In the early 1980s, the field of natural language processing was dominated by systems like PAM and BORIS, which operated under a "single-strategy" assumption. These models presumed that text comprehension followed a universal logic: as a reader encounters information, they generate as many inferences as possible and attempt to integrate them immediately. However, psychological research consistently demonstrated that human readers are not monolithic. Faced with the same ambiguous narrative, different individuals often arrive at radically different interpretations.
The 1983 program STRATEGIST was developed to address this discrepancy. It challenged the notion that "understanding" is a fixed process, instead proposing that comprehension is governed by a high-level "Inference Manager" that dictates how and when knowledge is applied.
The Problem of Ambiguity
Consider a simple two-sentence narrative: "Nancy went to see a romantic movie. She was depressed."
A reader might infer that the movie was tragic, causing her depression (Interpretation A). Alternatively, a reader might infer that Nancy was already depressed and sought out a romantic movie as a form of escapism (Interpretation B). Traditional AI models of the era typically produced only the first interpretation because they favored immediate, predictive connections. They lacked the flexibility to model the second interpretation, which requires a retrospective re-evaluation of the first sentence based on the second.
The STRATEGIST Architecture
The core innovation of STRATEGIST lies in its bifurcation of processing rules into two distinct categories:
- Content-driven rules: These represent universal world knowledge (e.g., "people go to movies for entertainment" or "sad events cause depression"). These rules generate potential inferences based on the semantic content of the text.
- Strategy-driven rules: These function as the "Inference Manager." They do not contain world knowledge themselves; instead, they govern the application of content rules. They decide whether to commit to an inference immediately or hold it in abeyance.
By separating the "what" (content) from the "how" (strategy), the authors could simulate different cognitive styles by simply toggling a processing parameter within a Conceptual Dependency (CD) representation framework.
Modeling Human Cognitive Styles
STRATEGIST specifically modeled two primary reading strategies identified in psychological literature:
- Perseverers: These readers make specific inferences immediately upon encountering a stimulus. They "persevere" with their initial hypothesis until it is explicitly contradicted by the text. This strategy relies heavily on predictive inferences—looking ahead to anticipate what comes next.
- Recencies: These readers are more cautious. When faced with ambiguity, they delay making specific inferences, leaving "loose ends" in their mental model. When new information arrives, they use it to look backward and resolve previous ambiguities. This relies on postdictive inferences.
Experimental Results and Insights
The authors evaluated STRATEGIST using targeted story fragments designed to elicit divergent interpretations. Beyond the "Nancy" story, the system processed a narrative about "Melissa and Tyler" involving a marriage proposal and crying.
The results provided a computational proof-of-concept for human variability. By switching strategies, STRATEGIST could generate the same range of interpretations found in human subjects. For the Melissa story, the "Perseverer" mode interpreted her crying as a sign of unhappiness with the proposal, while the "Recency" mode interpreted it as "crying for joy" because the proposal fulfilled a latent goal.
Key insights from the study include:
- Timing as a Variable: The difference in "understanding" is not due to a difference in knowledge, but a difference in the timing of commitment to that knowledge.
- The Interpretive Spectrum: The authors posited that these strategies exist on a spectrum. An extreme Perseverer might exhibit "paranoid" understanding (forcing all facts into a rigid preconception), while an extreme Recency reader might fail to connect ideas at all if they wait too long to make any inferences.
- Sensitivity to Order: The system successfully mimicked human behavior where reversing sentence order leads to reversed interpretations, proving that the Inference Manager is highly sensitive to information flow.
Limitations and Legacy
While a landmark in cognitive modeling, STRATEGIST had notable limitations. Its knowledge representations were relatively simplistic, lacking a deep "common sense" framework for complex goal-planning. Furthermore, while the program could switch between strategies, the authors lacked a formal theory for the exact linear order in which rules should be applied within those strategies. The system was also limited to very short (2-3 sentence) fragments.
Despite these constraints, STRATEGIST remains a pivotal example of how AI can be used not just to solve tasks, but to model the inherent variability of human cognition. It moved the field away from rigid logic toward a more nuanced view of intelligence as a strategy-driven process.
Summary updated