TL;DR
This paper introduces an "analogy ontology" that formally represents the concepts and relationships involved in analogical reasoning, based on Gentner's structure-mapping theory. This ontology acts as a bridge, enabling AI systems to seamlessly integrate powerful analogical reasoning with traditional first-principles reasoning, leading to more flexible and human-like AI.
What Problem Does This Paper Address?
Many AI systems rely on "first-principles reasoning," which uses logical rules and axioms to derive conclusions. While effective, these systems often lack the flexibility and human-like intuition that comes from drawing analogies. Conversely, existing "case-based reasoning" systems, which are designed to learn from past examples, often use simplistic descriptions and struggle to integrate deeply with logical reasoning. The challenge is to create AI systems that can effectively combine the rigor of logical deduction with the creative problem-solving power of analogy, especially when dealing with complex, large-scale knowledge bases.
Methods
The core contribution is the analogy ontology, a formal, declarative representation of analogical reasoning concepts. This ontology is grounded in structure-mapping theory, a psychological model of how humans understand similarity and analogy. The ontology defines key entities like:
- Cases: Collections of statements representing situations or problems.
- Matches: The result of comparing two cases (a base and a target).
- Mappings: The specific correspondences and inferred relationships between the base and target cases.
- Correspondences: Alignments between individual items (entities, expressions) in the base and target.
- Candidate Inferences: New information about the target case that is "projected" from the base case via the analogy.
The semantics of this ontology are enforced using procedural attachment. This means that when the reasoning system encounters a query involving these ontological terms, it triggers specialized analogical processing software (like the Structure-Mapping Engine - SME, and MAC/FAC for retrieval). The results from these analogical processes are then reified back into the ontology, making them accessible to the first-principles reasoner. This creates a two-way street: first-principles reasoning can initiate analogical queries, and analogical reasoning can provide insights that inform logical deductions.
Datasets
Not discussed in the paper. The paper focuses on the methodology and its application to existing knowledge bases (Cyc, SAIC, KM) rather than specific datasets for training or evaluation.
Key Insights and Ideas
- Analogy as a first-class citizen: Instead of being a last resort, analogy is presented as a fundamental reasoning mechanism that can be integrated alongside logical inference.
- Bridging the gap: The analogy ontology provides the crucial formal language and mechanism to connect the symbolic world of first-principles reasoning with the relational world of structure-mapping.
- Leveraging psychological theory: Grounding the ontology in structure-mapping theory ensures that the analogical processing is psychologically plausible, potentially leading to more understandable and trustworthy AI.
- Scalability: The approach is designed to handle large, complex knowledge bases (thousands of propositions), overcoming a significant limitation of previous analogy systems.
- Flexibility through procedural attachment: This technique allows specialized, efficient analogical engines to be seamlessly integrated into general-purpose reasoning systems.
Key Results
The paper demonstrates the utility of the analogy ontology through three example systems:
- Crisis Management Reasoner: Used for answering complex questions about international crises, leveraging analogies to historical precedents. For instance, it could determine how Iran's role in one scenario is analogous to Iraq's in another, providing justifications.
- Battlespace Analogical Critiquer: This system critiques military "courses of action" (COAs) by comparing a proposed sketch to a library of critique cases. It identifies potential violations of doctrine or unwise strategies based on past experiences.
- Comparison Question-Answering System: Integrated with knowledge capture systems (like KRAKEN), it helps domain experts build knowledge bases by answering "how is X similar to Y?" questions, facilitating knowledge acquisition.
While specific quantitative results (like accuracy percentages) are not the primary focus, the paper highlights successful integration and application in challenging domains, achieving a "B" grade in the HPKB challenge, which is considered a success given the problem complexity.
Limitations and Future Work
- Alignable and Non-alignable Differences: The authors acknowledge that the current ontology does not fully capture the nuances of "alignable" (differences within aligned items) and "non-alignable" (differences between unaligned items) differences, which is an area for future work.
- Formalizing Complex Concepts: Formalizing these difference concepts has proven subtle and requires continued effort.
- Broader Applicability: The authors suggest that their approach could be valuable for existing case-based reasoning systems that still use relational information, potentially enhancing their capabilities.
Summary updated