Most data integration tools operate on a simple assumption: if two databases or ontologies describe the same world, their components must have a one-to-one relationship. We map "Author" in one system to "Writer" in another and call it a day. However, real-world knowledge is rarely that symmetrical.
As noted by Alsayed Algergawy and colleagues in the OAEI 2019 report, the field of ontology alignment has hit a performance plateau because traditional algorithms struggle with "complex" correspondences—relationships where one concept in a source corresponds to a logical combination, transformation, or filtered subset of concepts in the target.
The Expressivity Gap
The core challenge is that different schemas model the same information at different levels of granularity. A single database column might represent a concatenation of three different ontology properties, or a specific class in one system might only be equivalent to a subset of a class in another, defined by a specific attribute filter.
Heiner Stuckenschmidt and Livia Predoiu argue that simple functional mappings cannot bridge this semantic gap. In Learning Complex Ontology Alignments, they highlight that effective integration requires discovering logical rules—such as identifying that a "Professor" in one system is equivalent to the union of "Associate Professor" and "Full Professor" in another. Without the ability to handle these 1:n or m:n relationships, automated systems suffer from low recall and fail to achieve true interoperability.
Bridging with Logic and Lattices
To move beyond simple string matching, researchers have introduced mediating structures that can house complex axioms. Pascal Hitzler and Luciano Serafini propose using "bridge ontologies" or Distributed First-Order Logic (DFOL) to represent non-symmetrical relationships like subsumption. This framework, detailed in Integrated View and Comparison of Alignment Semantics, allows for a rigorous mathematical comparison of how different mapping languages compose and merge knowledge.
Another powerful mechanism for discovering these non-atomic links is Formal Concept Analysis (FCA). Instead of comparing entities in isolation, systems like FCA-Map, developed by Songmao Zhang and Brynn Zhao, construct concept lattices. These lattices allow the system to identify "one-to-group" mappings by analyzing the shared lexical and structural contexts of entities. As shown in their OAEI 2016 results, this approach is particularly effective in complex domains like biomedicine, where different ontologies may represent anatomical symmetry or granularity in distinct ways.
From Manual Hints to Automated Discovery
Historically, these complex mappings were a manual bottleneck. Tools like MAPONTO, introduced by Yuan An and John Mylopoulos, attempted to ease this by taking simple user "hints" and automatically generating the complex logic formulas needed to bridge databases and ontologies.
The modern frontier involves moving toward fully automated discovery using machine learning. Recent work has shifted toward:
- Inductive Logic Programming (ILP): Learning the actual rules that define how property values in one system relate to classes in another.
- Representation Learning: Systems like BioHAN, proposed by Peng Wang, use Graph Attention Networks to aggregate features from distant neighbors in a hierarchy, helping to resolve structural heterogeneity.
- Large Language Models: The LLMATCH framework by Sha Wang leverages the reasoning capabilities of LLMs to perform "Rollup and Drilldown" strategies, managing many-to-many correspondences across multiple interrelated tables in enterprise environments.
By treating alignment as a problem of logical and structural synthesis rather than simple label matching, these approaches allow for the integration of data that was previously siloed by its own complexity.