When an AI system diagnoses a patient, flags a financial anomaly, or calculates an optimal delivery route, it often relies on rigorous mathematical frameworks—such as formal logic, automated theorem proving, or graph-based reasoning. These methods provide mathematical guarantees of correctness. However, the raw outputs they generate (such as proof trees, minimal unsatisfiable sets, or dense automata traces) are completely unreadable to humans.
To bridge this gap, a growing body of research focuses on formal-to-natural language translation. The core insight is that we can leverage the rigid, invariant structure of a formal proof and map its components directly to linguistic templates or controlled natural language. This preserves the mathematical soundness of the underlying system while matching human cognitive constraints.
The Gap Between Rigor and Readability
Automated reasoning systems excel at finding exact solutions, but their internal representations are optimized for computation, not comprehension. As Chunxi Ji argues in On the Computation of Necessary and Sufficient Explanations, while formal logical explanations provide mathematical rigor, they fail to be useful in practice without a mechanism to translate these logical reasons into natural language or intuitive visualizations.
This friction manifests across several domains:
- Automated Theorem Proving: Systems like Vampire can successfully find contradictions in complex logic puzzles, but the resulting optimized symbolic outputs are too dense for human experts to verify manually, as highlighted by Jeremy Avigad et al. in An Impossible Asylum.
- Software Verification: Formal semantic analysis yields low-level symbolic constraints that software engineers cannot easily interpret or act upon when resolving code conflicts (Anita Sarma et al. in Development context driven change awareness and analysis framework).
- Symbolic Regression: Even mathematically concise models, such as the evolved tree networks designed by William G. La Cava et al. in Learning concise representations for regression by evolving networks of trees, still require significant domain knowledge to decode unless they are transformed into linguistically grounded representations.
How Structural Mapping Works
Why does formal-to-natural language translation work? It works because formal logic structures—like trees, paths, and chains—contain explicit causal dependencies. Translation algorithms exploit these dependencies by replacing abstract predicates with human-centric semantics while maintaining the structural skeleton of the proof.
[Formal Proof Tree] ───► [Structural Mapping] ───► [Controlled Natural Language]
(A ∧ B → C) Map predicates to nouns "Because condition A and B
and implications to verbs are met, outcome C follows."
For example, Jan Noessner et al. in Interactive Data Integration with MappingAssistant demonstrate how formal proof trees can be systematically converted into natural language sentences. This allows non-technical domain experts to audit data integration rules and identify faulty mappings without understanding the underlying description logic.
Similarly, in biomedical query answering, Umut Öztok et al. in Generating Explanations for Complex Biomedical Queries leverage Answer Set Programming (ASP) to find the shortest chain of logical rules required to justify a query. By translating these raw logical chains into a controlled natural language, complex multi-database interactions are presented as readable sentences that maintain strict scientific rigor.
Application Across High-Stakes Domains
The necessity of this translation layer spans across several critical fields where black-box explanations are insufficient and formal verification is required.
AI Planning and Model Reconciliation
In human-robot interaction, an AI agent must explain why its plan is optimal or why a human's proposed plan might fail. Stylianos Loukas Vasileiou et al. have extensively explored this in classical, hybrid, and probabilistic planning scenarios (A Logic-based Explanation Generation Framework for Classical and Hybrid Planning Problems (Extended Abstract); On Generating Monolithic and Model Reconciling Explanations in Probabilistic Scenarios).
As Stylianos Loukas Vasileiou et al. note in On Generating Monolithic and Model Reconciling Explanations in Probabilistic Scenarios (Abstract Reprint), while minimal correction sets can mathematically reconcile a user's mental model with the AI's model, these mathematical proofs must be mapped to natural language to achieve actual user understanding.
Clinical Decision Support
When treating patients with multiple chronic conditions, clinical guidelines often offer conflicting advice. Frameworks that combine medical models with formal argumentation can resolve these conflicts mathematically. However, Kristijonas Čyras et al. in Argumentation for Explainable Reasoning with Conflicting Medical Recommendations emphasize that the resulting logical graphs are too technical for direct use by clinicians. Translating these formal symbolic proofs into natural language is required to make decision-support tools viable in clinical communication.
Graph-Based Recommendations and Routing
Explanations derived from structural paths offer clear audit trails, but they must be verbalized to be useful. In educational recommendation systems, Jibril Frej et al. in Finding Paths for Explainable MOOC Recommendation: A Learner Perspective show that while policy-guided path reasoning over knowledge graphs provides transparent reasoning paths, converting these raw graph paths into fluid natural language is essential for general learners.
Similarly, for physical routing, Aaron Schild et al. in Why is My Route Different Today? An Algorithm for Explaining Route Selection identify the exact traffic-heavy road segments causing a detour, but note that a gap remains between identifying these causal factors and presenting them as human-interpretable justifications.
Designing Better Intermediate Representations
A major focus of current research is designing intermediate representations that sit between raw logic and natural language. If the intermediate representation is too close to the formal logic, it lacks flexibility; if it is too close to free-form natural language, it risks losing logical precision.
For instance, Shuai Zhang et al. argue in InfeRE: Step-by-Step Regex Generation via Chain of Inference that intermediate steps derived purely from formal parsing trees lack the interpretability of human-centric intermediate representations.
To address this, researchers are exploring frameworks like probabilistic argumentation. As Francesca Toni et al. establish in Understanding ProbLog as Probabilistic Argumentation, formalizing the link between logic programming and argumentation provides a structural foundation for explanations. The ultimate goal, the authors note, is to use these argument trees as a stepping stone to generate interactive, natural language dialogues.
By treating natural language generation not as a separate post-processing step, but as a structured translation of formal logic, this research direction ensures that AI explanations can be both mathematically sound and humanly comprehensible.
Key Foundational Work
Tailoring Explanations for the User (1985) — contributed 1 concept elements to this direction.
An Integration of Resolution and Natural Deduction Theorem Proving (1986) — contributed 1 concept elements to this direction.