Introduction
In the early decades of Natural Language Processing (NLP), researchers often developed specialized systems to handle specific types of knowledge. Programs like SAM (Script Applier Mechanism) were designed for stereotypical scenarios, while PAM (Plan Applier Mechanism) focused on goal-oriented actions. Peter Norvig’s 1987 work challenged this paradigm, arguing that human readers do not switch between discrete "modes" of processing. Instead, they leverage a unified pool of world knowledge. Norvig proposed a general-purpose algorithm that performs diverse inferences—from script-based expectations to pronoun resolution—using a single, consistent framework.
The FAUSTUS System
The proposed system, FAUSTUS (Fact Activated Unified STory Understanding System), operates on a knowledge base represented in KODIAK, a rich semantic network formalism. Unlike its predecessors, FAUSTUS shifts the "intelligence" of the system from complex, task-specific code to the connectivity and density of the knowledge base itself.
Marker Passing and Path Shapes
The core of FAUSTUS is a marker-passing algorithm, a form of spreading activation. As a text is parsed into conceptual representations, "markers" are placed on the corresponding nodes in the KODIAK network. These markers spread to adjacent nodes, with their "energy" dissipating over distance. When markers from two different source concepts meet at a node, a "collision" is detected.
The system then analyzes the sequence of links forming the path between the two concepts. If the "path shape" matches one of several predefined inference classes, the system suggests a potential inference.
Key Insights: Connectivity as Inference
A central insight of this work is that most textual inferences are essentially discoveries of specific types of connectivity within a semantic network. This approach allows the system to handle linguistic nuances that previous systems often ignored.
Resolving Vagueness
FAUSTUS excels at "concretizing" vague relations. For example, the word "for" in "lawyer for the client" vs. "lawyer for the enemy" implies different underlying relationships. By passing markers, FAUSTUS can find the most direct and plausible path that defines these relations based on the existing knowledge of what lawyers, clients, and enemies are.
Results and Evaluation
FAUSTUS was evaluated against classic narrative benchmarks from the Yale school of AI. It successfully replicated the results of specialized systems like SAM and PAM without requiring their explicit control structures. For instance, it could perform "script-based" inferences—such as identifying a waiter's role in a restaurant—simply by finding the connections in the network.
Furthermore, the system demonstrated emergent behavior. In one instance, when processing a story about a president discussing Nicaragua, FAUSTUS inferred that Nicaragua was the president's home country. While factually incorrect in a real-world context, the inference was a logical consequence of the provided text and the system's internal knowledge, demonstrating its ability to generate non-programmed connections.
Limitations and Future Directions
Despite its success in unifying inference types, FAUSTUS faced challenges with contextual sensitivity and computational efficiency. Marker passing can lead to a "combinatorial explosion" of paths in very large networks, requiring sophisticated heuristics to prioritize "interesting" connections. Additionally, the system relied on external parsers to convert raw text into KODIAK representations, leaving the problem of initial conceptual analysis for future work.
Summary updated