Building Concept Representations from Reusable Components

Peter Clark, Bruce W. PorterAAAI 1997 · OpenReview

Abstract

Our goal is to build knowledge-based systems capable of answering a wide variety of questions, including questions that are unanticipated when the knowledge base is built. For systems to achieve this level of competence and generality, they require the ability to dynamically construct new concept representations, and to do so in response to the questions and tasks posed to them. Our approach to meeting this requirement is to build knowledge bases of generalized, representational components, and to develop methods for automatically composing components on demand. This work extends the normal inheritance approach used in frame-based systems, and imports ideas from several different areas of AI, in particular compositional modeling, terminological reasoning, and ontological engineering. The contribution of this work is a novel integration of these methods that improves the efficiency of building knowledge bases and the robustness of using them.

Introduction

Traditional knowledge-based systems often struggle with a "knowledge acquisition bottleneck." To provide meaningful answers to complex queries, these systems typically rely on exhaustive, manually encoded declarative representations. However, it is practically impossible for developers to anticipate every potential question a user might ask.

While standard AI techniques utilize inheritance—such as a "house-boat" inheriting properties from both "house" and "boat"—this approach falls short in two critical ways. First, some concepts are "systems" of relationships (like containment) that apply across many domains; forcing these into a rigid taxonomy is often unintuitive. Second, inheritance fails to capture emergent properties that arise from concept interactions. For instance, combining "apartment" and "dog" creates new constraints, such as size limits or exercise requirements, that are not simply the sum of the individual concepts' parts.

Compositional Modeling Framework

To address these limitations, the authors propose a shift toward modular, reusable "mini-theories" called components. This framework is built upon three core pillars:

  1. Components: Instead of monolithic frames, knowledge is organized into modular components. Each component consists of Participants (the objects involved), Axioms (the rules and relationships), and Roles (the interface labels). This structure allows components to be "plugged together" regardless of the underlying representation language.
  2. Access Paths: To maintain computational tractability and avoid the "explosion" of exhaustive inference, queries are represented as "access paths"—sequences of binary predicates. This focuses the system's search space only on the information strictly necessary to answer the specific question at hand.
  3. The Classify-Elaborate Cycle: This iterative algorithm is the engine of dynamic construction. When the system encounters an object, it first attempts to classify it to a more specific type (e.g., identifying a generic "tool" as a "hammer"). It then elaborates the representation by pulling in the axioms associated with that specific component (e.g., a hammer implies the existence of a head and a handle). This cycle repeats—new details trigger further classification, which brings in more axioms—until the query is satisfied.
The structure of a representational component, illustrating the interface (Roles and Participants) and the internal Axioms.

Case Study: Bioremediation

The effectiveness of this approach was evaluated using a complex model of bioremediation—the process of using microorganisms to remove toxic waste. The system demonstrated the ability to dynamically assemble a representation for "microbe-soil-treatment" by composing more general components such as conversion, treatment, and digestion.

Dynamic assembly of a complex concept (bioremediation) by mapping reusable components like Conversion, Treatment, and Digestion.

A key finding was the system's sensitivity to context. By altering a single detail—such as changing the environment from "soil" to an "oil slick"—the system triggered a different chain of classifications and elaborations. It correctly identified that an oil-based scenario would require a "sprayer" rather than a "rototiller," showcasing the power of goal-driven, context-aware representation construction. This qualitative proof-of-concept shows that the system can construct representations on the fly that were never explicitly pre-coded.

Conclusion and Future Work

By treating knowledge as a set of reusable components rather than a static monolith, this research offers a path toward more flexible and efficient AI systems. The goal-driven nature of the Classify-Elaborate cycle ensures that the system only builds what is necessary, avoiding the intractability of exhaustive reasoning. However, challenges remain, particularly in developing principled methods for component identification and managing consistency as knowledge bases scale. Future work will also need to address how the system handles conflicting assumptions or differing perspectives within the component library, as well as the trade-off between expressivity and logical completeness.

Summary updated

Abstract

Our goal is to build knowledge-based systems capable of answering a wide variety of questions, including questions that are unanticipated when the knowledge base is built. For systems to achieve this level of competence and generality, they require the ability to dynamically construct new concept representations, and to do so in response to the questions arLd tasks posed to them. Our approach to meeting this requirement is to build knowledge bases of generalized, representational components, and to develop methods for automatically composing components on demand. This work extends the normal inheritance approach used in frame-based systems, and imports ideas from several different areas of AI, in particular compositional modeling, terminological reasoning, and ontological engineering. The contribution of this work is a novel integration of these methods that improves the efficiency of building knowledge bases and the robustness of using them.
BibTeX
@inproceedings{clark1997building,
  author = {Peter Clark and Bruce W. Porter},
  title = {{Building Concept Representations from Reusable Components}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1997},
  url = {https://openreview.net/forum?id=H1ZnY0gd-B},
}
Figure