Semantic Heterogeneity

Occurs when different ontologies use different terms for the same concept or identical terms for different concepts. This includes synonymy (e.g., 'automobile' vs. 'car'), polysemy (e.g., 'bank' as a financial institution vs. river bank), and multilingual variations. String similarity metrics like edit distance and Jaccard coefficient serve as primary matchers to quantify textual likeness, but lexical overlap alone is insufficient for reliable alignment without structural or semantic context.

Arises when the same domain is modeled using different ontological commitments and axiomatization patterns. One ontology may represent a concept as a class, while another models it as an instance or property. Hierarchical depth varies—one taxonomy may be flat with few levels, another deeply nested. Tree edit distance quantifies the minimum operations needed to transform one hierarchical structure into another, providing a measurable signal for structural matching algorithms.

Occurs when ontologies cover overlapping but non-identical domains with different granularity levels. A medical ontology may model 'Heart' with detailed anatomical subclasses, while a pharmaceutical ontology treats it as a simple target organ. This granularity gap creates partial correspondences rather than exact equivalence. Upper ontologies like BFO mitigate this by providing shared high-level categories—continuants and occurrents—that serve as integration hubs for domain-specific extensions.

Stems from differences in expressivity and formalism between ontology languages. An RDFS vocabulary supports only class hierarchies and property domains, while an OWL 2 DL ontology enables complex class expressions with existential restrictions and disjointness axioms. Aligning across these expressivity gaps requires careful handling—mappings that are logically sound in a weaker formalism may introduce unsatisfiable classes when merged with a more expressive ontology, necessitating alignment repair techniques.

Reflects differences in design intent and application context even when formal semantics align. Two e-commerce ontologies may both define 'Price' but one includes tax and shipping while the other represents only the base cost. These implicit modeling assumptions are rarely documented in axioms, making them the hardest heterogeneity type to detect automatically. Resolution often requires domain expert intervention or analysis of instance-level data distributions to infer intended semantics.

The challenge of determining whether two entity references denote the same real-world object across knowledge graphs. The owl:sameAs property asserts exact identity, but misuse is rampant—studies show up to 20% of sameAs links in the Linked Open Data cloud are erroneous. Conservativity principle violations occur when mappings introduce unintended subsumptions, while alignment coherence measures evaluate logical consistency by checking for disjointness violations in the merged ontology.

The post-matching process of detecting and removing incoherent or logically inconsistent correspondences from a generated alignment. It restores the satisfiability of the merged ontology by enforcing the Conservativity Principle.

• Detects violations like disjointness clashes
• Uses tools like LogMap for logic-based repair
• Ensures the merged knowledge graph does not contain unsatisfiable classes

The process of mapping concepts across ontologies labeled in different natural languages. This technique bridges the lexical gap by leveraging multilingual embeddings or pivot languages to identify semantic equivalence despite surface-form divergence.

• Uses cross-lingual word embeddings (e.g., MUSE, LASER)
• Critical for global knowledge graph integration
• Often employs machine translation as a preprocessing step

A neural technique that learns low-dimensional vector representations for entities across different knowledge graphs in a unified space. Geometric proximity in this embedding space indicates semantic equivalence, enabling alignment without explicit lexical overlap.

• Uses models like TransE, RotatE, or GCNs
• Requires seed alignment pairs for training
• Effective for structurally heterogeneous graphs