Classification Networks: A Knowledge Representation Scheme for Curriculum Prescription

TL;DR

This paper introduces "Classification Networks," a novel knowledge representation scheme designed to formally separate curriculum prescription from content knowledge representation in Intelligent Tutoring Systems (ITS). By using features and dimensions to categorize subject matter, this approach aims to improve the clarity, simplicity, and maintainability of large-scale ITS development.

What Problem Does This Paper Address?

Developing sophisticated Intelligent Tutoring Systems (ITS) often faces a significant challenge: clearly distinguishing between what knowledge needs to be taught (curriculum prescription) and how that knowledge should be represented for the system to use (content knowledge representation). Historically, these two crucial design processes have been intertwined, leading to confusion and hindering the development of large-scale, maintainable ITS. The authors argue that this confusion stems from a lack of separate knowledge representation schemes for each process. Previous approaches often conflated pedagogical considerations with the underlying knowledge structure, making it difficult to manage complex curricula. This paper aims to address this by proposing a formal, explicit method for representing the curriculum structure itself.

Methods

The core of the paper introduces "Classification Networks," a system built upon two primary classificatory devices:

  • Features: These represent properties that define categories of subject matter. A feature can refer to both a property and the category it defines (e.g., "member-listing" as a feature for a category of set enumeration).
  • Dimensions: These formalize perspectives for categorizing subject matter. A dimension is a data structure containing "differentiation features" that partition a set of items into mutually disjoint categories. Dimensions can be "parallel," meaning they represent different perspectives on the same subject matter.

These features and dimensions are organized into a network. The paper also defines "feature patterns" (logical expressions of features) and "characterizations" (maximal feasible feature combinations that define basic categories). The authors propose a graphical notation for representing these networks, including concepts like "dominance" (where one feature implies another) and "context" (the scope to which a dimension applies).

The process of forming topics (units of tutorial structure) involves traversing this classification network, typically using a depth-first or breadth-first search, guided by "key features" that are deemed important for teaching.

Datasets

Not discussed in the paper. The paper focuses on a theoretical framework and its application to a hypothetical electricity lesson and elementary set theories, rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Separation of Concerns: The most critical insight is the explicit separation of curriculum prescription (what to teach) from content knowledge representation (how the system uses knowledge). This is presented as fundamental for scalable ITS.
  • Formalizing Curriculum Structure: Classification Networks provide a formal, explicit way to define the "curriculum skeleton," moving beyond ad-hoc or implicitly defined structures.
  • Features as Building Blocks: Features are presented as the fundamental units for categorizing and describing subject matter, allowing for a granular and systematic approach to curriculum design.
  • Dimensions as Perspectives: Dimensions allow for multiple ways of viewing and organizing the same subject matter, reflecting the diverse perspectives a human teacher might take.
  • Key Features for Topic Identification: The concept of "key features" helps prioritize what needs to be taught, guiding the generation of tutorial topics and preventing the system from getting bogged down in every minor detail.
  • "Glue" for Coherence: The authors suggest that the relationships and structures within the classification network can provide the "glue" that connects different pieces of content knowledge, enhancing coherence in teaching.

Key Results

The paper presents a conceptual framework and a methodology for building classification networks and deriving topic hierarchies. It does not report quantitative results from empirical studies. The primary "result" is the proposed classification network scheme itself, which is demonstrated through examples like an electricity lesson and elementary set theories. The authors state that they have implemented a small ITS based on this approach for elementary set theories, using a simple overlay student model, but further development is planned.

Limitations and Future Work

  • Current Implementation: The authors acknowledge that their current ITS implementation is small-scale and uses a simple overlay student model. They plan to develop a second student model to diagnose inter-topic transitions.
  • Scope of Tutorial Objectives: The current tutorial objectives do not explicitly include teaching the contexts of task application or diagnosing student performance based on features, though the authors note this is possible with their framework.
  • Mathematical Foundations: While the paper mentions that further details on the mathematical foundations and the development of a classification network editor can be found in [Yum, 1993], these details are not elaborated upon within this paper.
  • Future Work: The authors suggest that features can be used for indexing subject tasks, teaching feature patterns and contexts, and diagnosing student performance. They also imply that the classification network could be used to generate tutorial strategies that remind students of relevant learned features.

Summary updated

BibTeX
@inproceedings{1993classification,
  title = {{Classification Networks: A Knowledge Representation Scheme for Curriculum Prescription}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1993},
}
Figure