Introduction
In the early 1980s, a significant hurdle in Artificial Intelligence was the inability of natural language processing (NLP) systems to "change their minds." Human readers naturally make fallible assumptions to bridge gaps in a narrative, revising these beliefs as new information arrives. For instance, if a character "puts a quarter in a slot," a reader might assume they are playing a game, only to revise that belief to "buying a soda" upon reading the next sentence. Paul O'Rorke's 1983 paper introduces RESUND, a story-processing system designed to handle such revisions through the application of non-monotonic reasoning and Truth Maintenance Systems (TMS).
The Problem: Rigid Inference in Story Processing
Early AI story processors, such as SAM and PAM, relied heavily on rigid "scripts" or "schemata." While effective for standard scenarios, these systems struggled with belief revision. Once a system committed to a specific interpretation—such as a character being at a horse race—it lacked a formal mechanism to retract that assumption when faced with contradictory evidence. O'Rorke argues that a robust model of understanding requires a way to track the justifications for inferences, allowing for rational retraction and recovery from errors.
Methodology: The RESUND System
The RESUND system integrates schema-based story processing with techniques from AI research on problem-solving. It treats comprehension not as a static data structure, but as a dynamic process of transitioning from one set of justified beliefs to another.
1. Reasoned Understanding and Justifications
Instead of merely recording facts, RESUND records the justifications for every belief. Each inference is linked to the specific rules and prior beliefs that supported it, creating a web of dependencies. This allows the system to trace the origin of any conclusion and identify the "weakest link" when a conflict occurs.
2. Non-Monotonic Dependencies and TMS
RESUND utilizes non-monotonic logic, where the addition of new information can invalidate previous conclusions. By incorporating a Truth Maintenance System (TMS), the system can retract a belief if its underlying justifications are undermined by new evidence, ensuring the internal consistency of the story model.
3. Dependency-Directed Backtracking
When a conflict arises—such as a character being assigned mutually exclusive roles—RESUND employs dependency-directed backtracking. It traces the conflict back through the chain of justifications to identify the underlying assumptions responsible for the contradiction, allowing the system to rationally revise its interpretation.
4. Preference Policies
To resolve conflicts, the system applies a hierarchy of "preference policies." These policies prioritize certain types of inferences over others. For example, a "definitional" rule about the world is generally preferred over a "default" assumption made in the absence of specific data. Inferences like "criteriality" (assuming a whole event based on a sub-event) are treated as highly fallible and are prime candidates for retraction.
Evaluation: The "Window Text" Example
The paper evaluates RESUND using "window texts"—short, intentionally ambiguous prose. A primary example involves a man placing money at a window, a woman returning money, and the subsequent purchase of popcorn.
Initially, the system might infer a "horse race" schema based on the man "plunking down $5 at the window." However, as the story progresses to mention "buying popcorn" and "getting inside," a conflict is detected with the horse race interpretation. Through backtracking, RESUND identifies the horse race assumption as the source of the conflict. It then replaces it with a "movie theater" schema, re-interpreting the woman not as a betting clerk, but as a friend paying her own way. Another example, the "Slot Machine Text," demonstrates how the system handles ambiguity between a vending machine and a video game.
Insights and Limitations
O'Rorke's work highlights that contradictions in text are not failures but valuable signals that trigger re-evaluation. However, the 1983 implementation had limitations. The system was still under development, and the preference policies used a simple total ordering that might be too brittle for complex narratives. Furthermore, the model focuses purely on logical reasoning, ignoring the roles of affect, emotion, and motivation in human comprehension. The question of how well dependency-tracking scales to novel-length texts also remains.
Conclusion
RESUND represents an early and influential attempt to bring the rigor of non-monotonic reasoning to the field of natural language understanding. By treating understanding as a process of justified belief maintenance, O'Rorke laid the groundwork for more flexible and human-like AI story processors.
Summary updated