Abstract
This paper presents the Semantic Operating Language (SOL), a computational model designed to simulate human long-term memory as a dynamic network of nodes and labeled relations. Developed by David Rumelhart and Donald Norman in 1973, the system addresses core challenges in cognitive psychology and early artificial intelligence, including reconstructive memory, referent ambiguity, and causal inference. By decomposing complex verbs into semantic primitives and utilizing sophisticated foregrounding strategies, SOL moves beyond static data storage toward a reconstructive model of human cognition.
Introduction
In the early 1970s, artificial intelligence faced significant hurdles in processing human language, which is inherently ambiguous and context-dependent. Rumelhart and Norman identified three primary areas where standard logic-based systems failed: the reconstructive nature of human memory, the resolution of referent ambiguity (e.g., identifying "the girl" in a specific context), and the ability to perform common-sense causal leaps. To address these, they proposed a model where memory is an active process rather than a passive file system.
System Architecture and Methods
The SOL system, implemented in ALGOL, consists of four integrated components:
- Nodespace: A database distinguishing between "primary nodes" (general concepts) and "secondary nodes" (specific instances).
- Supervisor: Manages direct access to the nodespace.
- Parser: An augmented recursive transition network that maps English strings to network structures using Fillmore’s Case Grammar.
- Interpreter: Executes strategies embedded within the network itself.
Semantic Primitives and Verb Decomposition
A cornerstone of the SOL model is the use of semantic primitives. Rather than treating verbs as atomic units, the system decomposes them into underlying building blocks such as *LOC (state of being in a location), *CHANGE (transition between states), and *CAUSE (agency or reason).
As shown in Figure 1, the verb "put" is not a single node but a complex structure involving a CAUSE that results in a CHANGE of location (LOC) for the object. This allows the system to understand the shared logical structure between different actions, facilitating better paraphrasing and inference.
Evaluation and Linguistic Puzzles
The model was evaluated using hand-crafted linguistic scenarios designed to test specific cognitive hypotheses. These included the "Peter" sentence, the "Three Drugstores" problem for causal reasoning, and the "Telescope" ambiguity.
The "Telescope" puzzle—"I see the girl with the telescope"—tests whether the system can use its internal knowledge base to resolve whether the telescope is an instrument used by the speaker or an object possessed by the girl.
Key Insights
Memory as an Active Process
The authors argue that memory is "active" because the definition of a word is a "procedural definition"—a small program that dictates how the system interacts with the rest of the network.
Foregrounding and Referent Resolution
To manage context, the system employs "foregrounding." When a specific context is established (e.g., "Fred’s kitchen"), related items are prioritized in the search list. This allows the system to resolve definite articles like "the" by searching through a hierarchy: unique instances, foregrounded items, recent short-term memory, and finally the broader network.
Results and Progress
The SOL system successfully demonstrated:
- Disambiguation: Identifying correct referents in ambiguous sentences by querying internal states (e.g., checking if a specific girl possesses a telescope).
- Structural Generation: Creating complex network graphs from surface-level English that capture the "deep structure" of actions.
- Causal Modeling: Using
*CAUSEprimitives to provide a framework for goal-oriented reasoning, such as inferring why someone might visit multiple stores.
Limitations and Future Directions
While foundational, the 1973 progress report noted several limitations. The "intelligent search" for deep network traversal was not fully optimized, and the short-term memory model—a simple five-sentence buffer—lacked a psychologically realistic decay mechanism. Furthermore, the vast amount of world knowledge required for general human-level reasoning remained a significant challenge for the era's technology.
Summary updated