Introduction
Case-Based Reasoning (CBR) systems solve new problems by retrieving and adapting solutions from past experiences. However, as an AI’s memory grows, finding the most relevant case becomes a significant challenge. Traditional retrieval methods often index events by their surface-level features (e.g., "a bank robbery"). David B. Leake argues that this is insufficient for generating useful explanations. To provide a relevant explanation, a system must understand exactly why an event was surprising—the specific nature of the anomaly.
This paper proposes a specialized indexing vocabulary that categorizes anomalies—the gap between expectation and reality—to help AI systems retrieve and adapt relevant explanations more efficiently.
The Explanation Retrieval Pipeline
The core of the proposed approach is a three-step process that shifts the focus from raw data to the type of expectation failure. Instead of searching for similar events, the system searches for similar surprises.
- Anomaly Detection: The system identifies a conflict between its internal model and the observed reality.
- Characterization: An "anomaly characterizer" maps the conflict to a specific category in the indexing vocabulary (e.g.,
SURPRISING-PLAN-CHOICE). - Retrieval: The system uses this characterization to search its library for past explanations that addressed similar types of failures.
Hierarchical Indexing Vocabulary
The vocabulary is organized into a hierarchy that allows for flexible retrieval. At the top level are broad categories of failures, which branch down into more specific sub-categories and eventually link to individual explanations.
For example, a PLAN-EXECUTION-FAILURE might contain sub-categories like PLAN-CANCELLATION or PLAN-DELAY. Each category is associated with a "knowledge structure"—a set of slots (such as Actor, Goal, and Surprising Plan)—that directs the system to the most relevant features of the situation during retrieval. This structure also enables "near-miss" retrieval: if an exact match isn't found, the system can move up the hierarchy to find a more general explanation that might still be adaptable.
Implementation and Results
This theory was implemented in ACCEPTER, a story-understanding program within the broader SWALE project. Because the research stems from the symbolic AI era, it was evaluated on a curated set of approximately 180 anomalies and explanations derived from real-world news, such as the Space Shuttle Challenger explosion and unexpected celebrity deaths.
Key findings include:
- Search Space Reduction: Indexing by the source of the failed expectation significantly narrows the search space, avoiding intractable search problems.
- Vocabulary Sufficiency: ACCEPTER successfully characterized a wide range of anomalies using nine top-level categories. The author estimates that roughly 15 categories would cover most everyday human-centric stories.
- General Strategies: Even in novel situations where no specific case exists, the anomaly category can suggest general explanation strategies, such as looking for external agents dictating a surprising plan.
Conclusion and Future Work
By indexing the "why it's weird" rather than just the "what happened," Leake provides a framework for more relevant and efficient case retrieval in explanation-based systems. While the specific case retrieval was successful, future work remains in fully implementing fallback "general explanation strategies" and expanding the vocabulary to handle the full complexity of human experience.
Summary updated