Architecture and evolution of semantic networks in mathematics texts

Nicolas H. Christianson, Ann Sizemore Blevins, Danielle S. BassettarXiv 2019 · arXiv

Introduction

How do we quantify the "learnability" of a textbook? While educators intuitively understand that the sequence of topics matters, formalizing the "expositional flow" of information has remained elusive. This research utilizes network science and algebraic topology to map the "shape" of knowledge within linear algebra textbooks, revealing how topological structures correlate with student satisfaction.

Quantifying Expositional Flow

The primary challenge addressed is the lack of rigorous tools to visualize how concepts connect and evolve in a reader's mind. By treating a textbook as a growing semantic network—where nodes represent mathematical concepts and edges denote their co-occurrence in sentences—the authors investigate whether specific topological features can distinguish high-quality educational materials.

Methodology: From Text to Topology

The researchers developed a multi-step pipeline to transform static text into a dynamic graph:

  1. Concept Extraction: A modified Rapid Automatic Keyword Extraction (RAKE) algorithm identifies significant mathematical phrases (e.g., "vector space," "eigenvector") while filtering out common language.
  2. Network Construction: Concepts are treated as nodes, with edges formed when two concepts appear within the same sentence.
  3. Expositional Growth: To simulate a reader's progress, nodes and edges are added chronologically as they appear in the text.
  4. Persistent Homology: This topological approach detects "knowledge gaps" or cavities in the network. These range from 0D gaps (disconnected topics) to 1D gaps (circular connections missing a central link).
The multi-step pipeline for transforming textbook text into a dynamic semantic network.

Key Insights: The Shape of Knowledge

The study analyzed ten popular linear algebra textbooks, comparing their network structures against student ratings from Goodreads. Several key architectural patterns emerged:

Core-Periphery Architecture

Mathematical knowledge in these texts is not a flat web. Instead, it features a "dense core" of foundational concepts—such as matrices and vectors—introduced early and connected to most other topics. Specialized modules, like "dual spaces," form a "periphery" that connects to the core but rarely to other peripheral modules.

A visualization of a semantic network where nodes represent mathematical concepts and edges represent their co-occurrence.

The "Gap-Filling" Strategy

Textbooks are remarkably efficient at "killing" knowledge gaps. When a new concept creates a topological hole, authors typically provide the "missing link" shortly thereafter. The research found a strong negative correlation between the longevity of these gaps and textbook ratings. Books that leave "hollow" conceptual structures open for too long are generally rated lower by students.

Results and Correlation with Quality

The findings suggest that the faster a book fills its conceptual holes, the more favorably it is received. Specifically, there was a strong negative correlation (ρ ≈ -0.82) between the Normalized Average Cycle Lifetime (gap persistence) and Goodreads ratings. High-quality textbooks establish a "logical anchor" early on and minimize the time topics spend in isolation.

Limitations and Future Directions

While promising, the study is limited by a small sample size of seven rated textbooks and potential noise from OCR processing of math-heavy PDFs. Future work could employ more sophisticated NLP techniques, such as Large Language Models, to better identify semantic relationships and explore how these topological patterns manifest in classroom lectures and affect student performance.

Summary updated

BibTeX
@misc{christianson2019architecture,
  author = {Nicolas H. Christianson and Ann Sizemore Blevins and Danielle S. Bassett},
  title = {{Architecture and evolution of semantic networks in mathematics texts}},
  year = {2019},
  eprint = {1908.04911},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1908.04911},
}
Figure