Introduction
In 1980, researchers at SRI International introduced NANOKLAUS, a pioneering system that addressed one of the most significant hurdles in early Artificial Intelligence: the knowledge engineering bottleneck. While modern AI typically relies on massive, static datasets, NANOKLAUS proposed a "learning by being told" paradigm, allowing domain experts to teach a computer new concepts through interactive natural language dialog.
The Knowledge Engineering Bottleneck
During the early development of expert systems, creating a robust knowledge base required a "knowledge engineer"—an intermediary who could translate domain-specific expertise (e.g., in medicine or chemistry) into formal logical structures. This process was slow, expensive, and prone to translation errors. Haas and Hendrix sought to eliminate this intermediary by enabling subject matter experts to "tutor" the system directly using English, even if they had no background in programming or formal logic.
The NANOKLAUS Architecture
NANOKLAUS (part of the Knowledge-Learning and -Using Systems series) was designed as a pilot implementation to test the feasibility of direct knowledge acquisition. Its architecture integrated several sophisticated components for its time:
1. Pragmatic Grammar and Syntax
The system utilized a LADDER-style pragmatic grammar. This approach looked for specific syntactic patterns to trigger acquisition procedures. For instance, a sentence structure like [A] [NEW-WORD] [BE] [A] [KNOWN-NOUN] would signal to the system that a new concept was being defined in relation to an existing one.
2. Many-Sorted First-Order Logic
Knowledge was represented using a logic system that categorized entities into "sorts" or classes. This allowed the system to maintain a structured, typed understanding of the world rather than a flat list of facts.
3. The "Tangled" Hierarchy
A central feature of NANOKLAUS was its support for a "tangled" hierarchy, which allowed for multiple inheritance. This meant a single concept could have multiple ancestors, reflecting the complex, overlapping ways humans naturally categorize information (e.g., a "Submarine" being both a "Vessel" and a "Weapon System").
4. Mixed-Initiative Dialog
Unlike passive systems, NANOKLAUS engaged in mixed-initiative conversation. If a tutor provided ambiguous or incomplete information, the system would ask clarifying questions. For example, if told "A ship is a physical object," the system might inquire whether "Ships" and "Animate objects" are mutually exclusive to ensure the logical hierarchy remained consistent and free of contradictions.
Simultaneous Acquisition of Language and Logic
A core insight of the paper is the inextricable link between language and reasoning. NANOKLAUS mimics the human process of learning a new concept and the vocabulary to describe it simultaneously. When a new concept is introduced, the system updates both its internal logical hierarchy and its lexicon in tandem.
The authors noted that verbs presented a particular challenge. To truly "learn" a verb like assign, the system had to acquire not just the action, but its transitivity, passive forms, and its interaction with specific prepositions (e.g., "assign a unit to a task group").
Results and Demonstrations
The system was evaluated through a tutorial session focused on the naval domain. Starting with a small set of "seed concepts" (such as THINGS, RELATIONS, and MEASURES), NANOKLAUS successfully demonstrated its ability to:
- Define units of measure: It learned the concept of "length" and associated units like feet and meters.
- Construct hierarchies: It built a classification of naval vessels, including carriers, submarines, and destroyers.
- Process instances: It could handle specific data points, such as "The JFK is a Kittyhawk."
- Perform deduction: It answered questions based on its newly acquired knowledge, such as calculating a ship's length in feet even if the data was originally provided in a different context.
Limitations and Legacy
Despite its successes, the authors were candid about the system's "fragility." The initial grammar was limited, and the system struggled with complex linguistic features like verb particles ("pick up") and sentential objects. Furthermore, while NANOKLAUS excelled at basic deduction, it lacked the capacity for deep reasoning regarding causality, mass terms (like "water"), or learning by analogy.
The ultimate vision for the KLAUS series was the creation of an "intelligent agent" capable of not only learning about a domain but also operating external software, such as databases and simulators, on behalf of the user. This vision prefigured many of the goals of modern autonomous agents and interactive AI interfaces.
Summary updated