Two Approaches to Ontology Aggregation Based on Axiom Weakening

Daniele Porello, Nicolas Troquard, Rafael Peñaloza, Roberto Confalonieri, Pietro Galliani, Oliver KutzIJCAI 2018 · DOI

TL;DR

This paper tackles the challenge of combining multiple ontologies from different agents into a single, consistent ontology. They propose and compare two novel methods: one that aggregates first and then repairs inconsistencies using "axiom weakening," and another that builds the ontology incrementally through a turn-based negotiation process. Both methods leverage a "reference ontology" derived from agent preferences to guide the repair process, aiming for a more faithful representation of collective knowledge.

What Problem Does This Paper Address?

In today's interconnected world, different agents (e.g., research groups, organizations) often develop their own ontologies to describe specific domains. When these ontologies need to be combined for a larger, shared understanding, inconsistencies frequently arise. This is because each agent might have slightly different perspectives, priorities, or even errors in their individual ontologies. Simply merging them can lead to a logically contradictory knowledge base, rendering it useless. Traditional methods for resolving these inconsistencies, like removing entire axioms, can be too coarse and discard valuable information. This paper addresses the need for more nuanced approaches to ontology aggregation that can handle inconsistencies while preserving as much of the original knowledge and agent preferences as possible.

Methods

The paper introduces two main approaches to ontology aggregation, both relying on a novel technique called axiom weakening:

  • Vote Aggregation with Axiom Weakening:
    1. Aggregation: Agents submit their consistent ontologies. A voting mechanism (e.g., majority rule) is used to select axioms for a preliminary, potentially inconsistent, aggregated ontology.
  • Reference Ontology Computation: To guide the repair, a "reference ontology" is computed. This is done by analyzing agents' preference profiles (rankings of axioms) to find a "collectively best consistent subset" (CBC) of the original agenda. This CBC represents a consensus on important axioms.
  • Repair: If the aggregated ontology is inconsistent, "bad axioms" (those contributing to inconsistency) are identified. These bad axioms are then replaced with weaker versions of themselves, derived using the reference ontology. This process continues until the aggregated ontology is consistent.
  • Turn-Based Negotiation with Axiom Weakening:
    1. Reference Ontology Computation: Similar to the voting approach, a reference ontology is computed first based on agent preferences.
  • Incremental Construction: Agents take turns proposing their most preferred, yet-to-be-treated axiom from their individual ontology.
  • Repair during Negotiation: If adding an agent's proposed axiom would make the collective ontology inconsistent, the agent weakens that axiom (using the reference ontology) until it can be added without causing inconsistency. This process continues until all agents have proposed their chosen axioms.

The core novelty lies in axiom weakening, a fine-grained repair mechanism that replaces an axiom with a logically weaker but still relevant statement, preserving more information than simple axiom removal.

Datasets

The experiments were conducted using 7 ontologies sourced from BioPortal. The paper does not introduce new benchmarks but rather uses existing real-world ontologies to evaluate the proposed methods.

Key Insights and Ideas

  • Axiom Weakening is Key: The paper strongly advocates for axiom weakening as a superior method for repairing inconsistent ontologies compared to coarse methods like axiom removal. It allows for a more faithful representation of agent preferences and collective knowledge.
  • Reference Ontology for Consensus: Deriving a "reference ontology" from agent preferences is crucial for guiding the axiom weakening process. This ensures that the repair is not arbitrary but reflects a collective understanding of axiom importance.
  • Trade-offs between Voting and Turn-Based: The experiments reveal distinct performance characteristics for the two approaches. The voting approach offers better control over which axioms are rejected (leading to higher "strict happiness"), while the turn-based approach might lead to higher "tolerant happiness" by allowing agents to incorporate weaker versions of their preferred axioms.
  • Social Choice Theory Connection: The paper draws parallels to social choice theory, framing ontology aggregation as a collective decision-making problem. This perspective helps in understanding desirable properties like Pareto efficiency and anonymity.

Key Results

The experimental results, visualized in Figure 1 and Table 2, generally confirm the authors' hypotheses:

  • Both the voting and turn-based approaches, using axiom weakening, generally lead to higher agent happiness (both tolerant and strict) compared to baseline methods like "unanimity" (intersection of all axioms) and "removal" (random axiom deletion).
  • The turn-based approach tends to yield higher "tolerant happiness" than the voting approach, suggesting it better incorporates agent preferences even when strict consistency is compromised.
  • The voting approach shows advantages in "strict happiness," likely because it allows a majority to prevent axioms they strongly disagree with from being included, even in their weakened form.
  • For the "pseudo" dataset, no significant difference was observed between the removal and turn-based approaches.

Limitations and Future Work

  • Expert Input for Weakening: While axiom weakening is powerful, selecting the specific weaker axiom can be subjective. The paper acknowledges that in practical applications, expert input might be needed to choose among different weakenings of a given axiom.
  • Parameter Sensitivity: The authors suggest exploring the effect of various parameters, such as the sparsity of agent ontologies, on the relative effectiveness of the approaches.
  • Exploring Different Weakening Operators: Future work could involve investigating a wider range of axiom weakening operators.
  • Negotiation over Axioms: An interesting extension would be to allow agents to negotiate not just on which axioms to include but also on the quality and informativeness of the axioms themselves, potentially leading to more sophisticated aggregation mechanisms.

Summary updated

BibTeX
@inproceedings{porello2018two,
  author = {Daniele Porello and Nicolas Troquard and Rafael Peñaloza and Roberto Confalonieri and Pietro Galliani and Oliver Kutz},
  title = {{Two Approaches to Ontology Aggregation Based on Axiom Weakening}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/268},
}
Figure