Finding Ontological Correspondences for a Domain-Independent Natural Language Dialog Agent

Hamid Haidarian Shahri, Donald PerlisAAAI 2008

Introduction

Interacting with complex systems, such as GPS units or specialized robotics like Mars rovers, often requires users to master rigid, artificial interfaces. To move toward more intuitive, natural language interaction, systems must overcome the "vocabulary gap"—the discrepancy between human terminology and a machine's internal logic. This paper presents a novel approach to achieving semantic convergence through automated ontological correspondence.

The Semantic Convergence Problem

The core challenge lies in mapping a "User Ontology" (the agent's model of human concepts) to a "Robot Ontology" (the device's specific command structure). For a domain-independent agent to function across various applications without manual rule-writing, it must automatically determine when two differently named concepts—such as "magazine" and "journal"—refer to the same entity.

A Multi-Metric Approach to Concept Matching

The authors propose a concept-matching algorithm that calculates similarity scores by aggregating four distinct metrics. This approach allows the agent to look beyond simple string matching and understand the functional role of a concept within its hierarchy.

Architecture of the Natural Language Dialog Agent system showing the interaction between user and robot ontologies.

The four metrics utilized are:

  1. Lexical Similarity: Measures string distance (e.g., Jaro-Winkler) between concept names.
  2. Extensional Similarity: Analyzes the overlap of specific instances or "individuals" belonging to the concepts.
  3. Extensional Closure Similarity: A broader metric including instances of all subclasses, capturing the full "reach" of a concept.
  4. Global Path Similarity: Compares the structural position of concepts by examining their hierarchical paths from the ontology root.

By employing an ontology reasoner (Pellet), the system can interpret relationships like subClassOf that are invisible to standard text-processing tools.

Evaluating Performance

The algorithm was tested using two independent bibliographic ontologies (Karlsruhe and INRIA), containing 24 known corresponding concepts. The results highlight the limitations of relying solely on lexical matching. While string similarity provides a baseline, it cannot achieve high recall without introducing significant noise.

Precision, Recall, and F1 scores across different similarity metrics, showing the performance boost of the multi-metric approach.

As shown in the evaluation, incorporating structural and extensional metrics (Global Path) significantly improves Recall and the overall F1 score while maintaining high Precision. This confirms that structural context can resolve ambiguities where names differ but functional roles align.

The Cost of Semantic Reasoning

A critical finding of the research is the computational trade-off involved in "thinking" versus "reading." Lexical matching is highly efficient, taking approximately 172 ms. However, metrics requiring an ontology reasoner to traverse hierarchies and retrieve instances increase computation time to roughly 1100 ms. This order-of-magnitude jump represents a significant bottleneck for real-time dialog systems.

Conclusion and Future Directions

The multi-metric approach successfully bridges the vocabulary gap by leveraging the rich structural information inherent in ontologies. While the current system relies on experimental similarity thresholds, future iterations could utilize machine learning to optimize these parameters automatically. By providing a framework for "tentative opinions," the agent can engage in clarification dialogues with users, further refining its understanding of human intent in complex domains.

Summary updated

BibTeX
@inproceedings{shahri2008finding,
  author = {Hamid Haidarian Shahri and Donald Perlis},
  title = {{Finding Ontological Correspondences for a Domain-Independent Natural Language Dialog Agent}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure