TransBox: EL++ Ontology Embedding

• TransBox is an EL++-closed ontology embedding framework that maps OWL concepts and roles to geometric boxes, ensuring closure under complex EL++ constructs.
• It employs explicit geometric operations for conjunction, existential restrictions, and role chains to facilitate robust reasoning over nested axioms.
• Experimental results demonstrate that TransBox outperforms comparable methods in handling many-to-many relations and complex axiom predictions in diverse ontologies.

TransBox is an EL++-closed ontology embedding framework designed to geometrically represent OWL ontologies, with the key property that it can embed all complex expressions in the Description Logic EL++ via explicit geometric composition. TransBox addresses core challenges in ontology representation learning, including closure under EL++ constructs and handling of many-to-one, one-to-many, and many-to-many relations, enabling robust reasoning over atomic and highly complex axioms (Yang et al., 2024).

1. Formal Setting and Problem Definition

OWL ontologies, critical to domains such as healthcare and bioinformatics, encode both relational facts and complex schema through Description Logic. TransBox operates within EL++ (the description logic underpinning OWL 2 EL), whose signatures consist of:

• $\mathcal{N}_C$ = atomic concepts (classes)
• $\mathcal{N}_R$ = roles (relations)
• $\mathcal{N}_I$ = individuals

EL++-concepts are composed via:

$$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$

with TBox axioms (e.g., $C \sqsubseteq D$, $r \sqsubseteq s$, $r_1 \circ r_2 \sqsubseteq t$) and ABox assertions ($A(a)$, $r(a,b)$).

The principal embedding challenge is to find mappings of concepts and roles to geometric entities such that all EL++ expressions—including arbitrary nesting and role chains—are faithfully represented and compositionally closed; that is, the output of applying any EL++ constructor remains representable in the embedding space.

2. TransBox Embedding Space and Geometric Operations

2.1 Embedding Parameterization

• Concepts: Each concept $C$ is mapped to a (possibly degenerate) axis-aligned box $\mathcal{N}_R$0, parameterized by center $\mathcal{N}_R$1 and offset $\mathcal{N}_R$2:

$\mathcal{N}_R$3

• Roles: Each role $\mathcal{N}_R$4 is associated with a translation box $\mathcal{N}_R$5 representing all possible translation vectors:

$\mathcal{N}_R$6

2.2 EL++ Constructors as Geometric Operations

All EL++ concept constructors correspond to closed geometric operations in box space:

• Conjunction (Intersection):

$\mathcal{N}_R$7

• Existential Restriction ($\mathcal{N}_R$8):

$\mathcal{N}_R$9

Result: $\mathcal{N}_I$0 is an axis-aligned box with center $\mathcal{N}_I$1, offset $\mathcal{N}_I$2.

• Role Chain ($\mathcal{N}_I$3):

$\mathcal{N}_I$4

Result: Center $\mathcal{N}_I$5, offset $\mathcal{N}_I$6.

2.3 Closure and Expressivity

TransBox is proven EL++-closed: the box embedding space contains all necessary elements (empty set, $\mathcal{N}_I$7, all singletons) and is closed under intersection, existential role application, and chain composition (Yang et al., 2024). Every (possibly deeply nested) EL++ concept or role can be embedded as a box; hence, subsumption and instance checking for arbitrary expressions reduces to geometric containment operations.

3. Handling Relation Arity and Robustness

TransBox generalizes beyond vector translations (as in TransE) by giving each role an explicit box region, supporting many-to-many, many-to-one, and one-to-many correspondence:

• For $\mathcal{N}_I$8, $\mathcal{N}_I$9 allows $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$0 to link to a set of possible $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$1 within $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$2.
• The volume and orientation of the box regulate role cardinality and directionality.

For high-dimensional robustness, TransBox extends box coordinates to $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$3: each box can have some coordinates that are fully empty, preventing degenerate empty intersections when boxes differ in non-overlapping high-dimensional axes.

4. Training Objectives and Scoring

The scoring function for axioms is closely tied to geometric relationships:

• For $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$4:

$$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$5

Penalties are imposed for any coordinate where $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$6 is non-empty but $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$7 is empty.

• For role chains, the composite box is calculated as above, and scored analogously.
• Negative sampling is used: for a true axiom $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$8 and negative $$C, D ::= \top \mid \bot \mid A \mid C \sqcap D \mid \exists r.C \mid \{a\}$$9, the loss is

$C \sqsubseteq D$0

with regularization to keep centers near the unit sphere.

Zero loss across all axioms implies a sound EL++ model encoded by the boxes.

5. Experimental Results and Comparative Analysis

TransBox has been empirically evaluated on multiple real-world ontologies:

Ontology	Hits@10 (TransBox/BoxEL/ELBE)	AUC (TransBox/BoxEL/ELBE)	Median Rank (TransBox/BoxEL/ELBE)
GALEN	0.05 / 0.01 / 0.00	0.73 / 0.54 / 0.51	727 / 959 / 995
GO	0.41 / 0.01 / 0.09	0.95 / 0.67 / 0.61	30 / 982 / 1035
Anatomy	0.55 / 0.00 / 0.08	0.99 / 0.60 / 0.68	7 / 1020 / 995

TransBox consistently outperforms existing EL++-closed methods for complex axiom prediction, particularly on complex (length 4–10) expressions extracted via random name “forgetting.” On normalized-axiom prediction (atomic or $C \sqsubseteq D$1), it is competitive, though non-EL++-closed Box$C \sqsubseteq D$2EL can slightly surpass it for atomic-only queries.

Ablation confirms that both semantic enhancement (training existential $C \sqsubseteq D$3 as $C \sqsubseteq D$4) and intersection enhancement are necessary for optimal complex axiom performance.

6. Applications, Limitations, and Future Directions

6.1 Use Cases

• Ontology Learning: The EL++-closure property allows embedding of novel complex concepts generated via logical composition, supporting candidate ranking.
• Ontology-Mediated Query Answering: Arbitrary EL++ queries are reducible to box-inclusion checks, allowing direct geometric computation.

6.2 Limitations

• Current method is limited to EL++ (no support for full negation, universal roles, or disjunction); $C \sqsubseteq D$5 is handled but not general $C \sqsubseteq D$6.
• As embedding dimension $C \sqsubseteq D$7 increases, naive boxes become sparse; mitigation via coordinate-dropping is only partially effective.

6.3 Prospective Directions

• Extension to more expressive description logics (e.g., ALC) is necessary for broader OWL support.
• Integration with neural LLMs could allow grounding of textual concept descriptions within the EL++-closed framework while maintaining compositional closure (Yang et al., 2024).

7. Theoretical Properties and Complexity

TransBox is provably sound: zero loss for all axioms in an ontology entails that the induced geometric model exactly satisfies the ontology. Parameter complexity is $C \sqsubseteq D$8 for $C \sqsubseteq D$9 concepts, $r \sqsubseteq s$0 roles, and $r \sqsubseteq s$1 individuals in $r \sqsubseteq s$2.

This geometric approach ensures both expressiveness and tractability for large-scale, compositional EL++ ontology embedding.