RESEARCHER: An Overview

Michael LebowitzAAAI 1983

Introduction

Presented at AAAI-83, RESEARCHER is an early artificial intelligence system designed to comprehend and store complex physical descriptions from technical documents. By analyzing patent abstracts related to computer hardware, RESEARCHER constructs a hierarchical, "generalized" memory. This allows the system to develop an understanding of prototypical devices—such as a disc drive—based on specific textual examples.

The Challenge of Physical Descriptions

In the early 1980s, Natural Language Processing (NLP) research primarily focused on "scripts" or "plans" for human events (e.g., dining at a restaurant). RESEARCHER addressed the significant gap in AI's ability to process descriptions of complex physical objects. Technical texts like patents present unique difficulties:

  • Structural Complexity: Objects consist of nested parts with specific spatial relationships.
  • Linguistic Ambiguity: Phrases describing connections or positions are often difficult to parse accurately.
  • Learning from Text: The system must go beyond simple storage to create prototypical models of objects through generalization.

Methodology

RESEARCHER utilizes several novel components to represent and learn from technical text:

Memettes

A frame-based data structure called a "memette" represents objects. These can be "unitary" (a single part) or "composite" (a collection of parts).

Canonical Relation Fields

To handle the variety of spatial prepositions, RESEARCHER maps descriptions into a small set of canonical fields. This abstraction allows the system to treat different words with similar spatial meanings consistently.

Table of properties used to map spatial descriptions into canonical fields.

Generalization-Based Memory

Inspired by Roger Schank’s work, this module compares new descriptions to existing ones. When the system identifies recurring structural features across different instances (e.g., multiple disc drives having covers), it automatically generates a prototype in its memory.

Memory-Based Parsing

The parser is guided by "Memory Pointers" (nouns) and "Relation Words" (prepositions/verbs). It employs heuristics, such as a "focus memette," to resolve where new parts belong within a complex hierarchy.

Key Insights

  • From Events to Objects: RESEARCHER shifted the focus from event-based logic to structural logic. Just as "giving" has a canonical structure in event logic, "enclosure" has a canonical structure in physical logic.
  • Learning as Generalization: Learning is viewed as the automatic creation of prototypes. By identifying shared structural features in patents, the system builds an internal taxonomy of machines without explicit instruction.
  • Heuristic Focus: To solve prepositional phrase attachment, the system assumes the most recently mentioned "active" object is the likely candidate for the next relation.

Key Results

The authors demonstrated RESEARCHER's capabilities through qualitative traces of the system processing technical snippets.

  • Successful Parsing: The system correctly identified parts like "breather filters" and assigned spatial relationships, such as one being "concentrically positioned about" another.
  • Prototype Creation: Given descriptions of two slightly different disc drives, RESEARCHER successfully generated a "Generalized Enclosed Disc Drive" memette. This prototype captures commonalities (motor, spindle, cover) while maintaining unique differences as variants.
The hierarchical structure of a generalized enclosed disc drive prototype created by RESEARCHER.

Limitations and Future Directions

  • The "Level Problem": Difficulty matching objects described at different levels of abstraction (e.g., a "read/write head" vs. a "head assembly").
  • Shape Representation: The system lacks a robust method for describing 3D shapes.
  • Functional Modeling: RESEARCHER understands spatial placement but not functional purpose. Future work would require functional models to resolve deeper linguistic ambiguities.
  • Scaling: In 1983, the computational power required for "level hopping" during memory comparison was a significant constraint.

Summary updated

BibTeX
@inproceedings{lebowitz1983researcher,
  author = {Michael Lebowitz},
  title = {{RESEARCHER: An Overview}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure