high contradiction
medium
low
Privacy: runs locally via Transformers.js (optional). Heuristic fallback needs no download.
Model: not loaded (heuristic ready)
Document (claims extracted as sentences)
—
How it works
- Segments long text into claims (sentence split with Arabic ؟ ، ؛ handling).
- Pairwise similarity: embeddings from
Xenova/multilingual-MiniLM-L12-v2(if enabled) or lexical Jaccard+3gram. Only comparable pairs checked. - Heuristic signals: negation flip (not / لا ..), antonym pair (increase/decrease, مفتوح/مغلق), number mismatch & date mismatch with shared context. Scored 0..1 → low/medium/high.
- Optional NLI verifier:
Xenova/nli-deberta-v3-basere-scores flagged pairs and keeps only contradiction (p > 0.55). Falls back gracefully. - Private by design: embeddings & NLI run in WebAssembly in your browser (Transformers.js). Installable PWA, works offline after model cached.
- Limitations: heuristic may flag near-contradictions (nuance) — review, don't auto-reject. Not a fact-checker (needs external knowledge).
Tip: check policy drafts, LLM outputs for self-contradiction, translation consistency, meeting notes.