NeurOWL: An LLM-Based Neural-symbolic Framework for Incomplete OWL Ontology Reasoning

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "NeurOWL: An LLM-Based Neural-symbolic Framework for Incomplete OWL Ontology Reasoning".

Jane: Here's the gist of the paper. The gist The NeurOWL framework is an end-to-end neuro-symbolic system designed to verify and abduce missing axioms in incomplete OWL ontologies by integrating LLMs with ontology embeddings, which addresses the challenge of plausibility reasoning without requiring predefined hypothesis candidates.

Tom: Problem and Motivation Ontologies expressed in the Web Ontology Language (OWL) provide a formal framework for representing structured knowledge and enable automated reasoning through Description Logics (DL) Owing to their precise and shareable representation of machine-interpretable semantics, OWL ontologies are widely used in knowledge-intensive domains such as biomedicine (Ashburner et al.

Jane: 2000) and healthcare (Bos and Donnelly 2006).

Tom: A fundamental reasoning task in ontology engineering is subsumption reasoning: determining whether a subsumption relationship holds between two concepts and providing explanations for such entailments For complete and consistent ontologies, subsumption can be determined by standard reasoners (Glimm et al.

Title and authors: Tom: So we’re looking at a paper called "NeurOWL: An LLM-Based Neural-symbolic Framework for Incomplete OWL Ontology Reasoning". Jane, what’s the core idea here in plain terms?

Jane: Basically, it tackles the problem of checking if a specific relationship between two concepts in an incomplete knowledge base actually holds true. It uses a combination of neural networks and formal logic to do that verification and also figure out what's missing.

Tom: That’s the key part, right? It’s not just asking if something is true or false, but if it’s plausible, and if so, suggesting the exact rules needed to make it true.

Lu: What I find interesting is that they frame this as a problem of finding missing axioms H such that when you add them to an incomplete ontology O', the target subsumption alpha becomes entailed in O' union H >

Meng: So, instead of just saying "this is wrong," the system tries to give you the actual missing pieces needed. That sounds more useful for building real systems.

Lalam: I see it as bridging the gap between formal ontology structures and the flexible understanding of language models, which is something we’ve been trying to do across our own work too.

Tom: Exactly. It’s integrating those two very different worlds—the strict logic and the fuzzy understanding from language models—into one system to handle incomplete knowledge bases.

Jane: It sounds like they're moving beyond just predicting what *might* be true, toward actually finding the minimal set of assumptions that *could* make it true.

The paper's summary: Tom: Okay, so the summary of NeurOWL says they propose an end-to-end neuro-symbolic framework for explainable subsumption reasoning over incomplete OWL ontologies. It supports both positive and negative entailment scenarios and doesn't require a predefined hypothesis space.

Jane: So, it’s designed to be flexible; it doesn't need you to hand it a list of things it can guess missing axioms from, which is a big deal because most previous methods required that fixed vocabulary.

Lu: They do this by automatically identifying meaningful missing components that are both compatible with the existing information and plausible in real-world contexts by integrating textual information with formal semantics through LLMs and ontology embeddings.

Meng: So it’s not just using the LLM to guess randomly; they are grounding those guesses in the structure of the ontology itself using those embeddings. That sounds like a way to make the suggestions more reliable.

Lalam: I think that’s where their approach differs from some other work; they are systematically searching for intermediate concepts C such that A is a subclass of C and C is a subclass of B, which can be used to construct missing axioms and explanations if no direct connection is found.

Tom: That search process sounds like it gives you a roadmap for finding the missing links rather than just throwing random axioms at the wall.

Jane: And they mention that by iteratively applying NeurOWL, you can recover any subset- and entailment-minimal solutions as a set of missing axioms, which is really powerful because it lets you refine your results step by step.

The paper's improvements: Tom: The authors highlight several improvements. First, they introduce a generalized TBox abduction task that doesn't require predefined missing axioms candidates and handles cases where the given subsumption is incorrect.

Jane: So it’s more robust than just checking if a statement is entailed; it can flag an assertion as potentially wrong and then suggest how to fix it.

Lu: They also propose NeurOWL itself, which is this end-to-end neuro-symbolic framework utilizing ontology embedding and LLM. They call their embedding method OnT, which they extend from Yang et al. two thousand twenty-five by introducing role embeddings as rotations and additional loss functions to encode logical properties <ref:2607.15776#pg2>.

Meng: I’m interested in the practical side of that OnT approach; if you fine-tune it on a specific incomplete ontology, does that give you a significant performance boost over using it out of the box?

Tom: Yeah, they show empirical results demonstrating effectiveness and robustness under varying degrees of incompleteness on datasets like FoodOn and Gene Ontology. The main result is that NeurOWL achieves strong performance in concept subsumption reasoning while also generating meaningful explanations for incomplete input ontologies.

Lalam: They also show how to optimize this by fine-tuning the embedding model on the specific incomplete ontology, demonstrating that it can adapt to the underlying logical structure through fine-tuning on complex datasets like Snomed∃.

Conclusion: Tom: So, wrapping up, NeurOWL is this end-to-end neuro-symbolic framework for explainable subsumption reasoning over incomplete OWL ontologies that jointly performs verification and abduction without needing a predefined hypothesis space.

Jane: It’s powerful because it automatically finds meaningful missing components compatible with existing information and plausible in real-world contexts, which means it doesn't just check for simple entailment; it infers necessary knowledge for those incomplete ontologies.

Lu: Plus, the iterative application of NeurOWL lets you recover any subset- and entailment-minimal solutions as a set of missing axioms, allowing the AI to progressively refine its understanding by recovering the weakest set sufficient to derive the target conclusion.

Meng: From an engineering standpoint, they show that using different stages in their framework—like Stage 2a or 2b—can be substantially higher than just using Stage 3a alone, which helps us focus computational effort on what actually matters for specific tasks <ref:2607.15776#pg2>.

Lalam: I just want to say that this work shows how we can use these LLMs and embeddings together not just for prediction, but for actual explanation in a way that’s grounded in the formal structure of the ontology.