NeurOWL uses AI to find gaps in formal knowledge

On July 17, 2026, a team from the University of Manchester and Zhejiang University introduced NeurOWL, a system designed to locate missing knowledge in formal ontologies. It combines a logical reasoner, numerical concept representations and language models to answer a specific question: when a relationship is absent from a knowledge base, is it really false, or is a rule missing that could justify it?

A concept map with gaps

An ontology is more than a list of words. It is a collection of formal statements about concepts and their relationships: for example, that one category belongs under another or that an entity can have a certain property. The OWL standard represents that knowledge in a form from which software can draw logical consequences. It is particularly useful in domains with large vocabularies, including biomedicine and food data.

Those structures are rarely complete. Knowledge changes, expert curation takes time and some connections have simply not been written down. A conventional reasoner can prove what follows from the available rules, but it cannot establish a relationship that depends on a missing axiom. NeurOWL does not turn a guess into a fact. It tries to decide whether a guess is semantically plausible and suggest which missing statement should be reviewed.

Logic first, language second

The system works in three stages. First, it uses a description-logic reasoner to test whether a relationship can already be derived from the ontology. If not, it looks for intermediate concepts that might bridge the two terms. Embedding models rank candidates by semantic proximity, and a language model validates the possible links.

When that route does not find a useful bridge, the method broadens the search using semantic information from embeddings. Its output is not just yes or no. It tries to return potentially missing axioms and a justification combining those proposals with rules that were already present. That traceability matters because an AI suggestion should not enter an ontology used by other applications without review.

What the authors tested

The preprint evaluates NeurOWL on deliberately incomplete versions of two real ontologies: FoodOn, for food and agriculture, and SNOMED CT, a major clinical terminology. The authors removed selected rules to create cases where a relationship could be valid in the original ontology but was no longer provable in the pruned version. They also built negative cases to distinguish a genuine gap from an incorrect relationship.

In those tests, the paper reports F1 scores of up to 96.8 for assessing the plausibility of subsumption relationships, and scores of up to 89.3 and 79.3 on its explanation metrics, depending on the setting. Those are high figures, but they describe an evaluation environment built from known ontologies. They do not yet show that the system will discover new, correct knowledge in any real-world database.

A review aid, not a machine that rewrites science

The practical appeal is clear. In a health ontology, an omitted relationship may affect search, record integration or decision-support tools. In a food ontology, it can make it harder to connect information on ingredients, production and nutrition. A system that prioritises connections worth reviewing could save experts time.

The decisive step remains human. Text labels for two concepts may sound close while a formal relationship is still wrong, too broad or unsuitable in a clinical context. NeurOWL offers an interesting combination of symbolic rigour and linguistic capability, but specialists who understand the domain and the consequences of adding an axiom must assess its proposals. Its value lies in making the hypothesis and its justification explicit, rather than hiding them behind a confident-looking answer.

Three states, not two

In a description-logic ontology, failure to prove a relationship does not automatically make it false. It may be false, information may be missing, or the relationship may fall outside what the available rules can express. That distinction, related to OWL’s open-world assumption, is why ontology completion is not the same as filling empty cells. A proposal should remain explicitly “not entailed” until sufficient evidence exists.

Abduction adds another layer. It starts with a conclusion to be explained and searches for axioms that, combined with existing knowledge, would make it derivable. Several explanations may be logically valid. The one that sounds most natural is not necessarily best: an axiom may explain the local case while being so broad that it introduces many unwanted consequences. Review must test both the immediate explanation and everything it changes elsewhere.

Testing a suggestion before accepting it

The first step is to run a reasoner on a copy and check that the proposed axiom produces the intended consequence without inconsistency. The second is to seek counterexamples: sibling concepts that the same rule would connect even though domain knowledge says otherwise. The third is to review definitions, synonyms and provenance with specialists; linguistic similarity can join terms that belong to different levels or contexts.

The fourth step is to record the decision. The suggestion, NeurOWL version, axioms used, explanation, reviewer and reason for acceptance or rejection should stay together. A later review can then distinguish validated knowledge from an automatic suggestion. The preprint produces explanations containing potentially missing axioms; an operational implementation also needs this governance history.

The study removed rules from known ontologies and asked whether the system could reconstruct relationships whose original answer was available. That design measures recovery in a controlled setting. It does not reproduce every difficulty of discovering a genuinely new relationship: the source ontology may contain errors, external text may be ambiguous and specialists may disagree. A high reconstruction score is evidence about the method, not authority to edit a clinical terminology.

Separating language, logic and domain evidence

A team can organise review in three columns. The language column asks why names and definitions seem connected. The logic column shows exactly which axioms yield the conclusion. The domain column provides scientific or professional evidence authorising the relationship. If only the first is filled, there is similarity; if the first two are filled, there is a formal hypothesis; only the third provides a basis for curating it as knowledge.

The transferable skill is to treat every suggested relationship as a versioned hypothesis. Before adding it, demand the logical explanation, global consequences, domain evidence and accountable decision-maker. NeurOWL may shorten the search; validity comes from the process that turns a suggestion into a reviewed axiom.

Teams should also protect the stability of systems reusing the ontology. Adding one axiom can change derived classifications, queries and mappings without touching those systems. A proposal should pass through a candidate version, competency tests expressing expected questions and a readable change note. If an unintended consequence appears, the axiom can be withdrawn before it reaches records or decisions. Curation does not end when a relationship is accepted; it includes checking which downstream consumers inherit its effects. This is especially important when a terminology update crosses organisational boundaries and the original reviewer cannot see every application.