Abstract
Knowledge Graph (KG) embeddings often fail to capture the formal logic defined in underlying ontologies. This paper introduces EmEL++, a geometric embedding model that maps the $\mathcal{EL}^{++}$ description logic into a vector space. By representing classes as $n$-dimensional balls and relations as translations, EmEL++ facilitates logical reasoning through geometric operations. Experimental results on large-scale ontologies like SNOMED CT and GALEN demonstrate that EmEL++ significantly outperforms traditional KG embeddings and previous ontology-specific models in subsumption reasoning tasks.
Introduction
Traditional Knowledge Graph embedding models, such as TransE and DistMult, are effective at link prediction but remain "logic-blind." They treat entities as points in space, ignoring the hierarchical and relational constraints defined by Description Logics (DL). For instance, if an ontology specifies that "every Surgeon is a Doctor," a standard embedding might not preserve this relationship unless it is explicitly and frequently linked in the data. EmEL++ addresses this by treating logic as geometry, allowing for efficient reasoning within high-dimensional vector spaces.
Methodology: Geometric Interpretation of Logic
The core innovation of EmEL++ is the representation of ontological constructs as geometric entities.
Classes as $n$-balls
Unlike point-based embeddings, EmEL++ represents each class $C$ as an $n$-ball (a sphere in $n$-dimensional space) defined by a center vector $\vec{e}_c$ and a radius $r_c$. This "volume" allows for a natural representation of class hierarchies: a broad class can have a large radius, while a specific subclass has a smaller radius contained within it.
Subsumption and Relations
The model uses specific geometric constraints to enforce logical axioms:
- Subsumption ($C \sqsubseteq D$): Represented as containment. The ball for class $C$ must reside entirely within the ball for class $D$.
- Disjointness ($C \sqcap D \sqsubseteq \perp$): Represented as separation between the two balls.
- Relations (Roles): Following the translation-based approach, a relation $R$ is a vector $\vec{R}$. For an axiom like $C \sqsubseteq \exists R.D$, the model attempts to translate the center of ball $C$ by $\vec{R}$ such that it lands within ball $D$.
Complex Role Axioms
EmEL++ extends previous models by incorporating loss functions for complex features:
- Role Inclusions ($R \sqsubseteq S$): Ensuring the translation vector for a sub-role (e.g., son of) aligns with its parent role (e.g., child of).
- Role Chains ($R_1 \circ R_2 \sqsubseteq S$): Ensuring the composition of two translations (e.g., part of $\circ$ part of) aligns with the resulting role vector.
Experimental Evaluation
The model was evaluated on four diverse ontologies: SNOMED CT (clinical terminology), Anatomy (phenotype-gene links), Gene Ontology (GO), and GALEN (complex clinical information). The primary task was subsumption reasoning—predicting whether one class is a subclass of another.
Results
EmEL++ demonstrated superior performance compared to both traditional KG embeddings and the previous state-of-the-art model, EL Embeddings (ElEm).
- Subsumption Accuracy: On the GALEN dataset, which is rich in complex role axioms, TransE achieved a Hits@10 of 0.00, whereas EmEL++ reached 0.11.
- Geometric Consistency: EmEL++ showed significant improvements in "Geometric Accuracy"—the physical containment of balls as required by the logic. For the Anatomy dataset, this metric improved from 0.09 (ElEm) to 0.48 (EmEL++).
Conclusion
EmEL++ bridges the gap between symbolic logic and vector space embeddings. By providing a geometric interpretation of $\mathcal{EL}^{++}$, it enables fast, $O(n)$ reasoning on massive datasets. While it does not yet match the absolute accuracy of symbolic reasoners like ELK, it offers a scalable approach for integrating formal logic into machine learning workflows. Future work will explore extending this framework to more expressive logics like $\mathcal{SROIQ}$.
Summary updated