Abstract
In early Artificial Intelligence, Explanation-Based Learning (EBL) often equated deductive validity with explanatory sufficiency. This paper argues that real-world explanations must also be relevant, believable, and tailored to specific goals. We present ACCEPTER, a system that evaluates candidate explanations for anomalies by checking their fit with world knowledge and their utility for the explainer's objectives. Through case studies involving unexpected deaths, the paper demonstrates how ACCEPTER filters implausible explanations and prioritizes those that satisfy the information needs of specific roles, such as a veterinarian or a detective.
Introduction
A fundamental challenge in early Artificial Intelligence was determining what constitutes a "good" explanation. While Explanation-Based Learning (EBL) was a dominant paradigm in the late 1980s, most systems operated under the assumption that any valid deductive proof was a sufficient explanation. However, in complex, real-world environments, logical validity is often insufficient. An explanation must also be relevant to the context and useful for the explainer's specific goals.
This paper identifies three primary flaws in the traditional EBL approach:
- The Single-Explanation Assumption: Systems often assumed only one explanation would be available, failing to account for the multiple hypotheses common in real-world anomalies.
- The Validity Trap: By treating explanations solely as deductive proofs, systems ignored the fact that many "valid" proofs are actually implausible or contextually useless.
- Context Blindness: Systems failed to recognize that different users (e.g., a veterinarian versus a detective) require different types of information for the same event.
Methodology: The ACCEPTER System
To address these issues, the author introduces ACCEPTER, a story-understanding program designed to evaluate explanations in the domain of unexpected deaths. ACCEPTER moves beyond simple logic by incorporating structured world knowledge and goal-directed heuristics.
The system operates through a specific evaluation cycle, as illustrated in Figure 1.
The evaluation process consists of four key stages:
- Anomaly Detection: Using Memory Organization Packets (MOPs)—structured schemas of world knowledge—the system generates expectations. When an event contradicts these expectations (e.g., a healthy racehorse dying suddenly), an anomaly is flagged.
- Relevance Testing: The system verifies if a candidate explanation addresses the specific expectation failure. It must identify the faulty belief and provide a path for revision.
- Believability Testing: Rather than relying on complex probabilistic models, ACCEPTER uses Explanation Patterns (XPs) to check the "weakest link" in an explanation. It compares hypotheses against stereotyped knowledge to reject implausible scenarios.
- Goal-Based Evaluation: The system assesses whether the explanation provides actionable information based on the user's "role theme." For instance, a veterinarian requires a medical cause, while a detective seeks actors and motivations.
Case Studies and Results
The theories behind ACCEPTER were evaluated using high-profile "reminding" cases from the 1980s, such as the sudden deaths of the champion racehorse Swale, basketball star Len Bias, and running advocate Jim Fixx.
The system demonstrated a qualitative ability to distinguish between explanations based on context. For the death of Swale:
- Veterinary Perspective: ACCEPTER accepted an explanation involving a "hereditary heart defect" because it traced a path to a medical organ failure, satisfying the veterinarian's information needs.
- Detective Perspective: The same heart defect explanation was accepted but for different reasons—it satisfied the need to categorize the death as natural rather than criminal.
- Filtering: The system successfully rejected an explanation based on the "Jim Fixx" case (death by jogging) for Swale. Its internal MOP for "jogging" specified a human actor, making the explanation incompatible with a horse.
Discussion: Key Insights
Several core heuristics emerge from the ACCEPTER framework:
- The "Weakest Link" Heuristic: An explanation's strength is determined by its least plausible hypothesis. Ranking competing explanations involves identifying and evaluating these dubious assumptions.
- Contextual Relevance: Explanations must fill specific "knowledge gaps." The same event (e.g., a car recall) requires different explanations depending on whether the focus is corporate ethics or engineering quality.
- Role-Themed Templates: Professional roles have pre-compiled information needs. These templates guide the search for relevant causes, such as physical mechanisms for medical professionals or motivations for legal investigators.
- Memory-Based Verification: Plausibility is assessed by how well a fact integrates into existing memory structures rather than through raw logical computation.
Conclusion
While ACCEPTER provides a robust framework for evaluation, several limitations remain. The "information needs" for different roles are currently hard-coded as templates; future systems should derive these needs dynamically from active goals. Additionally, the current believability ranking is relatively coarse, and the system focuses on evaluating provided explanations rather than generating them from scratch in large search spaces. Despite these limitations, the paper establishes a critical foundation for context-aware, goal-directed explanation in AI.
Summary updated