Introduction
In the early 1980s, knowledge representation (KR) research faced a significant challenge known as the "levels" problem. While systems like KL-ONE were adept at the epistemological level—defining the formal structure of concepts and roles—they struggled to bridge the gap to the linguistic level, where humans use natural language to refer to specific individuals and context. KL-Conc was developed at the CNR in Pisa as an extension of the KL-ONE family to address this gap. It serves as a specialized layer that translates formal knowledge into the conceptual operations necessary for natural language interaction, such as identifying objects in a crowd or tracking discourse context.
System Architecture
The researchers implemented a multi-layered architecture to manage the transition from abstract data structures to linguistic operations. This architecture consists of several key components:
- KL-Magma: A foundation layer implemented in MAGMA-LISP that handles abstract epistemological structures.
- KL-Conc: The primary interface layer. It utilizes macros of KL-Magma functions to manage real-world objects through operations like
ANYONE(finding instances) andNEWIND(creating individuals). - META-KL-Magma: A metadescription component that allows the system to infer hierarchy or seek clarification when encountering unknown categories.
- Dual Memory System: A structure maintaining both long-term memory (permanent concept hierarchies) and working memory (a time-ordered flow for reference resolution).
Conceptual Operations and Metadescription
KL-Conc treats linguistic elements like definite and indefinite articles as specific conceptual operations rather than mere grammatical markers. For instance, the system distinguishes between "a rose" (any instance) and "the rose" (a specific individual in working memory).
A key innovation is the use of "syntactic metadescription." This allows the system to understand the structural relationship between concepts even without knowing their specific content. For example, if a user mentions "cytoplasm in a cell," the system can use its Meta layer to capture the part-whole relationship structurally, even if it lacks a biological knowledge base.
Results and Implementation
The functional implementation of KL-Conc demonstrated several capabilities:
- Individual Management: Unlike its predecessors, KL-Conc can manage individuals with arbitrary names and track them through a temporal flow in working memory.
- Automated Hierarchy Expansion: Through the
NEWINDfunction, the system can automatically expand its knowledge base. When presented with a "red rose," it ensures the individual is correctly linked to both the generic "rose" concept and the "red" property. - Simplified Interaction: By providing high-level conceptual operations, the system allows users to interact with complex Structured Inheritance Networks (SI-Nets) without requiring deep knowledge of the underlying LISP implementation.
Conclusion and Future Work
KL-Conc represents an intermediate step toward more fluid natural language understanding. While it provides a more flexible interface than pure KL-ONE, it still relies on a predefined set of operations. Future research directions include tighter integration between syntax and semantics to allow systems to interpret natural language sentences into conceptual operations more autonomously, moving beyond simple noun phrases and property assertions.
Summary updated