Abstract
Building scalable Intelligent Tutoring Systems (ITS) requires a clear separation between pedagogical curriculum design and epistemological knowledge storage. This paper introduces Classification Networks, a formal representation scheme that uses features and dimensions to define a "curriculum prescription" layer. By decoupling the teaching logic from raw data, the framework enables the automatic generation of structured lesson plans and more maintainable tutoring architectures.
Introduction: The Challenge of Scaling ITS
A persistent hurdle in the development of Intelligent Tutoring Systems (ITS) is the "tangling" of pedagogical goals with epistemological structures. Historically, the human-centric logic of what to teach first has been fused with how the computer stores facts. This entanglement creates "black box" systems that are difficult to update; changing a teaching sequence often necessitates a complete rewrite of the underlying knowledge base.
To address this, the authors propose a skeletal "curriculum prescription" layer. This layer acts as the glue connecting various subjects without becoming bogged down in raw data, allowing for a more maintainable and scalable approach to automated instruction.
The Classification Network Framework
The core of this proposal is the Classification Network, a formal scheme to categorize subject matter using two primary components:
- Features: Specific properties or categories (e.g., "Ohm’s Law").
- Dimensions: Perspectives that partition the Universe of Discourse (UoD). For example, in the domain of electricity, one dimension might be "Law" (partitioned into Ohm’s and Kirchhoff’s), while another might be "Problem Type" (Qualitative vs. Quantitative).
Key Technical Concepts
- Context and Dominance: Dimensions can be nested. A "non-primary" dimension only applies within a specific context. For instance, "Current-Voltage relationships" might only be relevant within the context of "Ohm's Law."
- Feature Patterns: These are logical expressions (AND/OR) of features that define a specific "Topic."
- Key Features: Designers can distinguish between "Key" features (deserving dedicated lessons) and "Non-key" features (conceptual scaffolding used to link ideas).
Automated Curriculum Generation
One of the primary benefits of this representation is the ability to automatically generate lesson plans. The paper describes an algorithm that performs depth-first or breadth-first traversals of the network to assemble "Topics." These topics are objects containing slots for instructional content, relevant dimensions, and transitions to subsequent concepts.
This "Zoom Lens" approach allows the system to start with a wide-angle view of major ideas and progressively "zoom in" to finer details. The "index feature pattern" acts as explicit "glue," providing coherence to the student's learning path. Furthermore, the system can intelligently order topics based on complexity, prioritizing simpler "dead-end" concepts to manage cognitive load.
Implementation and Results
The framework was validated using two domains: Electricity (to illustrate parallel dimensions) and Elementary Set Theory (where a small-scale ITS was implemented). The results demonstrated that classification networks move beyond simple prerequisite trees, providing a mathematical and graphical notation for complex curriculum structuring. By decoupling the curriculum from specific tasks, designers can modify the "presentation skeleton" of a course without altering underlying problem generators or task simulators.
Discussion and Future Work
While the current implementation successfully automates topic generation and uses an "overlay" student model to track progress, several areas for future work remain. The authors aim to implement a second model for diagnosing "inter-topic transitions" to understand why students struggle to move between concepts. Additionally, while traversal is automated, the initial creation of the network still requires significant human expertise. Future iterations may also expand the diagnostic scope to include teaching the context of task application.
Summary updated