Figuring Out What the User Wants: Steps Toward an Automatic Yellow Pages Assistant

Anatole Gershman

Abstract

This paper presents the Automatic Yellow Pages Assistant (AYPA), an early natural language interface designed to bridge the "conceptual gap" between human problem descriptions and rigid database categories. Developed in 1980, AYPA utilizes a symbolic AI framework consisting of a hierarchical semantic network (YPNET) and conceptual dependency theory to perform intent recognition. By mapping natural language queries to underlying goals, the system demonstrates how world knowledge can be leveraged to navigate complex information directories more effectively than traditional keyword-based searches.

Introduction

In the early era of "Good Old Fashioned AI" (GOFAI), researchers faced the challenge of making databases accessible to non-expert users. Anatole Gershman’s work on AYPA (Automatic Yellow Pages Assistant) addressed a fundamental problem: users often describe their needs in terms of problems (e.g., "My windshield is broken") rather than the specific service categories used by directories like the Yellow Pages (e.g., "Glass-Automobile, Plate, Window, Etc.-Repairing & Installing"). The core challenge was not just parsing grammar, but intent recognition—understanding that a broken item implies a need for repair or replacement.

Methods: The Architecture of AYPA

The system's intelligence relies on two primary components: a structured knowledge base and a conceptual analysis pipeline.

YPNET: The Knowledge Base

At the heart of AYPA is YPNET, a hierarchical associative network. It represents concepts as nodes and defines their relationships through specific link types.

  • ISA links: These handle generalizations, such as a "Mechanic" being a type of "Person."
  • REF links: These handle refinements and specific instances.
  • Frames: Each node acts as a "frame" with slots for attributes like ACTOR and OBJECT. For instance, a *REPAIR frame expects an ACTOR (typically a person or professional) and an OBJECT (the item being fixed).
Figure 1: A segment of the YPNET semantic network showing hierarchical relationships and conceptual frames.

As shown in Figure 1, the network organizes concepts from abstract roots (PERSON, ACTION, PHYS-OBJECT) down to specific domains like automotive repair. This structure allows the system to inherit properties and constraints across the hierarchy.

Conceptual Analysis and Inference

AYPA maps English sentences into language-independent "primitives" based on Roger Schank’s Conceptual Dependency (CD) theory. The inference pipeline follows three main stages:

  1. Goal Finding: The system uses heuristics to infer intent. For example, mentioning "my car" suggests a maintenance goal, whereas "a car" might suggest an acquisition goal.
  2. Goal Sharpening: The system distinguishes between related actions, such as "buying" versus "renting."
  3. Goal Elaboration: It connects parts to wholes. If a user mentions a "broken carburetor" on a "1979 Ford," the system infers the specific object of the repair.

Key Insights

A major contribution of this work is the integration of world knowledge into the search process. AYPA "knows" that glass parts are typically replaced rather than repaired. If a user requests a "windshield repair," the system can automatically pivot to "replacement" categories. Furthermore, Gershman addressed the "patchy" nature of real-world categories by attaching specific search rules to nodes—for example, directing the search from a specific part to the whole vehicle if a direct category for the part does not exist.

Results and Limitations

The prototype was tested on a subset of the New York Telephone Manhattan Yellow Pages, specifically the automotive domain. Using a vocabulary of 150 words and a network of 85 concept nodes, the system successfully mapped user queries to 46 distinct categories.

However, the paper highlights significant hurdles for scaling:

  • Scalability: The 46 categories represented only about 1% of the full Yellow Pages database.
  • Manual Engineering: The knowledge base required labor-intensive manual definition of every node and link.
  • Domain Specificity: Inference rules were highly specialized; rules for car repair would not easily transfer to medical or legal domains.

Conclusion

AYPA represents a pioneering step in intent-based natural language interfaces. While limited by the manual knowledge engineering requirements of the time, it demonstrated that effective information retrieval requires more than keyword matching—it requires a conceptual understanding of the user's underlying goals and the context of their request.

Summary updated

BibTeX
@inproceedings{gershman1981figuring,
  author = {Anatole Gershman},
  title = {{Figuring Out What the User Wants: Steps Toward an Automatic Yellow Pages Assistant}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure