Abstract
Reference reconciliation is the critical task of identifying whether different data records refer to the same real-world entity. LN2R is a hybrid system that combines logical reasoning (L2R) with numerical similarity propagation (N2R). By leveraging ontological constraints and the interconnected nature of data, LN2R achieves high precision and recall. This paper details the system's architecture and its performance on the OAEI 2010 benchmark.
Introduction
In the Semantic Web, data is often distributed across multiple sources, leading to an "identity crisis" where the same entity is described differently. For instance, one source might refer to "Le Louvre" while another uses "muse du Louvre." Traditional string-matching methods often fail to capture the semantic context or the complex relationships between entities. LN2R addresses this by treating reference reconciliation as a global problem influenced by domain-specific rules (ontologies) and data dependencies.
Methodology: The LN2R Architecture
LN2R operates through two integrated components: L2R and N2R, which handle certainties and uncertainties respectively.
L2R: The Logical Engine
The L2R component translates ontology constraints into first-order logic (Horn rules). It utilizes properties such as class disjointness and property functionality to infer certain matches and non-matches. For example, if an ontology specifies that "Museum" and "City" are disjoint, L2R ensures that a museum instance is never reconciled with a city instance. This logical pruning significantly reduces the search space and provides high-precision "anchors" for the numerical engine.
N2R: The Numerical Engine
N2R handles the "fuzzy" matching by modeling similarity as a system of non-linear equations. Unlike local matching methods, N2R propagates similarity scores through the data graph. The similarity between two entities is influenced by the similarity of their attributes and related entities. This process is iterative, continuing until the similarity scores converge.
Evaluation and Results
LN2R was evaluated using the OAEI 2010 Instance Matching track. The results demonstrate the system's robustness across different domains:
- Person1: Achieved a perfect 100% F-measure (100% precision and 100% recall).
- Person2: Achieved a 93% F-measure, maintaining a very high precision of 99.4%.
- Restaurant: Achieved a 75.3% F-measure.
The authors noted that the lower score in the restaurant dataset was partially due to inaccuracies in the benchmark's gold standard, where LN2R correctly identified matches that were missing from the official ground truth.
Conclusion
LN2R represents a powerful hybrid approach to entity resolution. By treating ontologies as a source of truth rather than just schemas, it achieves near-perfect precision on benchmark datasets. Future work will focus on improving scalability for larger datasets through distributed inference and reducing the dependency on manual pre-alignment of schemas.
Summary updated