Six Problems for Story Understanders

Peter NorvigAAAI 1983

Introduction

In the early 1980s, the field of Natural Language Processing (NLP) was characterized by a proliferation of specialized knowledge structures. Researchers frequently introduced new constructs—such as Roger Schank’s scripts for stereotypical sequences or Robert Wilensky’s plans for goal-oriented behavior—and developed bespoke algorithms to process each one. In his 1983 paper, Six Problems for Story Understanders, Peter Norvig argues against this "algorithm-of-the-month" trend. He posits that the fragmentation of inference engines led to redundant, inconsistent systems that failed to capture the underlying unity of human comprehension.

Norvig’s alternative is FAUSTUS (Frame Activated Unified Story Understanding System), a unified system designed to handle multiple levels of narrative inference through a single, frame-based mechanism. By consolidating the processing of objects, actions, and expectations into a shared architecture, FAUSTUS sought to move AI toward a more elegant and globally consistent model of memory and reasoning.

The FAUSTUS Architecture

The core philosophy of FAUSTUS is that all knowledge should be represented as frames. Whether the system is processing a physical object, a complex social interaction, or a character's internal state, it utilizes a consistent set of inferencing tools. This approach emphasizes an "economy of mechanism," suggesting that a unified engine is less ad hoc and more robust than a collection of specialized modules.

Key Technical Components

To navigate the vast network of frames, FAUSTUS employs several distinct processes:

  • Spreading Activation: Rather than relying on explicit "triggers" to activate knowledge, the system spreads "energy" through a network of related concepts. When the activation level of a specific frame meets a threshold, it is brought into active consideration.
  • Hierarchical Composition: This allows the system to identify a "whole" based on its constituent parts. For example, the presence of an "aisle" and "tuna fish" in a narrative might invoke a "supermarket" frame.
  • The Triage System: To manage ambiguity, Norvig implemented a heuristic decision-making process. If an input has a single clear meaning, the system accepts it; if a small number of interpretations exist, it chooses the most likely; if many interpretations are possible, it waits for further activation energy to resolve the conflict.
  • Attrition: This process simulates human forgetting by gradually dropping candidate frames from active memory over time. This allows the system to settle on default interpretations while remaining open to new information.
  • Concretion: This is the automatic specialization of abstract concepts. If a story mentions a "chair" in a specific context, concretion allows the system to realize it is specifically a "lounge chair" or a "throne" based on surrounding frames.

In this ecosystem, FAUSTUS acts as the central reasoning engine, while external modules like PHRAN (a phrasal analyzer) and PHRED (a generator) handle the translation between natural English and internal frame representations.

Evaluation and Insights

As was standard for the era, Norvig evaluated FAUSTUS using hand-crafted "micro-stories" designed to test specific logical challenges. These included "garden path" narratives where initial assumptions are subverted. For instance, a story might begin in a grocery aisle but later reveal the protagonist is a janitor at a stadium. FAUSTUS demonstrated the ability to "back up" and re-evaluate its frames when faced with such contradictions.

The Metric of Richness

One of Norvig’s key insights is the use of "richness" as a metric for interpretation. When faced with multiple valid ways to connect information, the system prefers the interpretation that ties the most pieces of data together. Connecting a "cobbler" to "boots" is considered superior to treating them as unrelated entities because it creates a more cohesive internal model of the story.

Simulating Surprise

Norvig explains the human phenomenon of surprise through the mechanism of attrition. If a candidate frame (such as a character's profession) drops out of active memory due to low activation, the reader is surprised when that frame is later confirmed. This happens because the system had already settled on a more common default frame, and the sudden re-activation of the original candidate creates a cognitive shift.

Limitations and Historical Context

Despite its successes in qualitative summarization and frame switching, FAUSTUS faced several limitations. Norvig noted that his attempts to simulate "skimming"—prioritizing new input over deep inference—often led to erratic behavior that did not accurately reflect human reading patterns. Additionally, the system relied on hard-coded heuristics, such as a "triage" rule that defined a "small number" of choices as four or fewer, rather than using probabilistic or learned thresholds.

Furthermore, FAUSTUS was entirely dependent on the accuracy of the PHRAN parser. It could not "read" English directly; if the parser failed to translate text into the correct frame representation, the reasoning engine could not function.

Conclusion

Six Problems for Story Understanders remains a landmark in the "Good Old Fashioned AI" (GOFAI) era. It highlights the early challenges of context, ambiguity, and memory management that continue to occupy AI research today. While modern Large Language Models (LLMs) solve these problems through massive scale and attention mechanisms, Norvig’s work on FAUSTUS provides a foundational look at the necessity of unified architectures in the quest for true narrative comprehension.

Summary updated

BibTeX
@inproceedings{norvig1983six,
  author = {Peter Norvig},
  title = {{Six Problems for Story Understanders}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure