FastAPI service, and a Telegram bot, each exposing a per-source evidence trail. We also document and correct an evaluation error in an earlier version of this work. Code, scripts, and unit tests are released for full reproducibility.