Interaction-based ontology alignment repair with expansion and relaxation

Jérôme EuzenatIJCAI 2017 · DOI

TL;DR

This paper introduces novel methods for agents to automatically repair and improve ontology alignments, which are crucial for inter-agent communication. By allowing agents to adapt their alignments through interaction, expand their knowledge, and even provide less precise answers, the system achieves near-perfect precision and coherence, significantly improving upon previous approaches.

What Problem Does This Paper Address?

In multi-agent systems, agents often use different ontologies to represent knowledge. To communicate effectively, they rely on ontology alignments, which map entities (like classes or properties) between these ontologies. However, these alignments are rarely perfect and can lead to communication failures. This paper addresses the challenge of how agents can autonomously detect and correct these faulty alignments through interaction, aiming to achieve robust and accurate inter-agent communication. Previous work showed that agents could repair alignments through simple games, but the authors found that these methods didn't always achieve perfect results and that previous evaluations might have underestimated their effectiveness.

Methods

The core of this work lies in enhancing the interaction-based repair mechanism. The authors build upon a framework where agents play "games" involving their ontologies and alignments. When a communication fails (an instance is misclassified), agents use adaptation operators to modify the alignments.

The paper introduces several key innovations:

  • New Adaptation Operators: Beyond simple "delete" and "replace," they introduce "refine" (which adds more specific correspondences) and "addjoin" (which adds correspondences to superclasses), and "refadd" (a combination of both). These aim to improve coverage and recall.
  • Expansion Modality: This allows agents to proactively add new, random correspondences when existing ones are removed, especially if they cannot be directly replaced. This helps to fill gaps and potentially discover new correct alignments.
  • Relaxation Modality: Agents are given the option to provide less specific answers (e.g., a broader class) with a certain probability. This is crucial for testing "shadowed" correspondences – those that are not directly tested because a more specific, correct one already handles the instance. By sometimes using less specific answers, agents can uncover and correct faulty shadowed alignments.

These new modalities are designed to satisfy properties like safeness (preventing the same failure from recurring) and entailment (ensuring new correspondences are logically supported), though expansion and relaxation relax these constraints to achieve better results.

Datasets

The experiments were conducted using a simulated environment with abstract ontologies. The "world" is described by Boolean features, and ontologies are strict hierarchies. The paper uses a newly constructed, more comprehensive reference alignment with 86 correspondences across 4 ontologies, an improvement over previous work that used a smaller reference. The experimental setup is randomized and run over 10,000 games, with results averaged over 10 runs.

Key Insights and Ideas

  • Shadowing is a Major Hurdle: The paper highlights that incorrect correspondences can persist and even lead to incoherence because they are "shadowed" by correct ones. The game mechanics, which only trigger repair on failure, don't expose these shadowed errors.
  • Beyond 100% Precision is Possible (and Problematic): Contrary to initial hypotheses, agents didn't always reach 100% precision with basic operators due to shadowing. This led to the development of expansion and relaxation.
  • Relaxation Unlocks Shadowed Correspondences: The relaxation modality, by allowing agents to sometimes be less precise, is key to testing and correcting shadowed alignments, leading to higher precision and coherence.
  • Local Actions Lead to Global Convergence: The agents' local adaptation based on individual game instances, when combined with these new operators and modalities, leads to a global convergence towards correct and coherent alignments.

Key Results

  • The new operators ("refine," "addjoin," "refadd") significantly improve recall and F-measure compared to basic operators like "delete" and "replace."
  • The "refadd" operator, combining refinement and adding correspondences, showed particularly strong performance.
  • The expansion modality increased recall and overall F-measure for all operators, though it slightly decreased precision.
  • The relaxation modality was crucial for achieving 100% precision and full coherence across all operators, though it sometimes decreased recall.
  • The combination of relaxation and expansion with the "refadd" operator achieved the best overall results, showing high precision, recall, and F-measure.
  • Agents consistently converge to a state with 100% success rate, meaning all communication attempts are successful.

Limitations and Future Work

  • Initial Alignment Dependency: The system's performance is still somewhat constrained by the quality of the initial random alignments. The authors suggest that future work could explore agents starting from scratch without any pre-provided correspondences.
  • Operator Selection: Determining which adaptation operator or modality an agent should use in a given situation is a remaining challenge. Future work could involve agents selecting their own operators or the environment selecting agents with specific operator sets.
  • Shadowing and Functionality: The paper notes that even with relaxation, issues like "reverse shadowing" (where a more general correct correspondence prevents a more specific one from being added due to functionality constraints) and the inability to generate enough new correspondences can limit perfect recall.

Summary updated

BibTeX
@inproceedings{euzenat2017interaction,
  author = {Jérôme Euzenat},
  title = {{Interaction-based ontology alignment repair with expansion and relaxation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/27},
}
Figure