Deontic Textual Entailment

The core mechanism involves classifying the deontic modality of the conclusion relative to the premise. The system must distinguish between three distinct normative statuses:

• Obligation (O): The premise entails that the conclusion is a required action.
• Permission (P): The premise entails that the conclusion is an allowed action.
• Prohibition (F): The premise entails that the conclusion is a forbidden action.

This tripartite classification moves beyond binary entailment to capture the full spectrum of normative force.

DTE benchmarks rigorously test a model's ability to handle contrary-to-duty (CTD) scenarios, which are conditional obligations triggered by a violation of a primary norm.

For example, given the premise 'You must not disclose the data. If you do disclose it, you must encrypt it first,' the system must correctly infer that the obligation to encrypt is only activated upon the violation of the primary prohibition. This tests for non-monotonic reasoning capabilities.

A critical subtask is identifying when a premise contains normative conflicts and determining if a conclusion is entailed despite them. The system must apply resolution strategies:

• Lex Superior: A higher-authority norm overrides a lower one.
• Lex Specialis: A more specific rule overrides a general one.
• Lex Posterior: A later-enacted rule overrides an earlier one.

Failure to resolve conflicts correctly results in contradictory or incorrect entailment judgments.

Advanced DTE systems decompose normative statements into Hohfeldian jural correlatives to disambiguate complex legal prose. The system must map textual statements to the correct analytical pair:

• Right/Duty: A's right correlates to B's duty.
• Privilege/No-Right: A's privilege correlates to B's lack of a right to prevent it.
• Power/Liability: A's power correlates to B's liability to have their legal status changed.

This granular analysis prevents the conflation of distinct normative positions.

DTE models must implement defeasible reasoning, where an otherwise valid entailment can be retracted in the presence of an exception. The system evaluates whether a general rule is defeated by a rebuttal or undercutting condition.

For instance, the premise 'All contracts must be in writing, unless the value is under $500' requires the model to check for the exception before inferring an obligation. This prevents over-generalization from broad normative statements.

A key evaluation metric is the normative faithfulness score, which measures the gap between the deontic content of the source premise and the generated entailment judgment. The system is scored on:

• Precision: Does the conclusion introduce obligations not present in the premise?
• Recall: Does the conclusion capture all applicable duties?
• Hallucination Rate: Are any cited norms fabricated?

This metric ensures the entailment is strictly grounded in the provided text.

The formal logical foundation for reasoning about obligation, permission, and prohibition. Uses modal operators—typically O (it is obligatory that) and P (it is permitted that)—to model ideal normative states. DTE systems rely on these formal semantics to define what constitutes a valid entailment relationship between a premise describing a legal state of affairs and a conclusion asserting a normative consequence.

A conditional obligation that activates when a primary duty has been violated. For example: 'You ought not to cause damage; but if you do, you ought to repair it.' CTD structures are pervasive in legal texts and represent a critical challenge for DTE systems, which must correctly identify that the secondary obligation is entailed only in the non-ideal context of the violation, not in the ideal compliance scenario.

The algorithmic process of resolving situations where two applicable norms prescribe incompatible actions. DTE systems must incorporate resolution strategies to determine which obligation is entailed when norms collide:

• Lex superior: Higher authority prevails
• Lex specialis: Specific rule overrides general
• Lex posterior: Later enactment supersedes earlier

A structured labeling framework for tagging legal corpora with deontic categories. DTE benchmark datasets are constructed using annotation schemas that mark obligation, permission, and prohibition spans, along with their attributes—bearer (who is obligated), counterparty (to whom the duty is owed), trigger condition, and deadline. The quality of DTE model evaluation depends directly on the inter-annotator agreement and semantic precision of these schemas.

A quantitative evaluation score measuring how accurately a generated entailment judgment reflects the deontic content of its source material. Unlike standard NLI accuracy, this metric penalizes:

• Hallucinated obligations: Asserting duties not present in the premise
• Omitted exceptions: Failing to recognize defeasibility conditions
• Scope errors: Misattributing the bearer or counterparty of a norm This metric is essential for validating DTE system outputs in high-stakes legal applications.

A prompt engineering technique that guides language models to explicitly articulate stepwise deontic reasoning before producing an entailment judgment. A typical DTE CoT prompt structures the reasoning as:

1. Identify all normative statements in the premise
2. Classify each as obligation, permission, or prohibition
3. Check for applicable exceptions or CTD structures
4. Resolve any normative conflicts
5. Determine whether the conclusion is entailed This approach significantly improves DTE accuracy on complex legal texts.