Porting and Extensible Natural Language Interface: A Case History

Candace E. Kalish, Matthew B. CoxAAAI 1987

Abstract

In the early development of natural language interfaces (NLIs), systems were typically engineered as bespoke solutions, tightly coupled to the specific logic and databases of their host expert systems. This paper presents KING KONG, an NLI designed specifically for portability and extensibility. By utilizing a modular architecture and declarative knowledge representation, the authors demonstrate the feasibility of migrating a complex interface between radically different domains—from air mission planning to automatic programming—with minimal manual coding and a development timeline of only six months.

Introduction

During the 1980s, the primary bottleneck in deploying natural language interfaces was the "hard-wired" nature of their design. Moving an NLI from one backend system to another often required a complete rewrite of the linguistic and semantic processing layers. The KING KONG project sought to solve this by separating general linguistic knowledge (the rules of English) from domain-specific knowledge (the logic of the expert system). The goal was to enable developers who were not experts in linguistics to port the interface to new domains using a structured, menu-driven approach.

System Architecture

KING KONG’s portability is rooted in its four-tier modular architecture, which isolates stable linguistic rules from volatile domain data:

  1. Morphological and Syntactic Layer: This component handles the fundamental rules of the English language using a simple affix-stripping model and a modified Marcus parser. Because the syntax of English remains constant regardless of the application, this layer is preserved entirely during the porting process.
  2. Semantic Layer: This layer maps vocabulary to concepts using an "A-Kind-Of" (AKO) hierarchy and a matrix of relation types (e.g., SPACE, TIME). It provides the basic framework for understanding what objects and relations exist in the world.
  3. Contextual Layer (Scenes): A significant innovation of the system is the use of "Scenes." These are declarative data structures representing stereotypical contexts or tasks (e.g., "selecting a target"). Scenes track which objects are currently "in focus," allowing the system to resolve pronouns and understand user intent within a specific discourse.
  4. The Functional "Glue": This is the only procedural component of the system. It consists of specific Lisp functions that translate the NLI's internal representations into the unique commands or database queries required by the backend expert system.

The Porting Process: A Case History

To evaluate the system's portability, the authors attempted to migrate KING KONG from its original domain—KRS (an air mission planning system using FRL frames)—to a radically different domain: ISFI (an automatic programming system that generates code from constraint networks).

The porting process was designed to be largely menu-driven. A developer defines the new domain's objects, vocabulary, and scenes through a graphical user interface, which then generates the underlying declarative code. The authors identified a two-phase evolution during the port:

  • The "Dumb" Interface Phase: By simply defining the vocabulary and basic semantics, the system can understand individual sentences. However, it lacks the ability to handle discourse or resolve references.
  • The "Smart" Interface Phase: By defining the "Scenes," the developer enables the system to follow the flow of a conversation and infer the user's broader goals.

Results and Evaluation

The migration from the military planning domain to the programming domain was completed by a single team member within six months. The vast majority of this time was spent on knowledge engineering—defining the domain model—rather than software development. Manual coding was strictly limited to the "glue" functions necessary to invoke ISFI-specific commands.

Ultimately, the ISFI version of KING KONG achieved functional parity with the original KRS version, demonstrating an ability to handle context-dependent references and complex user goals despite the fundamental differences in the backend architectures.

Discussion and Limitations

While KING KONG proved successful for expert systems, the authors noted specific limitations. The "Scene" mechanism is highly effective for task-oriented environments where user goals are predictable and limited. However, this approach may struggle with "narrative" text where context is broad or unpredictable. Additionally, while the system is highly portable, it lacks deep reasoning capabilities; it cannot reason about the causal consequences of words in the same way that contemporary "Conceptual Dependency" systems could.

Despite these constraints, KING KONG represents a significant step toward making natural language interfaces a plug-and-play component for expert systems, reducing the knowledge engineering bottleneck through declarative design.

Summary updated

BibTeX
@inproceedings{kalish1987porting,
  author = {Candace E. Kalish and Matthew B. Cox},
  title = {{Porting and Extensible Natural Language Interface: A Case History}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure