AROMA Results for OAEI 2009

Jérôme DavidOM 2009

TL;DR

This paper presents AROMA, an ontology alignment system that leverages association rules and statistical interestingness measures. For OAEI 2009, AROMA was enhanced with an equality-based matcher to improve efficiency and recall, particularly on the anatomy track, demonstrating competitive results on benchmark and anatomy tasks.

What Problem Does This Paper Address?

Ontology alignment is the task of identifying correspondences between entities (classes and properties) in different ontologies. This is crucial for enabling semantic interoperability between diverse knowledge bases. However, achieving high precision and recall in this task is challenging due to variations in terminology, structure, and the inherent ambiguity of natural language descriptions. Existing methods often struggle with different types of alterations and can be computationally expensive. This paper addresses the need for robust and efficient ontology alignment methods, specifically focusing on improving recall and efficiency for real-world ontologies.

Methods

AROMA employs a hybrid, extensional, and asymmetric approach to ontology alignment. Its core methodology involves three stages:

  1. Pre-processing: Entities (classes and properties) are represented by a set of terms extracted from their annotations and descriptions. These terms are also associated with ancestor entities to preserve the hierarchical structure.
  2. Association Rule Discovery: The system discovers association rules between entity term sets using a statistical interestingness measure called "implication intensity." These rules are interpreted as potential subsumption relations.
  3. Post-processing: This stage cleans and enhances the discovered alignment. It includes deducing equivalences, removing cycles and redundant correspondences, ensuring an injective alignment (one-to-one mapping), and finally, enhancing the alignment using a string similarity-based matcher.

For OAEI 2009, AROMA introduced a significant modification: an equality-based matcher. This matcher considers two entities equivalent if they share at least one annotation. This is applied to unaligned pairs before the more computationally intensive string similarity matcher. The string similarity matcher itself was also refined, using a different weighting scheme for Jaro-Winkler similarity and a new collection measure to aggregate similarities between entity annotations.

Datasets

The paper evaluates AROMA on several tracks of the OAEI (Ontology Alignment Evaluation Initiative) 2009 competition:

  • Benchmark: A standard set of ontologies for evaluating alignment systems.
  • Anatomy: A track involving biomedical ontologies, known for their complexity and real-world relevance.
  • Conference: A track involving ontologies from the domain of academic conferences.

The authors note that AROMA did not participate in the multilingual library track due to its reliance on syntactical data without multilingual resources, nor the instance matching track as it is not designed for that task.

Key Insights and Ideas

  • Association Rules as a Foundation: Using association rules with an implication intensity measure provides a principled way to infer semantic relationships (subsumption) from textual descriptions, even with some "counter-examples" (imperfect matches).
  • The Power of Equality Matching: Introducing an equality-based matcher as a preliminary step significantly improves efficiency by quickly identifying obvious matches, reducing the load on subsequent, more complex matching stages. This is particularly beneficial for real-world ontologies with rich annotations.
  • Annotation Aggregation Matters: The choice of how to aggregate similarity scores across multiple annotations of an entity is critical. The paper highlights a shift to a new collection measure that favors the highest similarity pairs, aiming to better capture nuanced relationships.
  • Trade-offs in Precision and Recall: The paper acknowledges that returning both equivalence and subsumption correspondences can negatively influence precision. This suggests that for certain applications, a focus on semantic precision and recall might be more informative.

Key Results

  • Benchmark: AROMA achieved good precision values, particularly on tests with structural alterations. The modifications made for OAEI 2009 led to an enhancement on the 3xx tests, improving precision from 82% to 85% and recall from 71% to 78%.
  • Anatomy: AROMA demonstrated good performance on the anatomy track, running quickly (around 1 minute) due to its subsumption relation pruning. The enhanced recall on this track is attributed to the new equality-based matcher.
  • Efficiency: The new equality-based matcher improved time efficiency by reducing the number of unaligned entities before the terminological matcher.

Limitations and Future Work

  • Directory Track: AROMA struggled with very small sub-directories in the directory track, as its statistical method requires a substantial amount of textual data. The authors suggest reintroducing larger taxonomies for future editions.
  • Oriented Matching Track: The authors did not participate in the oriented matching track, deeming the proposed reference alignments incomplete and potentially misleading. They suggest improvements to the track design, such as removing certain classes/properties to create more naturally derivable subsumption relations.
  • Multilingual and Instance Matching: AROMA is not designed for multilingual alignment or instance matching, representing clear limitations for broader applications.
  • Future Work: The paper suggests evaluating results using semantic precision and recall, which could provide a more nuanced understanding of AROMA's performance, especially given its output of both equivalence and subsumption correspondences.

Summary updated

BibTeX
@inproceedings{david2009aroma,
  author = {Jérôme David},
  title = {{AROMA Results for OAEI 2009}},
  booktitle = {OM},
  year = {2009},
}
Figure