Abstract
With the goal to find scalable reasoning approaches, neuro-symbolic techniques have gained significant attention. However, the existing approaches do not take into account the inference capabilities of ontology languages that are based on expressive description logic (such as OWL 2). To fill this gap, we propose two approaches: an ontology-based embedding model for theories in EL++ description logic and a reinforcement learning-based solution for efficient tableau-based reasoning on description logic. We describe promising initial results of our efforts towards these directions and lay down the direction for future work.
Introduction
In the Semantic Web and Artificial Intelligence, "reasoning" is the process of automatically deriving new facts from existing knowledge bases, or ontologies. While traditional logic-based reasoners are mathematically precise, they often struggle to scale when faced with massive, highly expressive ontologies like those used in medicine (e.g., SNOMED CT).
This research identifies a critical gap: neural networks excel at pattern recognition but often ignore the strict constraints of formal logic. Conversely, symbolic reasoners are frequently bogged down by "non-deterministic" choices—situations where the algorithm must branch to explore multiple logical possibilities. This often leads to a "combinatorial explosion" and heavy backtracking. We propose a neuro-symbolic marriage to enhance the scalability of Description Logic (DL) reasoning.
Proposed Methodology
We propose a two-pronged strategy to address these scalability issues: EmEL++ for geometric embeddings and an RL-based framework for tableau reasoning.
1. EmEL++: Ontology Embeddings
This model focuses on the $\mathcal{EL}^{++}$ description logic, which is widely used in biomedical ontologies. It represents logical classes as n-balls (spheres in n-dimensional space) and relations as translational vectors.
Unlike previous models that focused primarily on simple link prediction, EmEL++ defines specific loss functions for every logical construct in $\mathcal{EL}^{++}$. The objective is to minimize these losses such that logical subsumption (e.g., "Class A is a subclass of Class B") is mirrored by geometric containment (the n-ball for A is inside the n-ball for B).
2. Reinforcement Learning for Tableau Reasoning
Tableau algorithms are the standard for DL reasoning, but they rely on building a search tree. When the algorithm encounters non-deterministic rules, it must choose a path; a wrong choice leads to costly backtracking.
We frame this search as a Q-Learning problem. An RL agent navigates the search tree, receiving rewards for reaching a valid conclusion (+10) and penalties for logical contradictions (-10) or backtracking (-1). Over time, the agent learns "confidence scores" to prioritize the most promising paths, effectively pruning the search space.
Evaluation and Results
Datasets
The EmEL++ model was evaluated on four major biomedical ontologies: GALEN (medical terminology), GO (Gene Ontology), ANATOMY, and SNOMED (clinical healthcare terminology). For the RL-based approach, we utilize OWL2Bench, a benchmark designed to stress-test OWL 2 DL reasoners.
Key Results
We compared EmEL++ against the previous state-of-the-art embedding model, ElEm. The results, summarized in the table below, show a significant improvement in accuracy across all datasets and metrics.
In the ANATOMY dataset, EmEL++ improved inference accuracy by nearly 5x compared to the baseline. The geometric approach effectively captures the hierarchy of the ontology by forcing the neural network to respect the full complexity of the $\mathcal{EL}^{++}$ schema.
Conclusion and Future Work
Initial results demonstrate that neuro-symbolic techniques can significantly enhance DL reasoning. While EmEL++ shows strong performance in $\mathcal{EL}^{++}$, future work will focus on extending these geometric techniques to more expressive logics involving negation or universal quantification.
Additionally, the RL-based tableau reasoning framework is currently under implementation. Future efforts will evaluate the agent's ability to generalize confidence scores across different ontology domains and sizes, aiming to prove substantial speedups in large-scale reasoning tasks.
Summary updated