Interactive Script Instantiation

Michael J. PazzaniAAAI 1983

Introduction

In the early 1980s, natural language (NL) interfaces were often limited by their rigidity. While systems like the GUS travel agent could fill slots in a predefined frame, they frequently faltered when users deviated from a linear path or provided indirect answers. This paper introduces SNUKA, a natural language interface designed for the KNOBS planning system. SNUKA shifts the paradigm from passive story understanding to active participation, facilitating what the author terms "Interactive Script Instantiation."

In this model, the computer acts as a "servant" to a human "planner." Together, they collaborate to build complex mission plans—such as military strikes or space shuttle schedules—by leveraging "scripts," which are stereotypical sequences of events that provide the necessary context for goal-oriented conversation.

The Challenge of Interactive Planning

The primary difficulty in collaborative planning lies in the fluid nature of human dialogue. Users rarely provide information in a perfectly structured format; they use ellipsis, offer indirect answers, and omit details they assume the listener can infer. To be effective, a system must understand the causal relationships between steps in a plan and maintain a coherent conversational flow even when the user is not explicit.

Methodology: Scripts as Active Models

SNUKA’s core innovation is the application of Schank’s Conceptual Dependency (CD) theory to the planning process. Rather than treating scripts as static data structures for parsing stories, SNUKA uses them as dynamic models of a plan-in-progress. The system employs several sophisticated techniques to manage this interaction:

  • Conceptual Dependency Parsing: Utilizing the APE-II parser, SNUKA translates user input into a language-independent representation of meaning, allowing the system to focus on concepts rather than specific syntax.
  • Inference via Script Patterns: When a user asks a question that the database cannot answer directly (e.g., "What aircraft can the wing supply?"), SNUKA uses script patterns to translate the query into a format the database can process (e.g., "What aircraft does the wing contain?").
  • Conceptual Completion: If vital information is missing—such as the starting location of a flight—SNUKA performs "completion inferences," searching the database to fill in the blanks automatically based on the script's requirements.
  • Meaning Template Expectations: To handle ellipsis (e.g., a user responding with just "No, F-16s"), SNUKA builds a "meaning template" based on what it expects the user to say next. This allows it to merge fragments into the broader conceptual plan without relying on simple keyword matching.
  • Conversational "Demons": The system utilizes software routines known as "demons." These trigger automatically when specific conditions are met, allowing SNUKA to monitor the conversation and take the initiative—for instance, suggesting a specific aircraft if only one meets the current constraints.

Domain Applications

SNUKA’s capabilities were demonstrated across three distinct expert domains, highlighting its versatility in handling complex, data-rich environments:

  1. Air Force Tactical Missions: Planning strikes and air support operations.
  2. Naval "Show the Flag" Missions: Managing the deployment of naval assets.
  3. NASA Space Shuttle Scheduling: Organizing intricate crew activities and mission timelines.

Key Insights and Results

The research highlights that script-based inference is significantly more efficient than pure deductive database retrieval. By limiting the "search space" to relevant scenes within a script, SNUKA avoids processing irrelevant facts. For example, when discussing an aircraft's range for a mission, the system does not need to consider its landing gear specifications, even if that data exists in the same database.

Qualitative evaluations through sample dialogues demonstrated that SNUKA could:

  • Enforce Constraints: The system identifies and corrects user errors, such as attempting to assign aircraft to a wing that does not possess them.
  • Interpret Indirectness: SNUKA understands that a statement like "Hit the target with D4" serves as a direct answer to a question about weapon systems.
  • Automate Planning Logic: Once a user provides enough constraints, the system can automatically generate secondary details like departure times, call signs, and transponder codes.

Limitations and Future Directions

While SNUKA represented a leap forward in interactive planning, the author identified several areas for improvement. The system’s output remains largely template-based, lacking a sophisticated natural language generation component. Additionally, while the system uses heuristics to track topics, it can struggle with entirely free-form conversations. Future work suggests the need for better intention recognition—understanding why a user is asking a specific question—and the integration of graphical interfaces, such as maps and tables, to complement the text-based interaction.

Summary updated

BibTeX
@inproceedings{pazzani1983interactive,
  author = {Michael J. Pazzani},
  title = {{Interactive Script Instantiation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure