Question-based Acquisition of Conceptual Indices for Multimedia Design Documentation

Catherine Baudin, Smadar Kedar, Jody Gevins Underwood, Vinod BayaAAAI 1993

Introduction

In complex engineering environments, finding the right piece of information within a vast library of technical reports, CAD drawings, and meeting notes is a significant challenge. Traditional keyword-based search—often called syntactic indexing—frequently fails because it relies on exact word matches. A designer might search for the "function of a hub," while the relevant document describes the "operation of the disk stack."

To bridge this gap, researchers developed Dedal, a system that utilizes conceptual indexing. Instead of matching words, Dedal maps documents to an underlying semantic model of the domain. However, building these models manually is a labor-intensive "acquisition bottleneck." Dedal overcomes this by learning from the users themselves: it turns their search queries and relevance feedback into new, reusable conceptual indices.

The Foundation: A Semantic Domain Model

At the heart of Dedal is a domain model that represents the components and attributes of the system being documented. Rather than a flat list of keywords, this model is a structured hierarchy of parts and their relationships.

The domain model for a shock absorber design, showing hierarchical relationships like 'part-of' and 'isa'.

As shown in the figure above, the model defines how components like the inner-hub relate to the disk-stack, which in turn is part of a rotary-damper. These relationships—such as part-of, isa, and attribute-of—allow the system to perform "intelligent" leaps during retrieval. If a specific query about a sub-component fails, the system can use these links to find information about the parent component.

The Acquisition Loop: Turning Questions into Knowledge

Dedal’s core innovation is its ability to "learn" through a question-based acquisition loop. When a user enters a structured query, the system follows a multi-step process to either find an exact match or generate a new index based on user feedback.

  1. Structured Querying: Users combine domain concepts (e.g., "inner-hub") with task-based topics (e.g., "function").
  2. Proximity Heuristics: If no exact match exists, Dedal uses the domain model to find related information. For instance, if the user asks for the function of a part, Dedal might look for the operation of its parent assembly.
  3. Relevance Feedback: The system presents these "best guesses" to the user.
  4. Index Acquisition: If the user identifies a specific segment as relevant, Dedal automatically creates a new index (I3) that links the user's original specific query to that document segment.
The Dedal acquisition process: a failed exact match leads to a heuristic search, user feedback, and the creation of a new conceptual index.

This process effectively allows the system to capture the "abstraction level" that users actually care about, bypassing the need for manual indexing by experts.

Experimental Results and Generalization

The authors evaluated Dedal using a multimedia archive of a shock absorber design project. The experiments focused on whether indices learned from one user could be reused by others.

  • Learning and Reuse: In an initial test, a designer unfamiliar with the project generated 71 new indices. When these were reused, 70% were found to be relevant.
  • Expert Validation: While relevance dropped to 40% for experts (due to the "multiple subject problem" where a query might be broader than the document content), the system still provided a strong starting point.
  • Generalization: Most impressively, an independent designer reviewed the 71 acquired indices and rated 86% of them as relevant. This high rate of generalization suggests that the questions asked by one user are often the same questions others will have, making the learned indices highly valuable for the entire community.

Conclusion

Dedal demonstrates a powerful paradigm shift in information retrieval: treating the search process itself as a source of training data. By leveraging a structured domain model and user feedback, the system successfully automates the creation of high-quality conceptual indices. While challenges remain—such as handling complex queries with multiple subjects and ensuring the domain model stays up-to-date—Dedal provides a scalable path forward for managing the complex documentation of modern engineering projects.

Summary updated

BibTeX
@inproceedings{baudin1993question,
  author = {Catherine Baudin and Smadar Kedar and Jody Gevins Underwood and Vinod Baya},
  title = {{Question-based Acquisition of Conceptual Indices for Multimedia Design Documentation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure