LDM: Link Discovery Method for new Resource Integration

Nathalie Pernelle, Fatiha SaïsRED@ESWC 2011 · OpenReview

Introduction

The Linked Open Data (LOD) cloud is a vast, decentralized repository of structured data. However, its utility is often hampered by extreme heterogeneity. Data providers frequently describe identical real-world entities using disparate schemas (ontologies) and rarely provide explicit links, such as owl:same-as. Integrating this data typically requires two distinct steps: ontology alignment and reference reconciliation. The Link Discovery Method (LDM) addresses the inherent "chicken-and-egg" problem where effective alignment requires reconciled data, and effective reconciliation requires aligned schemas.

Methodology: The Two-System Approach

LDM is an unsupervised approach that models similarities using two communicating, non-linear equation systems. This architecture allows similarity scores to propagate bi-directionally between the conceptual (schema) level and the instance (data) level.

  1. Conceptual Equation System: This system calculates similarity between pairs of concepts (e.g., "Article" vs. "Publication") by analyzing their labels, structural positions in the hierarchy, and the similarity of their constituent data instances.
  2. Instance Level Equation System: This system determines similarity between data references (e.g., specific authors or papers) based on shared attribute values and the similarity of the concepts to which they belong.

The workflow begins with an initialization phase using standard alignment tools to establish baseline lexical similarities. Subsequently, the algorithm enters an iterative resolution phase. It calculates instance similarities, updates concept similarities based on those results, and feeds the new concept scores back into the instance-level equations. This "ping-pong" process continues until the similarity scores reach a mathematical fix-point.

Experimental Setup

The authors validated LDM using an extract of scientific publication data from DBLP. They compared a local RDF dataset containing concepts like Articles, Conferences, and Persons against a corresponding extract from the LOD cloud. This setup served as a running example to demonstrate the mathematical convergence and the method's ability to handle distinct schemas.

Ontology schema for the DBLP dataset used to illustrate link discovery between heterogeneous sources.

A primary contribution of this work is the demonstration of mathematical convergence. In the DBLP example, the system reached a global fix-point in only three macro-resolutions using a precision threshold of $\epsilon = 0.05$.

Table showing the iterative convergence of similarity scores (x1 to x6) over three macro-resolutions.

The iterative process significantly boosted initial similarity scores. For instance, starting from low lexical match scores (0.1 to 0.3), the system successfully inferred correct links with high confidence. With a reconciliation threshold of 0.45, LDM identified owl:same-as links for papers (0.66), authors (0.62), and conferences (0.47) across the two schemas, even without deep hierarchical structures.

Conclusion and Future Work

LDM provides a robust theoretical framework for simultaneous schema and instance alignment. By exploiting mutual reinforcement, it overcomes the limitations of siloed integration methods. Future work will focus on large-scale empirical evaluations, handling non-equivalent properties, and introducing a third layer for direct property-to-property similarity computation.

Summary updated

BibTeX
@inproceedings{pernelle2011ldm,
  author = {Nathalie Pernelle and Fatiha Saïs},
  title = {{LDM: Link Discovery Method for new Resource Integration}},
  booktitle = {RED@ESWC},
  year = {2011},
  url = {https://openreview.net/forum?id=ScDPoomjaJH},
}
Figure