Introduction
Effective human explanation is rarely a one-way transmission of data. Instead, it is a collaborative process—a "model construction dialogue"—where the explainer helps the listener build and refine a mental model. Traditional automated systems often rely on rigid, pre-scripted templates that fail to adapt to a student's specific needs or evolving understanding. This paper introduces a hybrid architecture and a dynamic interface designed to move AI toward more flexible, interactive educational dialogues.
The Problem: Moving Beyond Rigid Templates
In early automated explanation systems, "schematic" discourse plans were the norm. These templates dictated a fixed sequence of information, making it difficult for systems to handle the "back-and-forth" of real teaching. Such systems struggled to account for what a user already knew or how the focus of a conversation shifted over time. To address this, the authors argue for a system that treats explanation as situated cognition, where knowledge is co-constructed through interaction.
The Hybrid Explanation Planning (HEP) Architecture
The proposed solution centers on the Hybrid Explanation Planning (HEP) architecture. Rather than following a single script, HEP decomposes the explanation process into five distinct, independent subtasks. This modularity allows the system to be "opportunistic," changing its strategy based on the dialogue's progress.
The subtasks include:
- User Modeling: Predicting concept familiarity using spreading activation and simulation models.
- Goal Refinement: Determining the specific type of knowledge required to answer a query.
- Model Selection: Choosing the most appropriate model (e.g., causal vs. structural) based on prioritized preferences like "Say-Something-New."
- Plan Criticism: Identifying if a chosen explanation introduces unfamiliar terms that require additional support (e.g., defining Ohm's Law).
- Ordering: Sequencing information to maximize coherence and functionality.
The IKit Interface: Enabling Deictic Reference
To support these dynamic dialogues, the authors developed the Information Kit (IKit). This "live information" interface ensures that every text snippet or diagram on the screen is linked to its underlying data structure.
This linkage enables deictic reference, allowing students to interact directly with previous explanations. For example, a student can double-click a phrase to trigger a menu of applicable actions or drag-and-drop text to construct new queries. This reduces the "semantic gap" between the user's intent and the machine's internal logic without requiring complex natural language processing.
Insights and Qualitative Results
The system was demonstrated using a domain model of basic electricity. Key results include:
- Interactive Refinement: The system successfully transitioned from simple causal explanations to detailed models involving Ohm’s Law in response to user follow-ups.
- Context Sensitivity: By tracking the "focus of attention," the system avoided redundant information and prioritized models that built upon existing dialogue.
- Direct Manipulation: The IKit interface proved effective in allowing students to steer the conversation through mouse-driven interactions, proving the utility of the "live information" approach.
Limitations and Future Outlook
While HEP and IKit represent a significant step forward, several challenges remain. The architecture currently lacks a high-level "lesson plan" or long-term tutorial strategy. Additionally, while the interface is interactive, the system does not yet "reason" with visual representations in a human-like way. Future work would benefit from empirical testing with students to quantify learning gains and the development of more robust natural language capabilities.
Summary updated