Abstract
EmELvar is a neurosymbolic reasoning framework that extends geometric ontology embeddings by introducing a learned variance parameter. This innovation allows the model to represent many-to-many relationships within the $\mathcal{EL}^{++}$ description logic, overcoming the limitations of previous one-to-one translation models. By moving beyond rigid mappings, EmELvar provides a geometric foundation for scaling neurosymbolic methods to more complex logical frameworks like $\mathcal{SROIQ}$.
Introduction
In Knowledge Graphs (KGs), embeddings are essential for transforming entities and relations into vectors for downstream tasks like link prediction. However, standard KG embeddings often ignore the formal semantics of the underlying ontology. Recent geometric models, such as EmEl, attempted to address this by representing classes as $n$-balls (spheres in multi-dimensional space) and relations as translations.
A significant limitation of these models is the assumption that relations are one-to-one. In real-world logic, relations are frequently many-to-many; for instance, the relation fatherOf can connect one individual to multiple children. Forcing such relations into a single fixed vector leads to poor reasoning performance and an inability to handle advanced features like cardinality. EmELvar addresses this by introducing "wiggle room" into the vector space.
Methodology
EmELvar evolves the EmEl framework by introducing a learned variance parameter ($\sigma$) for every relation.
Geometric Representation
Classes are represented as $n$-balls defined by a center vector and a scalar radius. This allows for the natural representation of subclass relationships: if Class A is a subclass of Class B, the sphere for A is contained within the sphere for B.
Learned Variance and "Wiggle Room"
To resolve the one-to-one restriction, EmELvar allows translations to land within a "hard bound" distance ($\sigma$) of a target. Instead of a rigid $C + R = D$ requirement, the model targets a volume of space that can encompass multiple related entities simultaneously.
Loss Functions and Regularization
The researchers updated loss functions for existential restrictions (e.g., "Every student must have a teacher") to incorporate this variance. A regularization term was also introduced to prevent the variance from growing infinitely during training, ensuring the model remains constrained to meaningful logical boundaries.
Evaluation
The model was evaluated using datasets from the SemRec 2021 (Semantic Reasoning Evaluation Challenge), specifically ontologies OWL2EL2 through OWL2EL5 based on the $\mathcal{EL}^{++}$ description logic. Before training, these ontologies were normalized into standard forms to ensure consistent application of loss functions.
Subsumption Task Results
The primary evaluation metric was a subsumption task, checking if the model could correctly rank subclasses. While Top 1 hit rates remained modest, the median ranks suggest that the model successfully clusters related classes within the high-dimensional search space.
As shown in the results, performance varied by dataset. On OWL2EL3, the model achieved a Top 100 hit rate of 0.15. For OWL2EL4 and OWL2EL5, the median ranks (124.6 and 139.6, respectively) indicate that related classes are being placed relatively close to one another despite the difficulty of the task.
Discussion and Future Work
The core insight of EmELvar is that many-to-many relationships can be modeled geometrically by treating relations as "fuzzy" translations targeting a volume rather than a point. This approach treats reasoning as a distance-based ranking problem.
However, limitations remain. The low Top 1 hit rates highlight the gap between approximate neurosymbolic reasoners and traditional symbolic reasoners. Furthermore, some datasets lacked the necessary subclass relations for a full evaluation.
Future work will focus on extending this variance-based approach to handle more complex properties, such as cardinality (e.g., "a car has exactly 4 wheels") and transitivity. The ultimate goal is to support the full $\mathcal{SROIQ}$ logic, which underpins the OWL 2 DL standard.
Summary updated