Abstract
Ontology mapping is a critical process for enabling semantic interoperability between diverse data schemas. However, traditional mapping tools often rely on local heuristic rules that can lead to structural inconsistencies. This paper proposes a global optimization approach using an Interactive Activation and Competition (IAC) neural network. By modeling mapping hypotheses as nodes and structural constraints as weighted connections, the system identifies a globally optimal set of correspondences. Evaluation on the OAEI 2007 benchmark demonstrates a 19% improvement in F-measure, particularly in scenarios where linguistic information is suppressed.
Introduction
Ontology mapping identifies semantic correspondences between different data schemas to enable systems to communicate effectively. For example, a mapping tool might identify that "Artists" in one database refers to the same concept as "Celebrities" in another.
A significant challenge in this field is that ontologies are not merely lists of terms; they possess complex hierarchical and logical structures. Most existing tools process elements in isolation or use simple "if-then" rules, which frequently results in "crisscross mappings"—where a child element in one ontology is incorrectly mapped to a parent in another. This research addresses the need for a global optimization strategy that can satisfy multiple complex constraints simultaneously.
Methods: The IAC Neural Network Pipeline
The authors propose a multi-stage pipeline centered around an Interactive Activation and Competition (IAC) neural network. The process involves:
- Preprocessing & Initial Similarity: Standard techniques such as edit distance, profile similarity, and structural similarity are used to generate a set of preliminary mapping hypotheses.
- Harmony-based Aggregation: These similarity scores are adaptively combined to favor unambiguous 1-to-1 mappings.
- The IAC Network: This core component models each mapping hypothesis as a node. Connections between nodes represent constraints: positive weights for mutually supporting hypotheses (e.g., matching parents) and negative weights for conflicts (e.g., 1-to-1 mapping violations).
- Activation Dynamics: The network iterates until it reaches a stable state, where the nodes with the highest activation levels represent the final, globally optimized mappings.
Constraint Satisfaction
The system implements 12 specific constraints to guide the network's convergence. These include structural rules (hierarchy and sibling relationships), property-specific rules (domain and range), and OWL-specific axioms (such as owl:sameAs). By treating mapping as a constraint satisfaction problem (CSP), the neural network allows the entire system to "settle" into the most logically consistent configuration, using biological-inspired competition to suppress weak or contradictory mappings.
Experimental Evaluation and Results
The approach was evaluated using the OAEI (Ontology Alignment Evaluation Initiative) 2007 benchmark tests (#248-#266). These tests are considered particularly difficult because they involve "artificially-made" ontologies where linguistic information is discarded, forcing the algorithm to rely almost entirely on structural constraint satisfaction.
The neural network approach provided a dramatic boost over preliminary results across all key metrics. The overall F-measure improved by 19%, driven by significant gains in both precision and recall.
On specific difficult tests like #254 and #262, the F-measure improved by as much as 0.37. This demonstrates that the network excels when structural information is the primary clue available for resolving mapping ambiguities.
Conclusion
By framing ontology mapping as a global constraint satisfaction problem solved via an IAC neural network, this approach effectively resolves structural contradictions that local matchers miss. While highly effective, the authors note that the system can be sensitive to unexpected structural changes in ontologies. Future work will focus on improving computational efficiency through parallelization and exploring the automated discovery of more complex, domain-specific constraints.
Summary updated