Introduction
In the field of Case-Based Reasoning (CBR), the "indexing problem" remains a fundamental challenge: how can a system efficiently retrieve the most relevant past experience to address a novel dilemma? This challenge is particularly acute in the social domain, where human relationships are governed by complex intentionality rather than simple surface features. This paper, originally presented at AAAI-91, introduces Abby, a CBR system designed to provide social advice by indexing stories through the lens of intentional causality.
The Indexing Problem in Social CBR
Standard retrieval methods, such as keyword searches or shallow feature matching, often fail to capture the nuances of social situations. For instance, a teenager’s complaint about parental restrictions on marriage is not merely about "marriage"; it involves themes of autonomy, safety, and authority. To provide contextually appropriate advice, a system must understand the underlying "why" behind social actions. Abby addresses this by representing situations as structured "intentional chains" that map the goals and motivations of social actors.
Methodology: Intentional Causality and Fixed Frames
Abby’s core innovation lies in its structured approach to indexing based on intentional causality. The system represents social situations using a fixed frame with specific slots: Theme, Holder, Partner, Thread, Goal, Plan, Action, and Effect.
These slots are populated with "microfeatures"—binary vectors representing social concepts. By using a fixed structure, the system avoids the computational complexity of open-ended graph matching. Retrieval is then performed using a variant of Tversky’s similarity metric, which calculates the overlap and differences between a user's problem (the "probe") and the stories stored in memory (the "labels").
Chain-Interactions
Social problems are rarely isolated; they often arise from the collision of multiple actors' goals. Abby models these as "chain-interactions" or "mutant chains." For example, a conflict might be represented as one person’s successful pursuit of a goal (e.g., a parent’s desire to protect) causing a negative effect on another person’s goal (e.g., a child’s desire for independence).
By explicitly encoding the relationship between agents (e.g., parent-of, lover-of) as primary features, Abby can perform "cross-contextual reminding." This allows the system to retrieve a story about a nosy mother to advise a teenager about forbidding parents, as both share the underlying theme of parental over-involvement.
Evaluation and Results
The author developed a custom corpus for the Abby lovelorn advising system, consisting of 100 to 250 hand-coded story labels paired with canned text. In a preliminary, informal study, Abby’s retrieval quality was tested against human judgment:
- Acceptability: 80% of Abby’s top-three retrieved stories were judged as "acceptable advice."
- Baseline Performance: A random sampling of stories from the same corpus yielded only a 20% acceptability rate.
These results suggest that stories serve as a form of "situation assessment," helping users reframe their own dilemmas by identifying deep situational similarities in the experiences of others.
Conclusion and Future Work
While Abby demonstrates the power of intentional indexing, several limitations remain. The system currently relies on manual coding of indices by experts and retrieves static "canned text" rather than generating tailored explanations. Future work aims to automate probe generation from natural language and scale the library to over 500 cases to further test the robustness of the intentional chain model.
Summary updated