Ontology alignment is the process of discovering correspondences between different knowledge schemas—for example, realizing that "Staff" in one database is the same as "Employee" in another. To do this, researchers typically rely on two signals: lexical similarity (the names look the same) and structural topology (the concepts are located in the same part of the hierarchy).
While structural matching is often presented as a robust, independent way to find mappings, a recurring finding across nearly two decades of research is that structure is not a standalone signal. Instead, structural analysis acts as a refinement layer that is fundamentally dependent on "anchors" established by initial lexical matches.
The Dependency on Lexical Anchors
The intuition behind structural matching is that if two nodes are equivalent, their neighbors should be equivalent too. However, to start this comparison, a system needs a "seed" or an anchor. Quentin Reul in the KOSIMap framework describes lexical similarity as the critical foundation that provides these initial anchors. Without them, structural and logical features have no starting points to establish reliable correspondences.
This dependency creates a "cold start" problem for structural matchers. As noted by Gang Wang and the NBJLM team, if a literal matcher fails to recognize a concept due to specialized vocabulary or naming variations, the structural matcher lacks the necessary context to evaluate where that node sits in the hierarchy. The graph, in isolation, is often too ambiguous to solve.
The Fragility of Weakly Informative Ontologies
The reliance on lexical cues becomes most obvious when those cues are removed. In the X-SOM framework, Carlo Curino and Giorgio Orsi observed a sharp decline in precision and recall when concept names were randomized. If structural matching were truly independent, systems would still be able to navigate the "topology" of the data. Instead, performance collapses, suggesting that structural matchers are often just propagating the certainty of lexical matches rather than discovering new semantic truths.
This challenge is formalized in the concept of weakly informative ontologies. Wenyu Wang, developing the Lily system, points out that many real-world ontologies lack descriptive labels or comments entirely. In these cases, traditional matching strategies struggle because the "semantic subgraphs" they rely on are essentially blank. Even advanced systems like SEMA, which uses Latent Dirichlet Allocation (LDA) to find hidden meanings, see a significant drop in recall when all lexical data is removed, as Vassilis Spiliopoulos and colleagues demonstrated.
Root Equivalence and Logical Inference
The problem is even more acute when using automated reasoners to find hierarchical relationships (like "subclass of"). Xi Chen 0003 argues that structural alignment is often "brittle" because it depends on identifying a primary root equivalence between top-level concepts.
Think of it like trying to align two family trees. If you can't be sure that the "Grandparents" in Tree A are the same as the "Ancestors" in Tree B, you cannot use logic to infer that their children are cousins. If the initial base link (the anchor) is missed, the system fails to derive any subsequent hierarchical relationships, leading to a total failure in mapping discovery.
Hybrid Strategies as a Solution
To overcome this, modern systems treat alignment as a multi-stage pipeline where structure and lexicon are tightly coupled:
- Lexical Anchoring: Systems like RiMOM by Juanzi Li and Jie Tang first estimate the "vibe" of the data to decide how much to trust linguistic vs. structural cues.
- Structural Propagation: Once anchors are found, similarity is "spread" to neighbors. Uthayasanker Thayasivam and the Optima+ team use structural "height" to partition these tasks, but they acknowledge that the effectiveness of this logic remains highly sensitive to the quality of those initial labels.
- Consistency Checking: Finally, logic-based reasoners (like those in KOSIMap) prune matches that violate structural constraints.
The consensus across these papers—from HMatch to SYNTHESIS—is that while structure provides the "logic" of an ontology, the "meaning" remains locked in the lexical labels. Without the lexical key to anchor the process, the structural engine has nothing to turn.