When attempting to integrate disparate knowledge bases, databases, or web schemas, the traditional approach is to compare them directly. Automated systems typically look for lexical overlap (e.g., matching the string "Renal_Failure" to "Kidney_Failure") or structural similarity (e.g., identifying isomorphic subgraphs in two taxonomies).
However, direct matching frequently hits a performance ceiling. As documented over years of benchmarking by the Ontology Alignment Evaluation Initiative (OAEI), traditional schema matching algorithms struggle when target ontologies exhibit low lexical overlap or highly divergent hierarchical structures (see Manel Achichi et al. et al., Results of the Ontology Alignment Evaluation Initiative 2017; Alsayed Algergawy et al., Results of the Ontology Alignment Evaluation Initiative 2019; and Ian Horrocks et al., Results of the Ontology Alignment Evaluation Initiative 2020). If two ontologies use entirely different vocabularies to describe the same domain, direct comparison heuristics are fundamentally blind to their relationships.
To break through this performance plateau, researchers developed indirect or mediated ontology alignment. Instead of comparing two target ontologies in isolation, this paradigm introduces a third, external resource—a mediator ontology, central hub, or background knowledge base—to act as a semantic bridge.
The Core Intuition: The Semantic Rosetta Stone
Imagine two people who do not share a language and do not possess a bilingual dictionary. Direct communication is impossible. However, if both individuals happen to speak a common third language, that intermediate language can serve as a pivot.
In ontology alignment, large-scale, highly structured background knowledge bases (such as WordNet, SNOMED CT, or BioPortal) serve as this intermediate language.
Ontology A (Source) <--- [Anchoring] ---> Mediator Ontology <--- [Anchoring] ---> Ontology B (Target)
\ | /
\ [Internal Structure] /
\ | /
+----------------------------> [Derived Alignment] <---------------------------+
Direct matching fails because it assumes the source and target schemas contain all the semantic information necessary to deduce a relationship. Mediated alignment acknowledges that schemas are often sparse or idiosyncratic. By projecting the source and target concepts onto a dense, pre-existing web of external knowledge, the system can discover hidden connections that are completely absent from the input datasets.
The Two-Step Mechanism: Anchoring and Derivation
As formalized by Zharko Aleksovski et al. in Exploiting the Structure of Background Knowledge Used in Ontology Matching, the process of mediated alignment operates in two distinct phases:
- The Anchoring Phase: Concepts from both the source and target ontologies are independently mapped to the external reference ontology. Because the reference ontology is typically massive and rich in synonyms, establishing these initial "anchors" is much easier than finding direct matches between two narrow, specialized schemas.
- The Derivation Phase: Once the anchors are established, the system traverses the internal semantic relationships (such as hypernymy, hyponymy, or part-of relations) within the reference ontology. By tracing these paths, the system infers new, indirect links between the original source and target concepts.
For example, if Ontology A contains the term Cardiopathy and Ontology B contains Heart Condition, a direct string matcher might fail. However, if both can be anchored to a central reference ontology where Cardiopathy is explicitly defined as a subclass or synonym of Heart Disease, the system can transitively compose a high-confidence mapping between them.
This transitive composition dramatically boosts recall. Ernesto Jiménez-Ruiz et al. demonstrated this in Extending an ontology alignment system with BioPortal: a preliminary analysis, showing that using a massive repository of biomedical ontologies as an automated "oracle" allows matching systems to resolve borderline cases that direct heuristics miss.
Why It Works: Overcoming the Sparsity Problem
Mediated matching has become a cornerstone of state-of-the-art alignment frameworks like the LogMap family (Ernesto Jiménez-Ruiz et al., LogMap family results for OAEI 2014; Ernesto Jiménez-Ruiz et al., LogMap family results for OAEI 2015; Ernesto Jiménez-Ruiz et al., LogMap family participation in the OAEI 2017) and RiMOM (Xiao Zhang et al., RiMOM Results for OAEI 2008). It succeeds because it shifts the problem from string comparison to topological placement.
1. Cross-Resource Lexical Pivoting
In computational linguistics, direct mapping between highly specialized formal logic systems and colloquial lexical databases is notoriously difficult. Clifton James McFate in Expanding Verb Coverage in Cyc with VerbNet leveraged WordNet as a central pivot to bridge this gap, mapping synsets from an intermediate layer to transfer knowledge between disparate hierarchies. Similarly, Michael Matuschek in High Performance Word Sense Alignment by Joint Modeling of Sense Distance and Gloss Similarity utilized a central structural pivot to align word senses across entirely different dictionaries, proving that structural distance within a mediator can validate textual similarities.
2. Shared Coordinate Projections
Modern deep learning approaches translate this topological intuition into vector spaces. In the MultiOM framework (Weizhuo Li et al., Multi-view Embedding for Biomedical Ontology Matching), an external reference ontology acts as a geometric anchor. By projecting the representations of two separate target ontologies into a shared coordinate system dictated by the mediator's structure, the system bypasses the "sparsity problem." Rare or poorly defined entities inherit geometric context from the mediator, preventing misalignments.
3. Contextual Logic and Query Relaxation
Rather than forcing a single global truth that might break the internal consistency of local schemas, mediation allows for localized, context-aware translation. Systems can use "bridge rules" to link local logic spaces through a central reference terminology (Frank van Harmelen et al., Using C-OWL for the alignment and merging of medical ontologies).
This is highly effective for applications like query relaxation. When a user queries a local medical knowledge base using colloquial or non-standard terms, mapping the local instance to a comprehensive external ontology like SNOMED CT allows the system to leverage external hierarchical relationships, expanding the query to find semantically relevant answers even when the exact terms are missing locally (Chuan Lei et al., Expanding Query Answers on Medical Knowledge Bases; Elena Zotova et al., Towards the integration of WordNet into ClinIDMap).
4. Extensional Mediation
Mediation does not always have to rely on abstract concepts; it can also happen through shared data populations. For instance, Konstantin Todorov et al. in Multimedia ontology matching by using visual and textual modalities demonstrated that shared instances—such as images and descriptions associated with concepts—can act as the mediating layer. This extensional approach bridges different knowledge schemas by identifying commonalities in the underlying data they describe, using multi-modal text and vision to prune away false matches.
By shifting the alignment task from a direct, brittle comparison to a mediated, two-step journey through a rich background knowledge base, these systems turn a lack of direct evidence into an asset, leveraging the collective intelligence of the web to bind isolated schemas together.
Key Foundational Work
Ontology Alignment: An annotated Bibliography (2005) — contributed 1 concept elements to this direction.
Using C-OWL for the alignment and merging of medical ontologies (2004) — contributed 1 concept elements to this direction.