The Role of Roles: Some Aspects of Real World Knowledge Representation

Harald Trost, Ingeborg Steinacker

Abstract

This paper addresses the "knowledge problem" in early AI by proposing a pragmatic framework for Natural Language Understanding (NLU). It introduces a strict distinction between concepts—what an entity is intrinsically—and roles—how an entity functions within a specific context. By utilizing Structured Inheritance Networks (Si-Nets), the authors demonstrate how decoupling these properties leads to ontological parsimony and a robust method for tracking dynamic changes in a knowledge base.

Introduction

In the development of knowledge bases for NLU, a recurring challenge is the ambiguity of semantic mapping. For instance, is "father" a fundamental type of person or a relational role? Treating every functional relationship as a unique concept leads to a "bloated" hierarchy that is difficult to maintain and epistemologically inconsistent. This paper proposes a set of pragmatics for Structured Inheritance Networks (Si-Nets) to represent knowledge clearly by separating intrinsic attributes from functional roles.

The Conceptual Level: Defining What an Entity Is

The authors define the Conceptual Level as the layer where generic concepts are established through their intrinsic attributes. In this framework, a concept like a "hammer" is defined by its physical components and properties.

Figure 3: Conceptual representation of a hammer, defined by its parts (handle and head) and weight.

As shown in Figure 3, the concept "HAMMER" is linked to its constituent parts—a "STICK" (handle) and an "IRON/BLOCK" (head)—and a "METRIC/UNIT" (weight). This represents the object's identity regardless of its use. This approach ensures ontological parsimony by keeping the "person" or "object" concept restricted to core features, while moving social or functional data to external roles.

The Role of Roles: Functional Flexibility

The core innovation lies in the Role/Restriction distinction. While an object has an identity, it also performs functions. A "Role" explains the function of a filler in a specific context, while a "Value Restriction" defines the range of possible fillers.

Figure 1: Functional representation of the action 'to hammer', where 'hammer' acts as an instrument role.

In Figure 1, the action "TO/HAMMER" is modeled. Here, "hammer" is not a concept but an instrument role. This role is restricted to fillers that are "HEAVY/SOLID/OBJ". This distinction explains why a stone can be used as a hammer: the stone (a concept) fills the "instrument" role in the hammering action. This contextual flexibility allows the system to understand that an individual is defined not just by their parent concept, but by the sum of all roles they play across the network.

Modeling Dynamic Knowledge and Change

Representing change (e.g., a person moving house) often leads to duplicating individual nodes for every new state. The authors reject this, proposing a "change-of-state" concept. This concept links an individual to its previous state, its new state, and a timestamp. This allows the system to maintain a historical record of an individual's roles and attributes without creating redundant entities, providing a structured way to represent the passage of time.

Conclusion

By separating "what it is" from "what it does," this framework provides epistemological clarity and structural efficiency. It prevents the combinatorial explosion of sub-concepts in a hierarchy and offers a robust method for temporal tracking. This qualitative framework serves as a foundation for more sophisticated NLU systems capable of purposeful dialogue in complex domains like medicine.

Summary updated

BibTeX
@inproceedings{trost1981role,
  author = {Harald Trost and Ingeborg Steinacker},
  title = {{The Role of Roles: Some Aspects of Real World Knowledge Representation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure