A Collaborative Interface for Editing Large Knowledge Bases

Loren G. Terveen, David A. WroblewskiAAAI 1990

Introduction

The HITS Knowledge Editor (HKE), introduced at AAAI-90, addresses the significant challenges of navigating and maintaining massive knowledge bases (KBs). As KBs like CYC scaled to tens of thousands of interconnected units, traditional data-entry interfaces became insufficient. HKE reimagines the interface as a "collaborative partner," sharing the cognitive burden of "programming" knowledge with the human user.

The Challenge of Scale

By the early 1990s, the CYC project had reached a scale where manual editing was a "formidable" task. With over 35,000 units and an average of 35 assertions per unit, users faced two primary hurdles:

  1. Information Overload: Finding specific attributes among thousands of possibilities.
  2. Structural Integrity: Ensuring new entries adhered to complex existing conventions without getting lost in the "data space."

Core Mechanisms of Collaboration

HKE implements collaboration through three primary mechanisms: Perspectives, Checklists, and Angels.

Perspectives: Hierarchical Filtering

To combat information overload, HKE utilizes Perspectives. Rather than displaying every possible slot for an object, Perspectives filter and order information based on the current task or user history. These views are hierarchical; for example, a "City" unit inherits display rules from broader classes like "Geopolitical Entity."

Figure 1: The unit CityOfAustinTX viewed through the lens of multiple perspectives (StateCapital, City, GeopoliticalEntity), which filter and organize relevant slots.

Checklists: Intelligent To-Do Lists

When creating or editing units, HKE generates Checklists. These are dynamic to-do lists that manage dependencies. If a user references a non-existent object, the system automatically adds a task to the agenda to create that object later, preventing the user from losing their current train of thought.

Angels: Proactive Monitoring

The system includes background processes known as Angels (part of the "Hank" system). These monitors watch user commands for potential issues, such as operations that would be computationally expensive or visually overwhelming.

Figure 6: The 'Angel' Hank intercepts a graph command that would have expanded too many nodes, modifying the request to maintain readability and notifying the user.

Key Insights

The HKE project highlights several important shifts in AI interface design:

  • Asymmetrical Collaboration: The system excels at bookkeeping and tracking thousands of slot names—tasks humans find difficult—while the human provides the high-level semantic knowledge.
  • The Repair Paradigm: When users encounter confusion, the system offers "repairs," such as menus indexing relevant KB sections or suggesting "model units" to copy.
  • Evolutionary Artifacts: Interface customizations are stored within the KB itself. As users improve Perspectives for specific classes, those improvements become available to the entire community, allowing the tool to evolve alongside the data.

Conclusion

HKE demonstrated that qualitative and performance gains are possible when an interface actively assists in managing complexity. By turning minutes-long operations into seconds and reducing the need for users to memorize vast vocabularies, HKE proved that collaborative interfaces are essential for the continued growth of large-scale knowledge representation systems.

Summary updated

BibTeX
@inproceedings{terveen1990collaborative,
  author = {Loren G. Terveen and David A. Wroblewski},
  title = {{A Collaborative Interface for Editing Large Knowledge Bases}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure