Coherence Across Components in Cognitive Systems - One Ontology to Rule Them All

Gregor Behnke, Denis K. Ponomaryov, Marvin R. G. Schiller, Pascal Bercher, Florian Nothdurft, Birte Glimm, Susanne BiundoIJCAI 2015

Introduction

In the development of sophisticated cognitive systems—such as digital health coaches or autonomous assistants—engineers often face a significant architectural hurdle: the "silo" problem. These systems typically comprise multiple specialized components, including planners for decision-making, dialog managers for user interaction, and explanation facilities for transparency. Traditionally, each component operates on its own distinct knowledge model. This paper, originally presented at IJCAI 2015, proposes a solution to this fragmentation by utilizing a single, unified OWL (Web Ontology Language) ontology as a central "brain."

The Challenge of Component Incoherence

When building a cognitive system, the reliance on disparate knowledge models leads to two primary issues: redundancy and incoherence. Redundancy forces domain experts to encode the same information multiple times across different formats (e.g., PDDL for planning and custom scripts for dialog). More critically, incoherence arises when these models diverge. For instance, if a planner categorizes "Skip Rope" as a cardio exercise while the dialog system identifies it as a leg-strength exercise, the system’s output becomes contradictory, undermining user trust.

The authors address this by investigating how specialized models for planning, dialog, and explanation can be automatically derived from a single declarative source of truth.

Methodology: Bridging Logic and Planning

The core of the proposed framework is the use of Description Logics (DL) and OWL to represent domain knowledge. The system integrates these logical foundations with Hierarchical Task Network (HTN) planning, which decomposes high-level goals (e.g., "Improve Fitness") into executable sub-tasks.

The onlysome Construct

A significant technical contribution of this work is the management of the "Open World Assumption" (OWA) inherent in ontologies. In standard logic, the absence of information does not imply falsehood; a reasoner assumes something might exist unless explicitly forbidden. However, planning requires a "Closed World" view to define exactly what a task consists of. To bridge this gap, the authors introduced a specific OWL macro called onlysome. This construct allows the system to define task decompositions strictly, ensuring the planner has a complete and unambiguous set of instructions.

Automated Inference and Reasoning

Rather than requiring human engineers to manually script every possible task decomposition, the framework employs the FaCT++ reasoner to infer new "decomposition methods." For example, if the ontology contains anatomical data stating that "Skip Rope" targets the gastrocnemius muscle, and a "Lower Body Training" objective requires targeting leg muscles, the reasoner can automatically conclude that a workout including skip rope is a valid method for achieving that objective.

Evaluation in the Fitness Domain

The authors validated their approach using a comprehensive fitness training scenario. To ensure a robust knowledge base, they imported 613 concepts from the NCICB (National Cancer Institute Center for Bioinformatics) corpus, providing a standardized taxonomy of human anatomy. The final ontology comprised 1,230 concepts and 2,903 axioms, covering 310 tasks, 9 training objectives, and 24 workout templates.

Results and Key Insights

The evaluation demonstrated that a logic-driven approach is both feasible and efficient:

  • Computational Efficiency: Despite the heavy lifting required by DL reasoning, the system generated the entire extended planning domain in just 3.6 seconds on standard hardware.
  • Knowledge Expansion: The system successfully expanded the initial domain. From a base of 310 tasks, it automatically inferred 203 new workout methods and 59 methods for training objectives, significantly reducing the manual engineering burden.
  • Logic-Driven Explanations: By linking the explanation facility directly to the ontology, the system provides justifications based on logical proofs. Instead of "canned" responses, the system can explain that "Task A is necessary because it warms up Muscle B, which is required for Exercise C."

Limitations and Future Directions

While the framework ensures coherence, the authors noted areas for improvement. The automatically generated explanations can sometimes be overly verbose, requiring further refinement for natural user interaction. Additionally, representing temporal sequences (the order of tasks) remains a challenge in OWL, which is optimized for hierarchical rather than sequential structures. The authors currently use a syntactic encoding workaround but aim to integrate temporal ordering more deeply into the logical reasoning process in future iterations.

Conclusion

By centralizing knowledge within a single ontology, this research demonstrates a path toward more reliable and maintainable cognitive systems. The ability to automatically derive consistent models for planning and communication not only reduces development overhead but also ensures that the system speaks and acts with a single, coherent voice.

Summary updated

BibTeX
@inproceedings{behnke2015coherence,
  author = {Gregor Behnke and Denis K. Ponomaryov and Marvin R. G. Schiller and Pascal Bercher and Florian Nothdurft and Birte Glimm and Susanne Biundo},
  title = {{Coherence Across Components in Cognitive Systems - One Ontology to Rule Them All}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure