SNePS Considered as a Fully Intensional Propositional Semantic Network

Stuart C. Shapiro, William J. RapaportAAAI 1986

Introduction

In the mid-1980s, artificial intelligence research primarily focused on "extensional" representation, where symbols mapped directly to physical objects in the real world. However, human cognition is fundamentally "intensional." We possess the ability to reason about non-existent entities, such as fictional characters, and maintain distinct mental representations for the same physical object under different descriptions (e.g., the "Morning Star" vs. the "Evening Star"). Traditional extensional systems struggled with these nuances, failing to model mistaken beliefs or subjective perceptions effectively.

Shapiro and Rapaport address this by proposing a "computational philosophy" implemented via SNePS (Semantic Network Processing System). Their goal is to model a "mind"—specifically that of a cognitive agent named CASSIE—where identity is defined by the manner of representation rather than external physical reality.

Methodology

The core of this approach is SNePS, a propositional semantic network where nodes represent concepts and propositions themselves, rather than mere labels. The system is governed by several key principles:

  1. The Uniqueness Principle: Ensures a strict one-to-one correspondence between nodes and concepts. If two parts of the network refer to the same concept, they must converge on the same node.
  2. Meinongian Theory of Objects: A philosophical framework distinguishing between objecta (simple concepts like "Lucy") and objectives (propositions like "Lucy is rich"), allowing for "objects of thought" regardless of their existence in reality.
  3. Inference Mechanisms: SNePS employs two types of reasoning:
  4. Path-based inference: Traversing hierarchical links to derive properties (e.g., "dogs are animals").
  5. Node-based inference: Applying explicit logical rules (e.g., "if $p$ is a person and $d$ is a pet, then $p$ loves $d$").
Basic CLASS and MEMBER arc structure used for inheritance reasoning.

The system's capabilities were evaluated through a simulated conversation with CASSIE. In this setup, CASSIE is provided with information about individuals (Lucy, John) and animals (Rover), and her internal belief structure is updated as the dialogue progresses.

Key Insights and Results

The research highlights several critical features of an intensional mind-model:

  • The Mind as the Model: SNePS represents CASSIE's beliefs rather than an objective world. This allows for "de dicto" beliefs, where CASSIE can track John's potentially incorrect beliefs about Lucy separately from her own knowledge.
  • Fine-Grained Representation and Displacement: The system distinguishes between different thoughts referring to the same entity and can discuss absent or non-existent objects with equal validity.
  • Internal Semantics: Meaning is derived structurally from a node's position and links within the network, rather than external grounding.

Qualitative results from the CASSIE conversation demonstrate sophisticated reasoning. CASSIE successfully performs class inheritance, identifying a "yellow dog" as an "animal." She also handles identity resolution; when informed "The dog is named Rover," she links her existing concept of the "yellow dog" to the name "Rover."

A complex SNePS network representing CASSIE's beliefs after a simulated conversation, showing relationships between people, animals, and properties.

Furthermore, CASSIE combines path-based reasoning with logical rules to conclude that "young Lucy loves yellow Rover," while maintaining distinct belief contexts to avoid assuming universal knowledge among all agents.

Conclusion and Future Work

While SNePS provides a robust framework for intensional modeling, several areas for improvement remain. The current handling of time (via STIME and ETIME arcs) is provisional and requires a more formal temporal logic. Additionally, CASSIE's response to missing information—answering "I don't know" instead of "No"—raises questions about how a computational mind should manage ignorance. Finally, while the logic scales conceptually, demonstrating these principles in a massive, real-world knowledge base remains a future challenge.

Summary updated

BibTeX
@inproceedings{shapiro1986sneps,
  author = {Stuart C. Shapiro and William J. Rapaport},
  title = {{SNePS Considered as a Fully Intensional Propositional Semantic Network}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure