A Viewpoint Distinction in the Representation of Propositional Attitudes

John A. BarndenAAAI 1986

Abstract

In the field of Knowledge Representation (KR), the modeling of "propositional attitudes"—mental states such as beliefs, desires, and intentions—is a cornerstone of artificial intelligence. This paper identifies a fundamental confusion in existing methodologies, specifically regarding the distinction between representations used by an agent for internal cognition and those used by a researcher to describe an agent's mental state. By analyzing the systems of Maida & Shapiro and Creary, the author demonstrates how failing to maintain this distinction leads to logical inconsistencies and the psychological implausibility of "imputing" complex theoretical structures onto the agents being modeled.

Introduction: The Ambiguity of Representation

A central challenge in AI is enabling machines to reason about the beliefs of others, particularly when those beliefs differ from reality or from the machine's own knowledge base. For instance, representing "John believes the morning star is the evening star" requires a system that can handle intensional concepts rather than just physical extensions.

The core problem identified is the ambiguity of the term "represents." When a node in a semantic network is labeled "John," does it denote the actual person in the physical world, or is it a theoretical "ambassador"—a handle for the agent's internal concept of John? Existing systems often conflate two distinct viewpoints:

  1. Viewpoint A: The representation as a basis for an agent's own cognitive processing.
  2. Viewpoint B: The representation as a theoretical tool for a researcher to describe an agent.

Critical Analysis of KR Schemes

The paper performs a conceptual audit of two major 1980s KR schemes: Maida & Shapiro’s semantic networks and Creary’s logic-based neo-Fregean approach.

Maida & Shapiro’s Semantic Networks

In this system, nodes represent "intensions" (concepts). However, Barnden identifies a "non-uniform dereferencing" problem. To process a proposition, the system must inconsistently decide when to look at what a node points to and when to treat the node as the object itself. This inconsistency poses a significant hurdle for automated reasoning engines.

Figure 2: Representation of a nested belief where Bill believes a specific proposition (John is taller than Mary).

Creary’s Logic-Based Scheme

Creary’s approach treats concepts as first-class objects within a logical framework. While powerful, Barnden shows that it suffers from the "Problem of Imputation." When a researcher uses complex functions to describe an agent's belief, the logic often implies that the agent itself is thinking in those complex terms. This is psychologically implausible; an agent can believe a proposition without possessing the complex theoretical machinery the researcher uses to describe that belief.

Figure 4: Representing a belief about a description, such as Bill believing Mike's favorite proposition.

Key Insights: Ambassadors and Imputation

Barnden introduces the concept of ambassadorial representation. An expression in a program should be viewed as an abstraction of a mental state rather than a direct pointer to a physical entity. Confusing "denoting a concept" with "being an abstraction of a concept" leads to what he terms "representational infelicities."

Furthermore, the paper highlights "lifting levels." While some systems successfully represent people as intensional concepts, they often fail to do the same for the propositions those people believe, leading to systematic confusion in nested attitudes.

Conclusion

The paper serves as a critical warning for researchers in Knowledge Representation. The primary takeaway is the "Viewpoint Requirement": any successful KR scheme must explicitly define whether it serves as an internal processing tool or an external descriptive tool. While the author hints at a new representation scheme designed to avoid these imputation difficulties, the primary contribution here is the rigorous identification of the conceptual traps inherent in modeling the minds of others.

Summary updated

BibTeX
@inproceedings{barnden1986viewpoint,
  author = {John A. Barnden},
  title = {{A Viewpoint Distinction in the Representation of Propositional Attitudes}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}

Loading related content...

Figure