Abstract
This paper introduces Conceptual Overlays, a framework for natural language understanding that treats context as a dynamic set of expectations rather than static background information. By modeling human-like "common-sense algorithms," the system (prototype: EX-SPECTRE-1) interprets ambiguous actions by mapping them to inferred goals and predicted future behaviors.
The Context Problem
In the mid-1970s, AI researchers recognized that sentences are rarely understood in a vacuum. A sentence like "Jake saddled his horse" means something very different if Jake just had his cattle stolen versus if he is going for a leisurely Sunday ride. The challenge Rieger addresses is the "context problem": how can a machine formally represent the "deep conceptual level" of context to resolve ambiguity? Specifically, how can a system understand volitional actions—things people do on purpose—by linking them to the underlying motivations and situational pressures established by previous events?
Methodology: The EX-SPECTRE-1 System
Rieger proposes a mechanism called Conceptual Overlays, implemented in a LISP-based prototype called EX-SPECTRE-1. The approach relies on several novel components to model human-like comprehension.
Action Overlays and the Expectancy Cloud
Action Overlays are data structures triggered by specific patterns, known as Interceptors. For example, a "theft" overlay is activated when the system detects an input like (PETE STEAL CATTLE FROM JAKE). Once an overlay is active, it generates a "cloud" of potential future actions (expectancies).
Stepwise Indexed Algorithms (SIAs)
These expectancies are pointers to Stepwise Indexed Algorithms (SIAs)—hierarchical, common-sense plans that describe how people typically achieve goals (e.g., a plan for "physical get-back" of stolen property). The system interprets a new input by searching upward through these algorithms. If "saddling a horse" is a known step in the "get-back" algorithm, the system identifies that as the interpretation of the action within the current context.
Expectancy Selector Function ($S_A$)
To prevent a combinatorial explosion of possibilities, this function uses a ternary transition network (answering "yes," "no," or "don't know") to assign "saliency" scores to expectancies based on the situation (e.g., "Is the stolen item valuable?" or "Are the actors friends?").
Key Insights
- Context is Active, Not Passive: Context isn't just a list of facts; it is a set of "what happens next" predictions that guide the interpretation of current events.
- Interpretation as "Inverse Execution": Rieger argues that the same common-sense algorithms we use to plan our own actions are used in reverse to understand the actions of others. To understand is to find the "goal" that a specific action is meant to satisfy.
- Saliency as a Heuristic: By using a discrimination network to prioritize which expectations are most likely, the system mimics human focus, allowing it to ignore irrelevant possibilities unless the "primary" interpretations fail.
Results and Limitations
The EX-SPECTRE-1 prototype successfully demonstrated the ability to account for different contexts (e.g., interpreting "John dived under his jeep" differently if he's a mechanic vs. if he's hiding from the cops) and detecting "peculiar" actions that violate the spirit of the current overlay. Rieger also posited a Path Length Conjecture, suggesting a universal metric for how much "common sense" a narrator expects a listener to fill in.
However, the system relies on manually encoded "common-sense algorithms," a massive knowledge engineering hurdle. Additionally, while the paper mentions "Overlay Switchers" for handling anomalies (like a "theft" turning into a "practical joke"), the mechanics of transferring accumulated information between overlays were not yet fully developed at the time of writing.
Summary updated