Standalone ontology extraction for first-order logic autoformalization

Develop fine-tuning objectives and prompting strategies that explicitly address Ontology Extraction as a standalone task, adhere to the Ontology Extraction/Logical Translation decomposition, and engage directly with ontology-learning methods and evaluation frameworks.

Background

The paper decomposes FOL autoformalization into Ontology Extraction (OE), which constructs a suitable ontology consisting of a labelled signature and possibly a theory, and Logical Translation (LT), which produces a formula relative to that ontology. Most existing methods instead treat autoformalization as raw translation and do not explicitly address OE.

Because direct metrics for evaluating labelled-signature quality are also absent, it is difficult to optimize or compare ontology-extraction systems independently. The paper therefore identifies the development of OE-specific training and prompting methods, together with closer integration with ontology learning, as an unresolved research agenda.

References

At the methodological level, almost no existing work explicitly targets OE as a standalone task; developing fine-tuning objectives and prompting strategies that adhere to the OE/LT decomposition, and engaging more directly with the ontology learning literature, represents a largely open research agenda.

— Natural Language to First-Order Logic LLM-based Autoformalization  (2610.12030 - Brunello et al., 8 Oct 2026) in Section 6, paragraph “On autoformalization methods”