Introduction
In the mid-1970s, Natural Language Processing (NLP) faced a significant hurdle: the "flatness" of traditional semantic networks. While these networks could represent concepts as nodes and relationships as links, they lacked the structural depth necessary for efficient word-sense disambiguation. Philip J. Hayes' 1977 paper addresses this by proposing a hybrid system that superimposes frame-like structures, termed "depictions," onto semantic nets. This architecture allows for "tightly controlled" associative searches, enabling a system to distinguish between different meanings of a word—such as "hand"—based on its immediate context.
Structural Framework: Depictions and Binders
The core of Hayes' proposal is the depiction, a subnet of a larger semantic network that describes a specific entity or concept. Unlike a generic frame, a depiction is a structured collection of nodes and links that define the internal composition and external relationships of a "depictee."
To move from generic concepts to specific instances, Hayes introduces binders. A binder is a mechanism that instantiates a generic depiction into a specific individual. For example, the generic "Human" depiction can be bound to a specific node representing "Fred," thereby inheriting all the structural properties defined in the human depiction for that specific instance.
Hierarchical Organization: Generalizers and SQN
To manage complexity and avoid data duplication, the system utilizes two orthogonal hierarchies:
- Generalizer Structure (ISA Hierarchy): This represents inheritance. A specific depiction, such as "Dog," can inherit properties from a more general depiction, like "Creature." This solves the symbol-mapping problem by allowing shared traits to be defined once at a higher level.
- SQN (Sine Qua Non) Structure (PART-OF Hierarchy): This represents essential part-whole relationships. As shown in Figure 1, an "Arm" is an SQN-offspring of a "Human." The term "Sine Qua Non" emphasizes that the existence of these parts is fundamentally tied to the whole entity.
By treating these hierarchies as independent but interacting trees, the system can navigate complex knowledge bases with high precision. For instance, it can reason that if Fred is a human, and humans have legs, then "Fred's leg" refers to a specific part of Fred, even if that specific leg hasn't been explicitly mentioned before.
Word-Sense Disambiguation in Practice
The primary application of this structured approach is resolving ambiguity. Hayes implemented these ideas in a system called CSAW. By "activating" certain depictions based on the preceding text, the system creates a context-sensitive filter.
Consider the word "hand." If the preceding sentence mentions a "clock," the "Clock" depiction becomes the primary context, making the "pointer" sense of "hand" the most accessible. Conversely, if a "gambler" is mentioned, the system prioritizes the "set of cards" sense. This efficiency is achieved through depiction chains—sequences of specific relationship types (Generalization, SQN, or Binders) that the system follows to find valid associations without an exhaustive search of all possible links.
Conclusion
Hayes' work demonstrated that adding structural layers to semantic networks significantly enhances their utility for language understanding. While the system required meticulous hand-crafting of knowledge—a significant bottleneck—it provided a robust logic for referent identification and sense selection. By separating generalization from part-whole relationships and using context as a search filter, Hayes laid the groundwork for more sophisticated knowledge representation systems in AI.
Summary updated