Abstract
In the early era of symbolic artificial intelligence, natural language understanding (NLU) relied heavily on "scripts"—predefined knowledge structures representing common human activities. However, these systems often struggled with "garden path" stories where initial assumptions proved incorrect. This paper introduces ARTHUR (A Reader THat understands Reflectively), a system designed to address the context-selection problem. By maintaining an inference-fate graph and applying a heuristic known as the Parsimony Principle, ARTHUR can dynamically re-evaluate and "supplant" erroneous inferences, allowing for a more flexible and human-like understanding of narrative intent.
Introduction
In the late 1970s and early 1980s, AI research was dominated by the work of Roger Schank and Robert Abelson, who proposed that computers could understand stories by utilizing "scripts" and "plans." If a character entered a restaurant, the system would activate a "restaurant script," predicting that the character would sit, order, eat, and pay. While effective for stereotypical scenarios, these systems were notoriously brittle. They lacked the ability to "change their minds" when a story took an unexpected turn.
This paper tackles the context-selection problem: the requirement for a system to recognize when an initial inference is no longer viable and to replace it with a more accurate context. While contemporary Large Language Models (LLMs) handle such shifts through massive statistical attention windows, the ARTHUR system provides a symbolic framework for explicit self-correction and reflection.
The ARTHUR System: Architecture and Mechanisms
ARTHUR is designed to process stories not just as a sequence of facts, but as a series of evolving hypotheses about character goals. Its core machinery consists of several key components:
1. The Inference-Fate Graph
Unlike previous systems that committed to a single interpretation, ARTHUR maintains a data structure called an inference-fate graph. This graph tracks every tentative inference generated during the reading process. Each inference is assigned a status: active, discarded, or supplanted. This allows the system to maintain a history of its own reasoning process.
2. The Parsimony Principle
To decide which interpretation is "best," ARTHUR employs the Parsimony Principle. This heuristic posits that the preferred representation of a story is the one where the fewest number of inferred goals account for the maximum number of actions. If a character's actions seem to serve two separate goals (e.g., going to a theater AND buying drugs), ARTHUR looks for a single goal that explains both (e.g., the theater was merely the location for the drug deal).
3. Supplanting
Supplanting is the specific process of error correction. When new information contradicts an active inference, ARTHUR doesn't simply fail; it searches for a new goal that better fits the entire sequence of events. The old inference is marked as "supplanted," and the new one takes its place as the primary context.
Experimental Evaluation: Garden Path Stories
As was standard for the era, the system was evaluated on a series of "deceptively simple" hand-crafted stories designed to trigger and then correct erroneous reasoning. These are referred to as "garden path stories," analogous to garden path sentences that mislead a reader's grammatical expectations.
- The Magazine Story: Mary picks up a magazine. ARTHUR initially infers the goal of
ENTERTAIN(reading). Mary then swats a fly. ARTHUR realizes the magazine was used as a tool (an instrumental use) rather than for its intended function. It supplants the reading goal with the goal ofPHYS-STATE -10(killing the fly). - The Newspaper/Tennis Story: Carl picks up a newspaper. ARTHUR assumes he wants to read. However, the next line reveals a tennis racket was hidden underneath. ARTHUR corrects itself, realizing the newspaper was merely a physical obstacle to be moved (a
PTRANSprecondition) rather than the object of interest. - The Theater Story: A character enters a theater, but instead of watching a movie, they meet a dealer. ARTHUR successfully re-evaluates the entire trip to the theater as a cover or a meeting location for a drug transaction, rather than a pursuit of entertainment.
Key Insights
Understanding as a Process
The author argues that a final, "correct" representation of a story is insufficient for true NLU. A model of understanding must include the process of arriving at that state, including the sidetracks and corrected errors. Because ARTHUR tracks its discarded inferences, it can answer questions in a human-like way: "At first I thought he wanted to watch a movie, but actually it's because he wanted to buy cocaine."
Functional vs. Instrumental Use
ARTHUR demonstrates a sophisticated ability to distinguish between an object's default "scripty" use (reading a magazine) and its "instrumental" use (using a magazine as a flyswatter). This distinction is vital for common-sense reasoning, as it allows the system to understand that objects can serve multiple roles depending on the character's immediate goals.
Limitations and Conclusion
While ARTHUR represented a significant step forward in reflective AI, it faced several limitations. The system relied on a pre-defined library of goals and scripts, making it difficult to scale to the infinite variety of real-world human behavior. Furthermore, the author noted a lack of "psychological plausibility" regarding surprise; ARTHUR treats all corrections with the same computational weight, whereas humans find some shifts more startling than others.
Despite these hurdles, the paper's focus on self-correction and the Parsimony Principle remains relevant. It highlights a fundamental truth in cognitive science: understanding is not just about accumulating facts, but about the constant, dynamic re-evaluation of our expectations in the face of a changing world.
Summary updated