Legal Entailment

Legal entailment is a directional relationship. It tests if a premise text (P) logically forces a conclusion hypothesis (H). This is distinct from semantic similarity; 'The defendant breached the contract' entails 'A party failed to perform a duty', but the reverse is not necessarily true. The system must recognize that entailment is asymmetric.

Unlike binary logic, legal NLI typically uses a three-way classification framework:

• Entailment: The hypothesis is necessarily true given the premise.
• Contradiction: The hypothesis is necessarily false given the premise.
• Neutral: The hypothesis may be true or false; the premise provides insufficient information. This schema prevents the model from hallucinating unsupported conclusions.

Standard factual entailment is insufficient for law. Legal entailment must incorporate deontic modalities—obligations, permissions, and prohibitions. A premise stating 'The tenant shall maintain insurance' entails the hypothesis 'The tenant has an obligation to maintain insurance'. The system must distinguish 'shall' (obligation) from 'may' (permission) to avoid catastrophic misreading of statutory duties.

Legal conclusions are rarely absolute. Entailment in law is defeasible—a conclusion can be overturned by exceptions or superior authority. A system must model that 'A contract is valid' is entailed by offer, acceptance, and consideration, but this entailment is defeated by a finding of fraud or incapacity. Static entailment without exception handling produces brittle, unrealistic legal logic.

Unlike standard NLI which operates on a single premise, legal entailment requires multi-document synthesis. A hypothesis like 'The plaintiff is entitled to damages' may require premises aggregated from a statute, two precedents, and a contractual clause. The system must resolve co-references across documents and fuse disparate factual and legal propositions into a unified premise set before computing entailment.

Legal entailment is governed by interpretive canons that act as formal inference rules. For example, expressio unius est exclusio alterius (the expression of one thing excludes others) dictates that a statute listing specific penalties entails the exclusion of unlisted penalties. A robust legal NLI engine must encode these canons as explicit constraints on the entailment decision function.