"t entails h" means a human reading t would conclude h is true. NLI is the task of predicting entailment / contradiction / neutral. Boring on the surface, load-bearing in production.
You built a summarizer. It produced a summary. How do you know the summary does not contain a hallucination?
You built a chatbot. It answered "yes." How do you know the answer is supported by the retrieved passage?
You need to classify 10,000 news articles by topic. You have no training labels. Can you reuse a model?
All three problems reduce to Natural Language Inference. NLI asks: given a premise t and a hypothesis h, is h entailed by t, contradicted, or neutral (unrelated)?
Hallucination check:t = source document, h = summary claim. Not entailment = hallucination.
Grounded QA:t = retrieved passage, h = generated answer. Not entailment = fabrication.
Zero-shot classification:t = document, h = verbalized label ("This is about sports"). Entailment = predicted label.
One task, three production uses. This is why every RAG evaluation framework ships an NLI model under the hood.
Full Access lesson
Keep going. The rest of this lesson is one step away.
Unlock all 503 lessons, video, code, quizzes and every future update. Pay in your own currency with Paystack.