Generating Interactive Explanations

Alison CawseyAAAI 1991

TL;DR

This paper introduces the EDGE system, a novel approach to generating interactive explanations that go beyond static text by managing user interactions and adapting to a changing discourse context. It achieves this through a two-level planning system: content planning for what to explain and dialogue planning for how to interact, enabling dynamic adjustments based on user understanding and interruptions.

What Problem Does This Paper Address?

Traditional text generation systems often produce static explanations, failing to account for the dynamic nature of human learning and communication. Explaining complex topics, especially in educational or advisory contexts, often involves a dialogue where an expert checks a novice's understanding, answers clarifying questions, and adapts the explanation based on the user's evolving knowledge. Existing text planners struggle with this, as the "correct" explanation can change mid-dialogue due to user input. This paper tackles the challenge of generating explanations that are not only coherent but also interactive, responsive to user input, and adaptable to a changing discourse context.

Methods

The core of the EDGE system lies in its two-level discourse planning framework:

  • Content Planning: This level determines what information to include in an explanation. It uses "content planning operators" that represent domain-specific knowledge and relationships between sub-topics (e.g., prerequisites, subskills). These operators allow for different discourse strategies (like compare-contrast) to be selected based on the user's assumed knowledge.
  • Dialogue Planning: This level handles the how of the interaction. It uses "dialogue planning rules" that define the conventional, domain-independent structure of explanatory dialogues (e.g., opening, exchanges, closing). These rules manage the overall flow, including initiating exchanges, framing topics, and focusing the user's attention.

These two levels work in conjunction. Dialogue planning sets content goals, and content planning determines how to achieve them. Crucially, the system plans incrementally, meaning the future explanation is not fully determined upfront. This allows the discourse context (user model, current focus) to dynamically influence the planning process. The system also explicitly models and updates:

  • A hierarchical discourse structure.
  • Currently salient objects and discourse segments (focus).
  • Assumptions about the user's knowledge.

Datasets

Not discussed in the paper. The paper mentions analysis of "human expert-novice explanatory dialogues of circuit behaviour" as a basis for the system's design, but does not detail specific datasets used for training or evaluation.

Key Insights and Ideas

  • Interactive Explanations Require Dual Planning: The separation of content and dialogue planning is crucial for managing the complexity of interactive explanations. Content planning ensures the right information is conveyed, while dialogue planning ensures how it's conveyed is appropriate for an interactive setting.
  • Incremental Planning is Key to Adaptability: By planning incrementally, the system can react to user input and context changes without needing to replan the entire explanation from scratch. This makes the system more efficient and responsive.
  • Discourse Context is Dynamic and Influential: The system's ability to update and utilize a model of the discourse context (especially the user model) is central to its adaptability. Changes in user understanding or interruptions directly influence subsequent planning decisions.
  • Handling Interruptions Gracefully: The system has a specific mechanism for dealing with user interruptions, allowing it to provide clarification and then resume the interrupted discourse coherently. This is a significant step beyond static explanation generation.

Key Results

The paper describes a fully implemented system (EDGE) that generates interactive graphical/textual explanations of circuit behaviour. While quantitative results are not detailed, the system was evaluated with a small number of naive users who found the approach helpful and the explanations coherent. The system successfully demonstrates:

  • The generation of interactive explanations.
  • The management of user interruptions.
  • The checking of user understanding.
  • The modification of the explanation based on changing context.

Limitations and Future Work

  • Limitations:
    • The content planning operators, while functional, could be improved to distinguish different relation types and discourse strategies more clearly.
  • Domain-independent relations (background, conclusion) are only implicit.
  • The current system's handling of focus shifts is a simple heuristic; more complex focus shifts are not fully addressed.
  • User questions are selected from menus, and sentence-level generation uses simple templates, which could be improved.
  • Future Work:
    • Combine the approach with other explanation planning systems that don't handle interruptions or checking moves.
  • Develop more robust mechanisms for dealing with situations where a partial future explanation plan becomes completely inappropriate and needs to be abandoned.
  • Extend the system to handle more complex kinds of focus shifts.
  • Integrate a more sophisticated natural language sentence generator.

Summary updated

BibTeX
@inproceedings{cawsey1991generating,
  author = {Alison Cawsey},
  title = {{Generating Interactive Explanations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure