The KM / KnEd System: An Integrated Approach to Building Large-Scale Multifunctional Knowledge Bases

Erik EilertsAAAI 1994

TL;DR

The KM/KnEd system introduces a frame-based knowledge representation language designed to handle the "context problem" in large-scale knowledge bases. By allowing for recursive value annotations and providing a specialized graphical editor, the system successfully scaled the Botany Knowledge Base (BKB) to over 150,000 facts while maintaining human-readable contextual relationships.

Addressing the Context Problem

Building massive, multi-functional knowledge bases (KBs) is a notorious challenge in AI. Traditional frame-based systems—which organize knowledge into "frames" (objects) with "slots" (attributes) and "fillers" (values)—often struggle with context.

For example, representing the statement "Texas Bluebonnets secrete nectar that contains a low concentration of sugar" typically forces a system to "flatten" data. A standard system might require creating an awkwardly named frame like Sugar-contained-in-Nectar-secreted-by-Bluebonnet. This stripping away of natural hierarchy makes the knowledge base difficult for humans to navigate and for machines to reason about contextually. The KM/KnEd system was developed to represent these "extra details" without losing the primary relationship structure.

The KM/KnEd Framework

The authors developed a two-part solution to manage this complexity:

1. KM (Knowledge Manager)

KM is a representation language that utilizes a slot-and-filler structure but adds a critical feature: recursive value annotation. Instead of creating a new frame to describe a specific attribute of a value, KM allows users to attach annotations directly to the value itself. These annotations can, in turn, be annotated further.

Recursive nesting of attributes within the Texas-Bluebonnet frame.

As shown in the figure above, this allows the system to link the "low concentration" attribute directly to the "Sugar" filler within the "Nectar" slot of the "Texas Bluebonnet" frame, mirroring natural language descriptions.

2. KnEd (Knowledge Editor)

KnEd is a graphical user interface (GUI) designed to make this complex, nested data manageable for human editors. Its primary tool is the text pane, which displays all relations associated with a frame in a structured list. KnEd provides tools for adding, removing, and copying values, as well as navigating the massive web of interconnected facts.

The structured text pane representation of the same frame.

Scaling to the Botany Knowledge Base (BKB)

The system was evaluated using the Botany Knowledge Base, a project started in 1987 at the University of Texas at Austin to support research in question answering and intelligent tutoring. By the time of publication, the BKB had grown to over 150,000 facts using the KM/KnEd framework. This successful management of a massive dataset proved that the recursive annotation model is both computationally and organizationally viable.

Key Insights and Results

  • Context via Nesting: The core innovation is that context is treated as an attribute of a relationship rather than a separate entity. This allows the KB to mirror the way humans naturally describe complex systems (e.g., a thing, which has a part, which has a property, which has a specific measurement).
  • The Interface Bottleneck: As knowledge bases scale, the bottleneck shifts from storage to human interaction. The authors argue that specialized editing tools like KnEd are "indispensable" because raw text representations of nested annotations quickly become unreadable.
  • Functional Utility: The BKB was successfully used in practical research applications, including automated modeling, suggesting the representation was robust enough for diverse AI tasks.

Limitations and Future Directions

While KnEd makes viewing easier, there is an inherent "complexity tax" paid for this level of detail; the paper acknowledges that such tools are a requirement, not an option. Furthermore, while the system supports "distributed editing," the paper does not detail specific version control mechanisms for simultaneous multi-user edits on a 150,000-fact base.

Ultimately, the authors conclude that as knowledge bases grow in ambition, the industry must move toward these more sophisticated annotation and visualization mechanisms to remain functional.

Summary updated

BibTeX
@inproceedings{eilerts1994km,
  author = {Erik Eilerts},
  title = {{The KM / KnEd System: An Integrated Approach to Building Large-Scale Multifunctional Knowledge Bases}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure