Ontology-Enriched Embeddings

• Ontology-enriched embeddings are vector representations that integrate formal ontological axioms and logical constraints into geometric frameworks.
• They support tasks like semantic similarity, link prediction, and zero-shot reasoning by enforcing deductive structure through tailored loss functions and geometric tests.
• Advanced models such as EmEL(var), Box²EL, and TransBox demonstrate significant improvements in reasoning accuracy and inference robustness over standard embedding approaches.

Ontology-enriched embeddings refer to vector-space representations of entities, concepts, and relations that are constrained or informed by the logical, hierarchical, and semantic structures explicit in ontologies. These approaches are distinguished from general-purpose knowledge graph embeddings by their rigorous encoding of the formal axioms and relationships native to ontological frameworks, especially those expressed in Description Logic (DL) or OWL. The core objective is to produce embeddings that are not only useful for typical machine learning tasks (e.g., semantic similarity, link prediction, search) but are also faithful to the deductive structure and model-theoretic semantics of the underlying ontology.

1. Formal Foundations and Representational Choices

Ontology-enriched embeddings are constructed to encode the intensional (axiom-based) and extensional (instance-driven) structure of ontologies. Let an ontology $\mathcal{O}$ have signature $(\mathcal{C}, \mathcal{R}, \mathcal{I})$ of classes, roles, and individuals, with a set of axioms in a logic such as $\mathcal{EL}^{++}$ or $\mathcal{ALC}$.

A typical embedding $\eta$ consists of:

• A map $$f_\eta: \mathcal{C} \cup \mathcal{R} \to \mathbb{R}^n$$, assigning geometric objects (e.g., centers, vectors, axes) in $\mathbb{R}^n$ to each concept or role.
• For concept regions: $n$-balls $(f_\eta(C), r_\eta(C))$ [Kulmanov et al., 2019; EmEL/EmEL++], axis-aligned boxes (intervals in each dimension) [Box$^2$EL, ELBE, TransBox], or ellipsoids [EIKE].
• For roles: translation vectors [TransE-style, ELEmbeddings] or higher-order regions (e.g., box of translation vectors for many-to-many) [TransBox, Box$^2$EL].

Faithfulness requires that the embedding geometry supports evaluation of logical entailment; i.e., if $\mathcal{O} \vDash \alpha$ (e.g., $C \sqsubseteq D$), then a geometric test (e.g., containment of $B(C)$ in $B(D)$) must hold for the embedding (Chen et al., 2024, Lacerda et al., 2023).

Technical variants address expressivity, e.g.:

• $\mathcal{EL}^{++}$: focuses on $n$-balls/boxes and translation mechanisms to capture existential and conjunctive axioms (Mohapatra et al., 2021, Jackermeier et al., 2023, Yang et al., 2024).
• $\mathcal{ALC}$ and higher: requires lattice-embeddings with saturation and order preservation to encode more complex axiomatics, including conjunction, disjunction, and negation (Zhapa-Camacho et al., 2023).
• Intensional/extensional hybridization: combines geometry with text-based PLMs (e.g., SBERT) for concepts (Wang et al., 2024).

2. Embedding Model Constructions and Semantic Losses

Ontology-enriched embedding models are trained by minimizing losses that enforce geometric analogs of ontological axioms. Consider, for $\mathcal{EL}^{++}$ (Mohapatra et al., 2021, Jackermeier et al., 2023, Yang et al., 2024):

Axiom Type	Geometric Constraint	Loss Function (example)
$A \sqsubseteq B$	$B(A) \subseteq B(B)$	$\max(0,\ \|f(A)-f(B)\| + r(A) - r(B) - \gamma)$
$A \sqsubseteq \exists R.B$	$x \in B(A)\implies x+f(R) \in B(B)$	$$\max(0,\ \|f(A)+f(R)-f(B)\| + r(A) - r(B) - \gamma)$$
$C \sqsubseteq D$ (boxes)	$Box(C)\subseteq Box(D)$	normed excess of lower corners; offset constraints
$C \sqsubseteq \exists r.D$	$Box(C)\oplus b_D \subseteq Head(r)$ ; $Box(D)\oplus b_C \subseteq Tail(r)$	Half-sum of above inclusion losses; regularization on bump
$C\sqcap D \sqsubseteq E$	$Box(C)\cap Box(D)\subseteq Box(E)$	Intersection-composed inclusion loss

Losses are regularized to prevent degenerate embeddings (e.g., norm constraints for class centers) and may include margin parameters. Extension to many-to-many relations (EmEL(var), Box$^2$EL, TransBox) incorporates learned variances (clouds or boxes of translations) to handle non-functional/arbitrary role mappings (Mohapatra et al., 2021, Jackermeier et al., 2023, Yang et al., 2024).

Semantic closure and deductive constraints are enforced by

• Incorporating the deductive closure of the ontology to generate all entailed positives and to avoid sampling negatives that are actually true (Mashkova et al., 2024, Mashkova et al., 2024).
• Negative loss terms applied only to truly false axioms (not entailed or derivable) (Mashkova et al., 2024).
• Joint modeling of instances, concepts, and text semantics in dual-space architectures (e.g., EIKE: ellipsoid+SBERT) (Wang et al., 2024).

3. Advances: Many-to-Many Roles, Complex Concepts, and Hybrid Architectures

Most early ontology embeddings assumed roles are one-to-one (functional); this was recognized as a severe limitation:

• Real ontologies feature roles such as partOf, memberOf, or parentOf, which are fundamentally many-to-many.
• EmEL(var) introduces per-role variances $\sigma_\eta(R)$, interpreting translated regions as “clouds” such that for a fixed head, any point within the variance region may map into several targets (Mohapatra et al., 2021).
• Box$^2$EL and TransBox generalize further by representing roles as regions in translation space (dual boxes or translation boxes), enabling accurate modeling of arbitrary arities for role assertions and role inclusion (Jackermeier et al., 2023, Yang et al., 2024).
• TransBox is explicitly $\mathcal{EL}^{++}$-closed: it supports full composition of logical constructors, including conjunction, existential restriction, and role chain, and is capable of representing arbitrary complex expressions via region-based compositionality.

Intensional/extensional hybrids (e.g., EIKE) and LLM-infused models extend representational power by fusing geometric constraints with textual semantics, capturing both structural (graph/axiom) and linguistic properties (Wang et al., 2024, Ronzano et al., 2024).

4. Experimental Methodologies and Performance

Ontology-enriched embedding models are commonly evaluated on tasks including:

• Subsumption axiom ranking: hiding a portion of $A \sqsubseteq B$ axioms and ranking candidates based on embedding geometry (Mohapatra et al., 2021, Jackermeier et al., 2023, Yang et al., 2024).
• Link prediction: predicting protein–protein or gene–disease associations via embeddings constructed from integrated multi-ontology graphs (Nunes et al., 2021).
• Zero-shot reasoning: matching or inferring properties for unseen or newly introduced classes via compositionality (Akça et al., 4 Dec 2025).
• Alignment: linking entities across ontologies by embedding both into a shared latent space and ranking candidate matches by cosine or other embedding-based similarity (Giglou et al., 30 Sep 2025).

Key metrics include Hits@K, mean/median rank, mean reciprocal rank, AUC, and macro/micro averaging to account for class imbalance.

Empirical findings show consistent superiority of ontology-enriched architectures over pure KG embeddings on reasoning-heavy tasks, with models such as EmEL(var), Box$^2$EL, and TransBox providing substantial improvements in ranking and recall over one-to-one approaches (Mohapatra et al., 2021, Jackermeier et al., 2023, Yang et al., 2024). Deductive closure and improved negative sampling further boost robustness and reduce false negatives (Mashkova et al., 2024, Mashkova et al., 2024).

5. Applications and Integration Scenarios

Ontology-enriched embeddings enable a breadth of applications:

• Knowledge base completion and reasoning—efficient inference over large-scale, richly axiomatized biomedical ontologies (Mashkova et al., 2024, Mashkova et al., 2024).
• Data integration and semantic alignment—cross-ontology mapping and semantic normalization for labor market analytics (Hihn et al., 5 Sep 2025), ontology alignment for interoperability (Giglou et al., 30 Sep 2025).
• Machine learning augmentation—ontology-constrained claim detection (Hüsünbeyi et al., 2024), semantic search, zero-shot and few-shot learning, especially in rapidly evolving label taxonomies (e.g., Gene Ontology) (Akça et al., 4 Dec 2025).
• Hybrid neuro-symbolic systems—integrating LLMs or context-aware textual models with ontological structure, e.g., via contrastive learning on ontology-derived pairs (Ronzano et al., 2024).

Tools such as mOWL provide unified pipelines for training and benchmarking geometric, sequence, and graph-based ontology embeddings (Chen et al., 2024).

6. Limitations, Challenges, and Future Directions

Several challenges persist in ontology-enriched embeddings:

• Full coverage of expressive OWL ontologies, including role negation, number restrictions, and transitive closure, remains difficult for geometric models (Chen et al., 2024, Yang et al., 2024).
• High-dimensional region representations are computationally demanding; closure under intersection becomes fragile in very high dimensions (Yang et al., 2024).
• Evaluation protocols require filtering for deductive closure to ensure that test predictions constitute genuine, logically unentailed knowledge (Mashkova et al., 2024).
• Seamless integration of embedding models with LLMs presents both opportunities and unsolved modeling questions (Ronzano et al., 2024, Akça et al., 4 Dec 2025).

Ongoing work targets:

• Expanding ontology-enriched embeddings to more expressive DL fragments and multi-ontology settings (Yang et al., 2024, Zhapa-Camacho et al., 2023).
• Improving compositionality for complex logical constructs and enabling robust generalization to newly-introduced concepts.
• Hybridizing geometric and neural components for robust, interpretable, and scalable semantic representation (Wang et al., 2024, Ronzano et al., 2024).

7. Representative Benchmark Results

A selection of quantitative results demonstrates the empirical impact of ontology-enriched embeddings:

Model / Dataset	Metric	Baseline	Ontology-Enriched Model	Improvement
EmEL++/GALEN	Hits@1	0.02	EmEL(var): 0.10	×5
EmEL++/SNOMED	Median Rank	87,000	EmEL(var): 43,000	2×
Box$^2$EL/GO	Hits@10	ELBE: 0.05	Box$^2$EL: 0.08	1.6×
TransBox/GO	Med. Rank	>900	TransBox: 30	≫30×
EIKE/YAGO39K	SubClassOf Acc	86.1%	EIKE: 90.45%	+4.3 pp
STAR-GO/GO (zero-shot)	Term AUC	—	STAR-GO: ≈0.90 (13/16 terms best)	Robust to unseen terms

These results exemplify both the technical rigor and practical gains that ontology-enriched embedding frameworks deliver across a spectrum of reasoning and ML tasks (Mohapatra et al., 2021, Jackermeier et al., 2023, Yang et al., 2024, Mashkova et al., 2024, Mashkova et al., 2024, Hihn et al., 5 Sep 2025, Wang et al., 2024, Akça et al., 4 Dec 2025).