Retrieving Semantically Distant Analogies with Knowledge-Directed Spreading Activation

Michael Wolverton, Barbara Hayes-RothAAAI 1994

Abstract

Traditional AI systems often struggle with "semantically distant" analogies—creative, cross-domain connections that lack surface-level similarities. This paper presents Knowledge-Directed Spreading Activation (KDSA), a method that utilizes task-specific heuristics to guide search through large knowledge bases. By treating "near-miss" matches as directional beacons, KDSA achieves significantly higher efficiency and scalability than standard spreading activation, enabling the discovery of innovative functional analogies.

Introduction

Analogy is a fundamental mechanism for creative problem-solving, allowing agents to apply solutions from one domain to another. A classic example is the wine press inspiring the printing press. However, retrieving such distant analogies in large-scale, multi-use knowledge bases (KBs) presents two major hurdles: unpredictability of the source domain and surface mismatch between functionally similar but structurally different concepts.

Traditional "blind" spreading activation methods are prone to combinatorial explosion as KB size increases. Conversely, manual indexing is often too rigid to capture the fluid, cross-domain connections required for innovation. This paper introduces KDSA to bridge this gap by transforming analogy retrieval into a heuristically guided search.

Methodology: Knowledge-Directed Spreading Activation

KDSA operates on the principle that a failed or partial match is not merely a computational loss but a valuable signal. The architecture, implemented in the Innovative Design by Analogy (IDA) system, integrates a spreading activation mechanism with a high-level decision cycle.

The KDSA architecture: A feedback loop where mapping evaluations inform search control to guide spreading activation.

As illustrated in Figure 1, the process begins by activating nodes related to a target concept. When an intermediate concept (IC) is retrieved, a Graph Matcher evaluates its similarity to the target. If the match is imperfect but shares critical functional attributes, the Search Control component uses it as a "beacon." By adjusting activation weights, the system steers the search toward concepts similar to the beacon, effectively "homing in" on relevant neighborhoods within the KB. Abstraction bridges, such as "Flow System," further facilitate jumps between disparate domains like highway traffic and plumbing.

Evaluation and Results

The effectiveness of KDSA was evaluated using a KB of approximately 1,050 nodes representing various mechanical and natural systems. Performance was compared against standard spreading activation (SA) across increasing KB sizes.

Performance comparison: CPU time for standard spreading activation vs. KDSA as the number of nodes increases.

As shown in Figure 2, KDSA demonstrates superior efficiency. While the CPU time for standard SA grows rapidly with the number of nodes, KDSA's growth is significantly more restrained. Theoretical modeling further suggests that while blind activation scales linearly or exponentially, KDSA's guided approach allows retrieval time to grow roughly logarithmically, making it suitable for KBs containing millions of nodes.

Qualitative results confirm the system's creative potential. For instance, IDA successfully redesigned a railroad crossing by identifying a semantically distant analogy in an on-off valve. By rewarding structural distance while maintaining functional similarity, KDSA prioritizes the discovery of novel, non-obvious solutions.

Conclusion

Knowledge-Directed Spreading Activation provides a scalable and robust solution for retrieving semantically distant analogies. By leveraging task-specific heuristics to interpret mapping failures as directional cues, KDSA overcomes the limitations of blind search in large knowledge bases. Future research will focus on generalizing search control rules and integrating the system with real-time sensory data.

Summary updated

BibTeX
@inproceedings{wolverton1994retrieving,
  author = {Michael Wolverton and Barbara Hayes-Roth},
  title = {{Retrieving Semantically Distant Analogies with Knowledge-Directed Spreading Activation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}

Loading related content...

Figure